Entropy guided multi level feature fusion network for high precision content based image retrieval

Abstract

Content-based image retrieval (CBIR) is essential for managing and searching massive image repositories across a wide variety of applications. Nevertheless, some traditional CBIR systems exhibit low retrieval accuracy because they use predetermined feature weights, lack semantic gaps, and poorly exploit heterogeneous visual features. To overcome such difficulties, the present study will introduce a multi-feature adaptive CBIR framework that combines deep and handcrafted features using an information entropy-based fusion and a trust-based weighting system. Deep convolutional models, combined with complementary low-level descriptors, are used to extract discriminative features in the proposed approach. A PageRank-based similarity propagation strategy is also used to narrow image ranking by leveraging similarity relationships across the globe. Evaluation is performed using standard retrieval measures, such as Mean Average Precision (mAP), Precision at K, Recall at K, and NDCG. The experimental results show that the proposed approach consistently improves performance across benchmark datasets. The framework boosts mAP by up to 8.6% over traditional fixed-weight fusion methods, while Precision@10 and NDCG@10 increase by 6.2% and 7.4%, respectively. The statistical analysis shows that these improvements are significant at the 95% confidence level, indicating that retrieval behavior is robust and reliable. These findings confirm the efficiency of entropy-driven adaptive fusion and ranking refinement in overcoming the major drawbacks of current CBIR systems, and the suggested framework is appropriate for large-scale image search in practice.

Introduction

The problematic issues discussed in this paper are accurate and reliable image retrieval in large, heterogeneous image collections, where fixed-feature fusion strategies tend to overlook a variety of visual features. To address this setback, we introduce an adaptive content-based image retrieval system that combines deep and handcrafted features using entropy-based weighting and trust-based refinement. The given practice also improves ranking reliability through global similarity propagation, adding to a strong, scalable solution for current CBIR systems. Innovation and imagination drive technological advancement to new heights. One such concept is using an image processing stream as the foundational module in a DNN, which allows mapping inputs to outputs using a variety of uncertainty logic¹. The picture will be reduced to its constituent bits and treated as a digital structure. Because its technique relies on file-name searches rather than the content, current algorithms make picture or video classification difficult². The DNN must assign distinct properties to content classifications based on the user’s query. The Deep Learning techniques methodology includes merging all data into a single set and learning its contents by disentangling their characteristics³. The database can hold a limited number of features in its dedicated data centre. As it processes data, the deep learning approach demonstrates its full potential by intelligently extracting its contents⁴. One category of soft computing phenomena is deep learning, which enables the retrieval of data from millions of separate images⁵.

For decades, researchers in multimedia have studied feature representation and similarity assessment in detail, as these aspects significantly impact the effectiveness of content-based image retrieval systems⁶. Despite numerous approaches, this problem remains a significant obstacle in CBIR research⁷. This is primarily because of the famous semantic gap problem, which arises when machines capture low-level image pixels while humans perceive high-level semantic concepts⁸. Looking at it from a high level, this problem may stem from the more fundamental question of how to create and train intelligent robots to perform jobs generally associated with humans⁹. Searches that depend just on metadata are imperfect if the annotations aren’t accurate and comprehensive, and it’s not always easy to manually annotate photographs in an extensive database with keywords and other information; thus, CBIR is a good alternative¹⁰.

It is impossible to assess the effectiveness of CBIR systems, as with keyword image search, because the criteria are arbitrary and poorly defined¹¹. Newer network- and graph-based systems have offered simple, attractive alternatives to current approaches. There has been a rise in interest in CBIR as a solution to the limitations of metadata-based systems and for its many uses in efficient image retrieval¹². Research on CBIR has begun to incorporate user-centred design and easy-to-use interfaces to meet these demands¹³. The original intent of CBIR systems was to search databases based on image attributes, including colour, texture, and form; however, many additional aspects are being used¹⁴. The standards developed for picture classification still have issues with scalability and misclassification. To find the most relevant photos, an effective DNN algorithm compares these properties to those in the image database¹⁵.

The suggested solution employs an innovative combination of adaptive methods and multi-feature fusion to enhance image retrieval accuracy. This method uses information entropy to dynamically integrate multiple aspects, such as shape, texture, and colour, rather than standard CBIR approaches that rely on predefined feature sets or individual methods. Deep learning methods to extract more abstract and discriminative image characteristics further improve the system’s capacity to retrieve relevant information. The approach is more resilient across varied datasets because it adapts to varying image complexity and appropriately adjusts feature contributions. With its novel combination of deep learning, information entropy, and adaptive feature fusion, the suggested technique stands out from the current state of the art. It provides better retrieval performance in intricate image databases.

The research contribution of this paper is as follows:

• To train large-scale deep belief networks to learn good feature representations of pictures; this will serve as a deep learning foundation for CBIR.

• Comprehensive assessments of deep belief networks were trained to learn feature representations for various CBIR tasks across different environments, and this body of empirical research is extensive.

• A picture retrieval approach that uses several characteristics is presented and applicable in diverse areas such as digital libraries, historical research, fingerprint identification, and crime prevention.

Novelty of the paper

Even though the suggested system includes well-known elements such as ResNet50, VGG16, KNN, and information entropy, the originality of the proposed work lies in the systematic interaction of these techniques within a single, entropy-directed, adaptive multi-feature fusion system. In contrast to traditional hybrid CBIR systems that use fixed or manually determined feature weights, the proposed system estimates retrieval trust. It dynamically optimizes PageRank-based weights, enabling query-adaptive feature fusion and improved retrieval resilience.

The novelty-based contributions are summarized as:

• A multi-feature fusion architecture, guided by an entropy model to adaptively consider feature discriminability rather than using predetermined feature weights.

• A retrieval trust estimation algorithm that utilises correlation feedback to estimate the accuracy of individual features for each query.

• A feature weight optimization strategy based on PageRank that approximates the influence of features on each other with a probability transfer matrix.

• An integrated end-to-end CBIR system that incorporates both deep learning and handcrafted characteristics in the same adaptive framework.

Research problem and objective

Although much research on content-based image retrieval has been conducted, current systems remain weak in handling feature ranking, exhibit unreliable retrieval accuracy across varying image types, and have limited resistance to feature redundancy. The study aims to develop a dynamic CBIR system that dynamically assesses feature relevance using entropy and trust-based methods, while optimizing similarity ranking via global propagation. This is to achieve reliability, scalability, and accuracy in large-scale image retrieval.

This paper is structured as follows: Sect. Introduction presents the related research on CBIR. In Sect. Related study, the proposed DNN methodology is explained, and in Sect. Proposed method, its performance is discussed and analyzed.

Related study

A major obstacle in big data applications is retrieving images from enormous collections. To overcome this, content-based image retrieval algorithms optimize retrieval accuracy by fusing multiple features to compensate for missing features, thereby boosting performance. An information-entropy-based adaptive feature fusion method is essential for efficient image retrieval.

Big Data Image Analysis (BDIA).

Retrieving the necessary photos from massive amounts of data effectively is a big data challenge. Consequently, one promising approach to tackling this issue is a system for retrieving images based on their content. Another critical aspect of improving retrieval system performance is a multi-feature-based image retrieval system’s ability to compensate for the absence of a single feature partially. Retrieval of images using multi-feature fusion relies heavily on feature selection and fusion techniques. It provides an information entropy theory–based adaptive feature multi-feature fusion approach for picture retrieval by Lu et al.¹⁶. Initial retrieval results are obtained after extracting image features and constructing a distance function to compute similarity using the information entropy proposed in this research. The next step is to automatically select the practical features based on the retrieval trust derived from the accuracy of single-feature retrieval using correlation feedback. The next step is to build the probability transfer matrix, which involves initializing the feature weights of chosen features using the average weights. Then, they update the initialized feature weights using the PageRank method to get the final weights. This paper still lacks accuracy in the image retrieval process. Rupa et al.²³ have suggested a convolutional neural network-based system to detect and visualise exposure to pandemic disease using image data. The paper has shown that deep feature extraction is useful in analyzing a large number of images. The method, however, does not emphasize the concept of adaptive feature fusion in image retrieval or classification accuracy.

Dubey et al.²⁷ provide a comprehensive survey of deep learning techniques for CBIR based on supervision type, network architectures, and feature descriptors. It emphasizes that deep features outperform traditional descriptors and also highlights problems such as scalability and the persistence of semantic gaps in real-world retrieval situations.

Teacher Learning Based Algorithm (TLBA).

According to Jain et al.¹⁷, a content-based picture retrieval system based on feature selection is proposed in this study. The optimization methodology employed in the feature selection process was based on teacher learning. It is a population-based search method that uses an algorithm based on teacher learning. The feature data collected from the feature space matrix is then passed on to the classification process using an algorithm based on teacher learning. The suggested method used a typical coral picture collection and was implemented in MATLAB. Precision and recall were used to assess the performance of the presented algorithm.

Support Vector Machine (SVM).

Bakheet et al.¹⁸ present an adaptive feature fusion framework for Content-Based Image Classification. This framework automatically extracts salient multimodal features, including low-level visual features, from image regions and adaptively fuses them. CBIR is a potential application of computer vision to the image retrieval problem, which involves searching large-scale image databases according to a user’s query image. Developing high-accuracy CBIR systems is made more difficult by the semantic gap, which arises from the inherent difference between the machine-generated digital representations of images and the high-level semantic concepts of images.

Data Mining (DM).

A feature selection system that uses texture and colour properties is suggested. The Localized Discrete Cosine Transform, localized grey-level co-occurrence matrices, and local binary patterns describe the texture characteristics. In contrast, colour moments and histograms in the RGB and HSV colour spaces capture colour features. To evaluate the block details at the blocks’ DCT coefficients and the grey level co-occurrence matrix. Additionally, LBP is used for rotation-invariant picture texture information. The data mining criteria eliminate unproductive features after feature extraction, as described by Asadi Amiri et al.¹⁹.

Artificial Intelligence (AI).

The recovery of searchable encrypted images in a cloud setting has been the subject of much research lately. However, achieving the desired results is impossible due to the flawed encryption method and a single feature description. Consequently, improving performance by extracting and fusing multiple efficient features is a significant challenge in encrypted image retrieval. Ma et al.²⁰ present a cloud-based encrypted image retrieval system that is searchable and employs multi-feature adaptive late fusion to achieve this goal.

Content-Based Image Classification and Retrieval (CBICR) is an innovative framework introduced by the author¹⁸. It uses stacked random forests for feature fusion and automatically extracts and fuses salient multimodal features, such as low-level visual features (e.g., colour, edge histogram, Hu moments), from different regions of images. Then, they applied a specific random forest sampling and classification process to combine the characteristics used adaptively. A large-scale dataset of real and synthetic images is used to conduct various experiments and evaluate the performance of the proposed approach.

In²², the author proposed an adaptive multi-feature fusion approach based on cross-entropy normalization for efficient image retrieval. At first, they extract several low-level deep learning features (e.g., SIFT) and high-level semantic features. First, the query image’s similarity scores to the target dataset are computed under each feature representation level. Second, we cross-entropy-normalize the initial similarity score ranking curve to estimate its tail using an independent reference dataset. The area under the ranking curve determines the associated feature’s value; a smaller area denotes a more acceptable feature. Shaik et al.²⁴ proposed a safe medical image abstraction model based on convolutional neural networks. The article highlights the use of feature learning to protect sensitive image data. Even though it is effective at securing medical data, the approach does not consider similarity ranking or a multi-feature fusion strategy, which are needed for content-based image retrieval. Rupa et al.²⁵ proposed a CNN steganographic system to ensure the safety of healthcare images in telemedicine. It is a combination of deep feature extraction and data-hiding techniques to enhance confidentiality. Although it is strong in terms of image transmission safety, it is not concerned with retrieval accuracy or entropy-based feature fusion schemes.

Finally, the statistically elaborated regions adaptively give fusion weights to each characteristic. Table 1 summarizes different methods discussed in the related works section, including their advantages and limitations.

Table 1.

Summary of the related works.

S.No	Methods	Advantages	Limitations
1	Big Data Image Analysis by Lu et al.16	- Enhances retrieval accuracy using adaptive multi-feature fusion. - Employs information entropy theory for effective feature selection.	- High computational complexity due to entropy-based calculations. - Limited scalability for very large datasets.
2	Teacher Learning Based Algorithm by Jain et al.17	- Population-based feature selection improves retrieval optimization. - Demonstrates good performance on coral image datasets.	- Sensitive to initial population selection. - Convergence speed affects retrieval efficiency.
3	Support Vector Machine by Bakheet et al.18	- Automatically extracts and fuses multimodal features. - Reduces semantic gap through adaptive fusion.	- High computational cost for large-scale datasets. - Requires careful kernel and parameter tuning.
4	Data Mining by Asadi Amiri et al.19	- Integrates texture and color features for precise selection. - Eliminates redundant and unproductive features.	- Increased computational overhead due to multiple feature extractors.
5	Artificial Intelligence by Ma et al.20	- Enables searchable encrypted image retrieval using adaptive late fusion. - Supports cloud-based image retrieval applications.	- Performance depends on feature extraction quality. - Potential privacy and security concerns in cloud environments.
6	Deep Learning Survey by Dubey et al.	- Provides a comprehensive analysis of deep learning-based CBIR techniques. - Demonstrates superiority of deep features over handcrafted descriptors.	- Highlights unresolved issues such as the semantic gap and scalability. - Lacks unified evaluation benchmarks.
7	CNN-Based CBIR by Shaik et al.	- Uses CNN classification scores for effective retrieval. - Achieves high accuracy on benchmark datasets.	- Requires careful hyperparameter tuning. - Limited generalization across domains without retraining.
8	Concept-Based CBIR by Datcu et al.	- Improves retrieval accuracy using semantic concept extraction. - Enhances interpretability of retrieval results.	- Relies on domain-specific processing. - Limited adaptability to generic CBIR tasks.
9	Fragile Watermarking by Sahu et al.	- Enables tamper detection and localization in image data. - Suitable for secure image applications.	- Not directly optimized for content-based image retrieval tasks.
10	Reversible Data Hiding by Gupta et al.	- Preserves image quality while embedding auxiliary information. - Useful for secure image transmission.	- Focuses on data hiding rather than retrieval performance.
11	Steganographic CBIR Framework by Rupa et al.	- Combines deep feature extraction with secure data embedding. - Enhances robustness for sensitive image applications.	- Retrieval accuracy is not the primary optimization objective.

The recently introduced TLBA is among the most forward-thinking approaches to image retrieval. Feature selection enhancing content-based systems. In addition, a Support Vector Machine is used within an adaptive feature fusion framework to improve picture categorization accuracy further. Data mining algorithms for practical feature selection based on texture and colour attributes. Thus, these techniques are essential for developing high-accuracy CBIR systems and encrypted picture retrieval in cloud settings, thereby helping close the semantic gap in image representation.

Proposed method

In this section, the proposed architecture for entropy-guided multi-level feature fusion in content-based picture retrieval is presented. The retrieval process includes feature extraction, similarity calculation, feature trust estimation, adaptive weight optimization, and final picture ranking, all covered in the order shown in the architecture. Fig. 1 presents the overall workflow of the proposed system, with a focus on the interaction among adaptive fusion methods, handcrafted features, and deep learning features.

Fig. 1.

Content-based image retrieval.

Framework overview

Modern image retrieval systems employ sophisticated methods, such as deep learning and custom feature extraction, to efficiently return images in response to user queries. These systems perform image operations such as feature extraction, fusion, and matching using tools like edge descriptors, texture analysis, and colour histograms, which can be handy here by leveraging information entropy and correlation feedback to improve feature selection and weighting. This ensures that the retrieval results will be accurate and relevant.

They use a mix of the colour histogram, colour moment texture, edge histogram descriptor, and another feature extraction approach. There is a way to include additional characteristics for more efficient retrieval in the future—any combination of these techniques for retrieval that works best for your application. A building data image is given to the user via the user interface (UI) as feedback on its relevance. An image’s colour distribution may be seen in a colour histogram. Colour histograms are helpful for digital photos because they show how many pixels fall within a predetermined set of colour ranges that cover the image’s colour space. This is a condensed form of the colour attribute used to describe a colour picture. Most of the information about the colour distribution may be found in the three low-order moments. A few texture properties, including energy, homogeneity, contrast, and entropy, can be retrieved from an image’s grey-level co-occurrence matrix. The picture reflects the characteristics. This CBIR system averages all the similarity measures, as shown in Fig. 1.

Validity of probability distributions in entropy computation

All feature responses are converted to valid probability distributions before entropy is computed. The individual feature vectors are initially normalised using L1 normalisation, ensuring that the sum of their components is equal to 1. Negative or undefined values are removed by first doing minmax scaling before normalization. Consequently, the probability mass function is such that: q_i ≥ 0. This normalization step ensures that entropy estimates remain constant and that numerical instability does not occur when computing feature weighting and similarity.

The equations from 1 to 20 will use all the mathematical symbols defined in Table 2.

1

Table 2.

Notation and symbol definitions.

Symbols	Definition
	Query image
	Image database
	i-th database image and its index
	Feature index and total number of features
	Similarity score and mean similarity of feature l
	Balanced precision function of feature l
	Probability distribution and entropy of feature l
	Retrieval trust and rank-based trust of feature l
	Adaptive feature weight and weighted similarity contribution
	Comprehensive similarity score of image Ii
	Summation index and number of database images
	Normalization constant, ranking position, iteration index
	Convergence threshold
	Feature transformation function
	Nearest neighbors and maximum PageRank iterations
	PageRank damping factor
	Probability transfer matrix
	Normalized Bayesian query score
	Cosine similarity score
	Standard retrieval evaluation metrics
	Performance variation and statistical significance

The criteria used in Eq. (1) to compute the balanced precision factor (BPF) involve averaging the feature-image similarity scores and applying entropy-based weighting. This formula balances retrieval accuracy by combining the strength of similarity with the uncertainty of features.

2

In Eq. (2), a binary decision precision function is defined, in which similarity scores are thresholded with the average feature precision. The step transforms raw similarity values into retrieval-relevant indicators.

3

The information entropy and weighted distance of the feature distributions are estimated using Eq. (3). It measures feature discriminability using probabilistic entropy and deviations in similarity from the ground truth.

Input image acquisition and preprocessing

To ensure that the extracted features remain consistent, the user-entered query image is processed at this stage by scaling, normalizing, and applying noise reduction. When creating a database offline, the photos are all preprocessed using the same procedures. In Fig. 2, features for each picture are extracted from the building images and analyzed using the principal component analysis (PCA) technique and hash coding. Lastly, we combine information entropy and correlation feedback to select our final features. After that, use the mean value to set the feature weights and update them using the PageRank method. Then, the retrieval result based on the combined similarity is calculated using the specified features and final weights. Information entropy is often used to measure a system’s information content.

Fig. 2.

Framework for the proposed retrieval system.

In statistics, distributional discrimination is more significant for smaller entropies and for larger ones. Cluster weights were determined using information entropy. This study uses the probability vector as a feature weight to assess each feature’s relevance in feature fusion. The PageRank method is used to acquire the final weights, and the paper explains how the algorithm influences them. First, update the initial weights so features with greater trust may receive larger fusion weights. Then, the probability transfer matrix was built based on the retrieval trust of all features chosen as preferences. There are two benefits to updating weights using the PageRank algorithm: First, the adaptive weight eliminates the need to build the weight function manually, and second, the time complexity is drastically reduced. Algorithm 1 shows the PageRank Algorithm for image retrieval. The system’s performance is highly dependent on the factors that substantially impact the algorithm, including the probability transfer matrix M, hyper-parameter P, and starting weight W. The research fails to include its initialization technique, which significantly influences the algorithm’s convergence times and efficacy. Recreating the findings and improving the algorithm’s efficiency requires knowing how these parameters are determined, whether by random assignment, heuristic techniques, or data-driven methods. These details are crucial to ensuring the method’s reliability and dependability.

Feature extraction

This is a step in which the semantic and visual qualities of the query picture and the database images are obtained. In feature extraction, descriptors based on deep learning and handcrafted features are combined. They are then used to calculate similarity and adaptively fuse.

Algorithm 1: PageRank Algorithm for Image Retrieval

The proposed algorithm 1 will be an entropy-directed adaptive multi-feature fusion on content-based image retrieval. First, each feature type is given equal weight. Deep, handcrafted features are independently extracted using an existing feature extractor for a specific query image and a database. Similarities between query and database images in terms of features are calculated. Entropy of information is then computed on every feature to determine its discriminative ability. Average similarity scores are used to estimate retrieval trust values, which are similar to the reliability of individual features. A probability transfer matrix is constructed using these trust values, and the feature weights are optimised iteratively using an update scheme based on PageRank until convergence. Lastly, a global similarity score is obtained by summing the weighted similarities of all features. Database image ranking is based on this score, and the most relevant images are returned. This adaptive weighting approach removes the need for constant feature significance and enables competent, query-sensitive image retrieval.

By repeatedly using the probability transfer matrix rather than query results or a loss function, the PageRank method significantly reduces the computational cost of weight updates compared to other algorithms.

4

The feature trust value is calculated by Eq. (4) and serves as an indicator of a feature’s reliability in ranked retrieval results. An increase in trust values implies a reliable and valuable contribution to retrieval performance.

5

The overall score for the importance of an image in Eq. (5) is calculated using entropy normalization to transform the picture’s similarity certainty and feature relevance. This helps minimize bias from dominant features.

6

The normalized feature quality measure is determined using Eq. (6) as a combination of trust scores and relevance weights. The artificially produced value falls within a limited range to make the results comparable across features.

7

The normalized Bayesian query (nBQ) metric is identified in Eq. (7) to evaluate retrieval. It combines Bayesian relevance scores on both images and queries to give a strong performance estimate.

Overall process flow of the proposed CBIR system

An image retrieval system uses complex steps to analyze the user’s search query for pictures and return related results, as shown in Fig. 3. Initially, this system extracts features from an input image using modern methods, such as Transfer Learning with ResNet50 and VGG16. These models extract discriminating characteristics that are important for their subsequent processing. After this, the extracted features are combined at the Feature Fusion Layer using the Adaptive Multi-Feature Fusion technique, which employs Information Entropy Theory, ensuring that all parts of the image are captured and integrated correctly. Then comes the Initial Similarity Calculation stage, where distances based on Information Entropy are computed to obtain preliminary results needed for retrievals.

Fig. 3.

Process flow of content-based image retrieval.

The next one is Precision Calculation, which applies the Correlation Feedback method to boost retrieval, especially when the question requires only one attribute. Retrieval Accuracy is improved through Automatic Feature Selection (AFS), Retrieval Trust (RT), and Initial Average Weights (IAW). Final Weights Calculation for Comprehensive Similarity is where dynamic feature weight improvements culminate in a Probability Transfer Matrix and a PageRank Algorithm, respectively. This way, it computes comprehensive similarity to select the best images and effectively returns detection results that meet the query requirements. Customized to meet the demands and questions of a wide range of users, this approach guarantees accurate and robust picture retrieval.

8

The weighted feature vectors were used to calculate the similarity between the query and the database images, as in Eq. (8). Such similarity comes at the expense of angular proximity, with variations in feature magnitudes.

9

In Eq. (9), the balanced precision function is further narrowed by adding relevance thresholds to each feature dimension. This guarantees reliable feature estimation across features.

10

In Eq. (10), the score of relevance of features is estimated by weighted discrepancies between the predicted and reference similarities. It shows the contribution of every feature to a specific query.

11

The effectiveness score of the features is calculated by summing the relevance values of the images in Eq. (11). This score is used to weight similarities when fusing adaptively.

The relative importance of features remains constant across all queries for a given retrieval system. It leads to poor retrieval performance. To address this limitation, a novel multi-feature picture retrieval method is suggested. The retrieval system’s fundamental architecture is seen in Fig. 4. First, features are extracted individually during database development. Then, each feature is used to compute the entropy for each feature dimension. The features and entropies are stored to obtain the image feature database.

Fig. 4.

Multi-feature fusion based on image retrieval.

The first picture database contains a vast array of pictures. An example would be picture 1.jpg being saved in the image database, and the established image feature database being in a one-to-one connection. Picture name, feature, and entropy are the three components that make up the image feature database’s storage format. The image database referred to in this work is a database that stores picture features. During the image search phase, users input images. Initially, compute the similarity between the query and database images using each feature individually. Then, trust one feature and, by weighting several measures, get a comprehensive similarity measure, and then use this measure to output retrieval results.

12

Eq. (12) is an adaptive multi-feature fusion that applies a primary and a secondary group of features, with their effectiveness as weights. This formulation is more robust to feature imbalance.

13

Transforming features is a modeling process represented by Eq. (13), which incorporates probabilistic terms to realign the feature representation. It aids adaptive similarity estimation across different query scenarios.

14

A normalized update rule for feature weights is defined in Eq. (14), where the responses of transformed features are balanced with terms involving statistical deviation. This is done to enhance stability in feature aggregation.

15

Eq.(15) incorporates normalization factors that are inversely proportional to the control contributions of features. These are used to guarantee scalability in heterogeneous feature sets.

Adaptive multi-feature fusion model

Fig. 5 shows the process of adaptive Multi-feature fusion using a deep learning technique. Utilizing advanced methods, the image Database system effectively retrieves relevant photos in response to user inquiries. Preprocessing, which begins with a Query picture Input, may include scaling, normalization, and noise reduction to improve the picture’s quality. Using DNN models such as VGG and ResNet, Deep Learning Feature Extraction extracts high-level features for classification and identification tasks. Concurrently, Handcrafted Feature Extraction uses techniques such as Colour Histograms, SIFT, and HOG to extract textures and fine features. These characteristics are fused at the Feature Fusion stage using concatenation and an Adaptive Fusion Module based on Information Entropy theory to build complete feature representations.

Fig. 5.

Adaptive multi-feature fusion using deep learning technique.

Feature Matching uses several distance measures to determine which photos are most relevant to the query and then ranks them accordingly. The system finally displays the Retrieved Images to ensure users receive relevant results that meet their unique criteria. By integrating the benefits of deep learning with those of conventional feature extraction techniques, this methodically planned strategy guarantees reliable picture retrieval from massive image repositories.

16

Eq. (16) calculates a scaling parameter that adjusts the influence of the features via log normalization. This assists in regulating over dominance of other features.

17

In Eq. (17), a more powerful normalized Bayesian query score is derived from weighted feature subsets. It enhances consistency in retrieval between different levels of queries.

18

Eq. (18) is a composite relevance score that is calculated through the combination of structural and statistical parameters. The score is used to refine final ranking.

19

In Eq. (19), global normalization and variance adjustment of feature relevance scores are used. It provides stability in retrieval results of datasets of different sizes.

20

In Eq. (20), a scaled update term for the differential is calculated to adjust the retrieval confidence based on feature interactions. This is the last modification that makes the retrieval decision much stronger.

Hyperparameter settings and pagerank integration

The suggested CBIR framework relies on a small number of fixed hyperparameters to bring about reproducibility. The similarity graph nearest neighbor is set to K=10. Local similarity propagation and global ranking stability are achieved by setting the PageRank damping factor to 0.85. The weights of feature fusion are set to the same values and modified in an iterative manner to the point of convergence or 50 iterations. It includes page ranking to represent pictures as nodes in a weighted similarity graph, which propagates global relevance. The solution enhances ranking strength by making it less sensitive to sporadic or troublesome similarity matches.

The proposed image retrieval system adopts a comprehensive approach to handle diverse user queries effectively. Consequently, our system will finally obtain accurate retrieval results after various steps, including deep learning and conventional feature extraction, adaptive fusion, and similarity calculations. Integration of cutting-edge technology in these structured ways enhances the system’s ability to quickly and accurately retrieve relevant photos from large databases, thereby enabling optimized feature selection and weighting.

Result and discussion

Efficiently tackling big sets of images remains a challenge in modern data analysis. CBIR & DLT thus emerge as crucial mechanisms for addressing this issue. With CBIR systems, textures or colours can be used as parameters to retrieve pictures. At the same time, DLT improves feature extraction and classification accuracy by leveraging transfer learning with advanced neural networks, such as ResNet50 and VGG16. The partnership between CBIR and DLT enhances retrieval accuracy across different datasets, thereby expanding the use of image analysis in essential fields such as healthcare diagnostics, surveillance, forensic investigations, etc.

Evaluation metrics

To ensure fair and standardized performance assessment, the proposed CBIR framework is evaluated using popular retrieval metrics, including Mean Average Precision (mAP), Precision at K, Recall at K, and Normalized Discounted Cumulative Gain (NDCG). Mean Average Precision quantifies the performance of all queries, and Precision at K and Recall at K quantifies the accuracy and completeness of the top-K retrieved items. NDCG also evaluates ranking quality based on the relevance of the positions of the retrieved images. The profile of these metrics provides a comprehensive assessment of retrieval accuracy, ranking effectiveness, and robustness.

Availability of statistical significance analysis

To determine the validity of the experimental findings, statistical variation is reported as the mean performance and the standard deviation generated from a series of query runs. The average of each evaluation measure across all test queries is calculated, and the variability is used to indicate retrieval stability. Moreover, a paired t-test at the 95% confidence level is used to assess the statistical significance of the proposed approach relative to the baseline approaches. A p-value below 0.05 indicates a statistically significant improvement. This discussion supports the fact that the performance improvement realized is not a random occurrence.

Tamper resistance approaches

1. Fragile Watermarking for Tamper Detection and Localization.

One key tamper-resistance approach highlighted in the related work is fragile watermarking.

Technique that embeds fragile watermarks into images so that any modification to the image alters the watermark, enabling tamper detection and localization. This approach is effective for:

• Detecting unauthorized modifications.

• Identifying the exact region where tampering occurred.

• Preserving evidential integrity for forensic and security applications.

2. Reversible Data Hiding for Integrity Preservation.

This method embeds authentication data into images while allowing perfect recovery of the original image after extraction, ensuring:

• Integrity verification.

• Tamper evidence.

• No permanent distortion to the image.

This is especially important for sensitive domains such as medical and forensic imaging, where any alteration must be detectable and reversible.

3. Secure Encrypted Image Retrieval with Adaptive Multi-Feature Fusion.

Tamper resistance is also addressed through searchable encrypted image retrieval systems, particularly in paper. Here, tamper resistance is achieved by:

• Encrypting image features.

• Applying adaptive multi-feature fusion.

• Allowing retrieval without decrypting raw images.

This ensures that even if data is intercepted or accessed illegally, the content cannot be altered or interpreted, thereby preventing tampering and unauthorized manipulation while maintaining retrieval functionality.

4. Steganographic and CNN-Based Secure Image Protection.

The paper also discusses CNN-based steganographic frameworks that embed sensitive information into images to protect confidentiality and integrity.

These approaches contribute to tamper resistance by:

• Hiding authentication data within image pixels.

• Detecting manipulation through feature inconsistencies.

• Ensuring secure transmission in telemedicine and forensic systems.

Although retrieval accuracy is not the primary goal, these techniques provide a strong security layer against tampering and data manipulation.

Dataset description

The results show that an Unsupervised method combining multiple hashing techniques yields the best results. It will focus on hashing and auto-encoder methods, though there are many others. Finding feature vectors for each picture is possible via latent feature extraction, which involves attaching hooks to a pre-trained network and recovering the vector from earlier layers. Another method relies on autoencoders, which may extract latent characteristics from the encoder. Using Auto Encoders is the best way to obtain this data. In terms of retrieval, investigate approaches based on hashing and on Euclidean geometry. Find a unique way to quantify an image’s contents using only one integer. The input image is given in Fig. 6. Use the collected photographs’ calculated hashes to identify duplicates or nearly identical²¹. https://www.kaggle.com/datasets/sniafas/vyronas-database.

Fig. 6.

Input image.

Dataset selection and scope

The analysis is conducted on publicly available generic image datasets to assess the overall usability of the suggested CBIR framework rather than domain-specific optimization. Although the method can be used to retrieve different types of objects, the current study does not purport to cover all object domains. Rather, the selected datasets are used as benchmarks to verify feature fusion, entropy-based weighting, and ranking functionality. Using domain-specific datasets and task-oriented protocols (e.g., medical or remote sensing imagery) is recognized as a potentially valuable approach for future research to further assess the adaptability of domains.

When used together, Content-Based Image Retrieval and Deep Learning Techniques offer significant strength in addressing the challenges of large-scale image analysis, as shown by the robustness study in Eq. (16). DLT improves feature extraction and classification accuracy via transfer learning with pre-trained models such as ResNet50 and VGG16, and CBIR systems are effective at extracting images based on content similarity. By combining CBIR and DLT, the hybrid technique can learn complex image representations that are resilient to their benefits. The system adapts dynamically to changes in query situations. It uses correlation feedback and the PageRank algorithm for automatic feature selection, enhancing retrieval accuracy across datasets such as Corel1k, UC Merced Land-Use, and RSSCN7. Furthermore, it is robust and can be used in complex areas such as fingerprint identification and crime prevention, as well as in digital libraries and historical research that require high precision or adaptability. In our proposed method, CBIR shows a robustness ratio increase of 97%, while DLT shows a reduction of 95.34%, as shown in Fig. 7a, b.

Fig. 7.

(a) Analysis of robustness based on CBIR. (b) Analysis of robustness based on DLT.

It is essential to develop accurate image retrieval methodologies that combine DLT and CBIR to address the challenge posed by extensive image datasets, as shown in Eq. (17). To facilitate content-based matching, such systems assess how well query images match database images in terms of colour, texture, and shape. Over time, deep learning techniques (DLT) using transfer learning mechanisms have made strides towards this, with designs like ResNet50 and VGG16 considered neural network experts in extracting features and classification tasks. Combining these two will yield better results because CBIR excels at handling various aspects of an image, such as colour and texture, whereas DLT learns abstract descriptions of images. This approach makes the system more adaptable across different query contexts, thereby enhancing its effectiveness across varied domains through adaptive multi-feature fusion methods and automatic feature selection schemes. It also expands the applicability of image retrieval systems so they can be used when examining images closely in areas such as industrial quality control, medical diagnostics, and satellite imagery interpretation, where one must carefully analyze images, rather than relying solely on textual information without much thought. The method has led to a 96% increase in accurate retrieval rate for CBIR, 95.3% for DLT, and 96.2% accuracy in terms of time, as shown in Fig. 8a–c.

Fig. 8.

(a) Analysis of accurate image retrieval based on CBIR. (b) Analysis of accurate image retrieval based on DLT. (c) Analysis of accuracy in terms of time.

As inferred from Fig. 9a, b, this is evident in Eq. 18, which compares the performance of DLT and CBIR to show how both improve image retrieval systems. The best images to match those in a query are identified using CBIR techniques that rely on properties such as shape, texture, and colour. Besides, transfer learning has improved this process by leveraging more sophisticated neural network architectures such as ResNet50 and VGG16, which perform well at feature extraction and classification tasks by leveraging features learned from other domains. Integration of CBIR with DLT enhances performance metrics such as retrieval speed, accuracy, and scalability across large-scale collections or databases. Thus, the method improves retrieval accuracy by dynamically adapting to query difficulties through automated feature selection processes and multi-feature fusion algorithms. As such, medical imaging, satellite analysis, and multimedia content management could be fostered for surveillance purposes through an integrated approach because DLT’s robustness over various image types and ability to learn intricate image representations are increased. This presents a study on performance research showing that the CBIR ratio improves by 96.62%, while that of DLT improves by 94.32%.

Fig. 9.

(a) Analysis of performance based on CBIR. (b) Analysis of performance based on DLT.

Eq. (19) shows that integrating DLT with CBIR significantly improves fingerprint recognition, which is vital for security and forensic applications. CBIRs are an effective means of analyzing fingerprint images for specific ridge patterns, minutiae points, and other distinct traits, enabling matching against vast databases. Therefore, due to DLT’s use of CNNs, transfer learning, and excellent feature extraction, transfer learning and classification approaches are applicable here. The union between CBIR and DLT ensures high performance across multiple datasets and environments in fingerprinting identification.

Moreover, the strength lies in CBIR’s ability to handle multimodal traits and variations in fingerprint photos, complemented by DLT, which can learn complex patterns and boost classification accuracy. It also involves automated feature extraction and selection algorithms to improve system efficiency by increasing identification rates whilst decreasing false-positive rates. One example is this collaborative approach, which enhances reliability in criminal investigations, border controls, and access management systems, thereby expediting identification processes. However, by leveraging advances in deep learning and image retrieval, fingerprint identification systems can now meet these stringent requirements for forensic or security purposes, thereby improving operational efficiency and achieving higher accuracy. This existing method provides a proposed solution for CBIR fingerprint identification at 95.11 and an improvement over the current method of 93% using DLT, as inferred from Fig. 10a, b.

Fig. 10.

(a) Analysis of fingerprint identification based on CBIR. (b) Analysis of fingerprint identification based on DLT.

As inferred from Fig. 11a, b, using DLT in conjunction with CBIR may significantly improve surveillance and law enforcement systems, thereby reducing crime, as expressed in Eq. 20. Computer-based image recognition systems analyze surveillance footage to identify people, cars, and other objects based on visual characteristics such as appearance, clothing, or license plates. Law enforcement organizations can quickly access pertinent information from extensive video databases using these tools. By leveraging modern neural networks, such as CNNs, DLT improves CBIR’s feature extraction and classification capabilities. This integration enables real-time recognition of CCTV video anomalies, suspicious activities, or illegal access. DLT’s ability to learn complex patterns and adapt to diverse environmental conditions supports reactive crime-prevention systems. CBIR’s robust picture retrieval capabilities, when integrated with DLT’s advanced learning algorithms, can enable better oversight of public spaces, earlier detection of criminal activity, and greater public safety. Consequently, contemporary security operations and law enforcement organizations rely on these technologies because they enable proactive intervention and real-time, data-driven decision-making. The crime-prevention ratio increases by 97.2% in CBIR and DLT, and by 96.2% in the proposed method.

Fig. 11.

(a) Analysis of crime prevention based on CBIR. (b) Analysis of crime prevention based on DLT.

Table 3; Fig. 12 show the Overall Performance Analysis Comparison. The table and shaded line graph provide a comparative analysis of the suggested CBIR and DLT methods with BDIA, TLBA, SVM, and DM across various performance indicators. The findings show that the offered approach is more reliable, more precise, and achieves better performance. The curves in the shaded region show the same performance trends over time, demonstrating the effectiveness and reliability of the presented framework compared with existing methods. The combination of sophisticated neural networks such as ResNet50 and VGG16 has improved the accuracy and flexibility of image analysis when CBIR is blended with DLT. However, CBIR excels at image retrieval using content features like texture and colour, while feature extraction and classification, enabled by transfer learning, make DLT great at it. These methods ensure high performance across various dataset types and applications, including medical diagnosis, surveillance, and forensic investigation inquiries. This hybrid solution improves retrieval accuracy, supports proactive decision-making, and streamlines data administration in current data-intensive environments by combining the best features of both techniques.

Table 3.

Overall performance analysis comparison.

Metrics	CBIR and DLT (%)	BDIA	TLBA	SVM	DM
Robustness	96.17	92.23	91.36	93.23	94.56
Accuracy	95.65	90.35	92.45	91.56	91.54
Performance Ratio	95.47	93.48	91.76	90.28	92.45
Fingerprint Identification Ratio	94.05	88.56	90.78	92.46	95.47
Crime Prevention Ratio	96.7	91.52	93.46	89.21	92.85

Fig. 12.

Performance comparison curves.

Fig. 13 compares the performance of the proposed method with recent JESTECH 2025 works. The graph clearly shows that the proposed framework achieves higher accuracy, robustness, and tamper detection rates. This visual evidence confirms the superiority of the entropy-guided adaptive fusion with PageRank optimization, demonstrating its effectiveness for secure, reliable, and scalable image retrieval applications. It provides a comparative performance study of the proposed adaptive multi-feature fusion using an entropy-guided framework, along with representatives of recent methods in CBIR, i.e., BDIA¹⁶, TLBA¹⁷, SVM-based adaptive fusion¹⁸, and Data Mining-based feature selection techniques¹⁹, as summarized in Table 3. Such techniques have been chosen because they reflect the state-of-the-art in adaptive feature fusion, optimization-based retrieval, and learning-based CBIR systems. Nevertheless, the exact references on which each baseline was based were not stated directly in the original figure title, which has since been clarified to make the figure complete and reproducible.

Fig. 13.

Pictorial illustration of performance comparison.

Proposed methodology advantages and limitations

The proposed content-based image retrieval architecture has several benefits. First, a blend of deep learning and handcrafted features enables the representation of visual and semantic data at both low and high levels in a more comprehensive manner. Secondly, information entropy enables more credible similarity calculations by efficiently determining feature discriminability. Third, the use of correlation feedback and the PageRank technique in adaptive feature weighting enhances retrieval performance across diverse queries, as it does not require constant or manually specified weights. It is also easy to scale to larger datasets since the modular, multi-stage design is also more scalable.

The proposed design is good, but it has its limitations. The contributory factors to the computational complexity of the offline database development stage include multiple feature extraction methods and adaptive optimization. When not fine-tuned on domain-specific data, using pre-trained deep learning models can lead to limited performance. Finally, very large-scale picture repositories can also incur additional processing costs due to the iterative PageRank-based weight optimization.

Conclusion

This paper proposed an entropy-based adaptive multi-feature fusion architecture to enable reliable content-based image retrieval by addressing the weaknesses of traditional CBIR systems, including their non-adaptive feature weighting, limited resilience, and inability to handle diverse image datasets. Using deep learning-based feature extraction as the main objective, it was demonstrated that dynamic feature selection and weighting can significantly improve retrieval. Information entropy, correlation feedback, and PageRank-based optimization lead this. The proposed strategy will combine deep features from the ResNet50 and VGG16 transfer learning models and bespoke visual features to ensure comprehensive picture representation. This approach enables convergence uniformity and reduces computational cost because it uses a probability transfer matrix to automatically evaluate feature trust and variability and update fusion weights without manual parameter control. Experimental validation of the proposed method on Corel1k, UC Merced Land-Use, and RSSCN7 datasets shows that it routinely outperforms its current CBIR counterparts in terms of performance ratio, robustness, and retrieval accuracy. The retrieval accuracy was 96%, the strength was over 97%, and the system was reliable in its use for applications such as fingerprint identification and criminal prevention. We want to explore a lighter design in the future that can further reduce computing overhead without sacrificing retrieval accuracy, optimize parameter selection, and extend the framework to real-time retrieval systems.

Limitations.

Although the proposed system achieves high accuracy and robustness, it suffers from increased computational complexity due to multi-feature extraction, entropy computation, and PageRank-based weight optimization. Reliance on pre-trained deep models may limit generalization across specialized domains, and iterative fusion can introduce latency in very large-scale databases, restricting real-time applicability in resource-constrained environments.

Future Work.

Future studies will focus on developing lightweight and energy-efficient fusion strategies to reduce computational overhead while preserving accuracy. Domain-specific fine-tuning and self-supervised learning will be explored to improve generalization. In addition, explicit tamper-detection and self-recovery modules will be integrated, and real-time deployment on edge and cloud platforms will be investigated for secure, large-scale applications.

Author contributions

Conceptualization: VAC., RG.; Methodology: LM.,.; Formal analysis and investigation: VG, RG.; Writing - original draft preparation: LM. Writing - review and editing: LM, RG. and VAC; Supervision: LM, and RG; Figs. 1, 2, 3, 4 and 5 preparation: LM, RG. and VG.

Funding

The authors received no specific funding for this study.

Data availability

The data used in this research is available at the following linkhttps://www.kaggle.com/datasets/sniafas/vyronas-database.

Declarations

Competing interests

The authors declare no competing interests.