Intelligent high-throughput recognition model for bitter gourd fruit morphology and tubercle characteristics

Abstract

Bitter gourd, an important crop with both economic and medicinal value, requires precise identification of fruit shape and tubercle phenotypes to enhance breeding efficiency. To address the low efficiency and high subjectivity of traditional methods, this study proposes an improved YOLOv8-CEFC model for high-throughput automatic detection of the bitter gourd fruit shape and tubercle characteristics. First, the model integrates the ConvNeXt V2 module into the backbone network, combined with a Fully Convolutional Masked Autoencoder (FCMAE) framework and Global Response Normalization (GRN) layers to enhance feature extraction capabilities. Second, an Efficient Multi-scale Attention (EMA) mechanism is introduced, capturing local tubercle textures and global fruit shape contours simultaneously through a parallel dual-branch structure, while also improving the model’s robustness against cluttered backgrounds and environmental noise. Finally, Focal-CIoU Loss is incorporated to replace CIoU Loss, reducing the impact of class imbalance on model accuracy. The results show that the model achieves precision, recall, mAP50, mAP50-95, and F1 scores of 93.9%, 94.4%, 96.3%, 93.6%, and 94.15%, respectively, which represent improvements of 2.0%, 3.5%, 1.1%, 3.4%, and 2.75% compared to the original YOLOv8n model. The performance gain of the model was further examined using the bootstrap method, which confirmed that the improvement is statistically significant. Further validation through confusion matrix analysis, PR curves, and ablation experiments confirms the effectiveness of the improvements. Compared to other mainstream YOLO models, YOLOv8-CEFC demonstrates more accurate identification, better stability, and higher detection efficiency. The proposed improved YOLOv8-CEFC model provides an efficient solution for phenotypic analysis in Bitter Gourd breeding and holds significant importance for advancing the intelligentization of crop breeding.

Introduction

Bitter gourd (Momordica charantia L.) is a climbing herbaceous plant in the gourd family (Cucurbitaceae), widely cultivated in regions of Asia, Africa, and South America. Bitter gourd has a unique flavor and is rich in nutrients, including high levels of vitamin C, amino acids, and minerals. It is traditionally used for its cooling, detoxifying, and vision-enhancing effects¹. The germplasm resources of bitter gourd display remarkable phenotypic diversity: fruit shapes range from fusiform and conical to ovoid, while surface tubercles exhibit complex variations in morphology—from shallow to deep protrusions, sparse to dense distributions, and ridged to granular configurations. This rich phenotypic diversity not only provides a valuable genetic foundation for selecting superior genotypes but also directly correlates with commercial value, market acceptability, and key nutritional quality^2,3. However, in traditional bitter gourd breeding processes, the identification and quantification of these key traits, such as fruit shape and tubercle traits, heavily depend on manual measurements and experience-based judgments. These phenotypic analysis methods are generally limited in scale, inefficient, and prone to significant errors⁴, which restricts the development of plant functional genomics and molecular design breeding⁵.

Phenotypic traits, as external manifestations of genotypes^6,7, are fundamental for germplasm resource research, classification, and precise breeding. Therefore, it is crucial to perform precise and efficient phenotypic identification for a large number of samples to screen for superior germplasm and facilitate resource conservation⁸. To overcome the limitations of traditional phenotypic analysis methods, the research and development of intelligent crop phenotypic identification technology is of great importance. This technology, based on morphological features, can provide efficient and standardized data support for germplasm classification and parent screening, thereby accelerating the intelligent management of germplasm resources. In recent years, with the development of computer vision technology, particularly target detection methods based on traditional image processing and deep learning⁹, new solutions have been provided for this field.

Traditional image processing methods mainly rely on features such as color, shape, and texture, combined with operations like segmentation, clustering, enhancement, and morphology to achieve target detection. In the area of feature and segmentation algorithm optimization, Luo et al.¹⁰ improved the clustering segmentation by combining point-line distance constraints to accurately locate the complex color and contour of grape fruit stem picking points. Cheng et al.¹¹ proposed the M-LP feature-weighted clustering method, which integrates Mahalanobis distance and shadow position features to improve cherry tree canopy segmentation accuracy. Kurtulmus et al.¹² innovatively combined color, circular Gabor texture, and eigenfruit features to achieve effective recognition (detection rate of 75.3%) of green citrus under natural light conditions. Rahimi et al.¹³ further applied an ANN classifier to enhance the accuracy of pineapple canopy recognition and counting. In terms of improving environmental robustness, Rabab et al.¹⁴ utilized a center-point radius model combined with NDVI threshold segmentation and overlapping pixel row adjustment to successfully extract overlapping perennial ryegrass plants in the field. Hannan et al.¹⁵ designed an adaptive chroma segmentation and perimeter fitting model, combined with dynamic thresholding and sliding window circumference detection, to effectively identify citrus in orchards under varying light conditions and overlapping conditions. Vitzrabin et al.¹⁶ improved the detection accuracy of red bell peppers in natural light by integrating an adaptive threshold algorithm from RGBD sensors with dynamic zoning and morphological optimization. Despite progress made by traditional image processing methods through algorithm optimization and environmental adaptability improvements, their feature extraction capability and recognition accuracy under complex environmental conditions still face challenges.

The rapid development of deep learning in the field of computer vision has provided new methods for plant phenotypic detection. Deep learning methods can autonomously learn and integrate multi-level features from images, greatly improving the accuracy of crop phenotypic recognition¹⁷. Target detection algorithms based on Convolutional Neural Network (CNN) have demonstrated outstanding performance in crop recognition tasks, with applications across several areas. For instance, in high-precision instance segmentation, Neupane et al.¹⁸ combined the Azure Kinect depth sensor with Mask R-CNN to achieve high-precision segmentation of mangoes, with an accuracy of 95.6% and a length measurement error of only 4.7 mm. Huang et al.¹⁹ proposed a deep learning-based segmental segmentation and morphological restoration method, which effectively addressed the occlusion and overlap issues in citrus fruits, achieving a 93.66% fruit recognition rate and 96.30% segmentation accuracy. In the field of target detection and recognition, Chen et al.²⁰ used HSV color space preprocessing combined with CNN to improve the accuracy, recall rate, and generalization ability of chili pest recognition. Li et al.²¹ built a rapeseed seedling detection system based on an improved YC-YOLOv7 model, achieving a recognition accuracy of over 90%. Hou et al.²² designed a network specifically for small leaf tip target recognition, achieving 91.6% precision and 85.1% mAP50. Additionally, Ji et al.²³ employed YOLOv5 and the lightweight network PP-LCNet to reconstruct the feature extraction module, significantly reducing the model’s parameter count and achieving lightweight weed recognition. For target counting tasks, Xu et al.²⁴ used Mask R-CNN and YOLOv5 to detect and count maize seedling leaves in complex field environments, achieving counting accuracy rates of 69.4% and 72.9% for fully unfolded and newly emerged leaves, respectively. Xie et al.²⁵ proposed a CNN-based deep regression model that improved leaf counting accuracy while reducing training costs. In classification and maturity evaluation, deep transfer learning techniques have shown good results, such as the use of the VGG19 model by DANH et al.²⁶ for classifying cherry tomato maturity, with an accuracy of 94.14%. However, when focusing on bitter gourd as a specific crop, relevant deep learning research is still scarce and its application is limited. Current research primarily focuses on disease recognition and health status assessment, such as Villanueva et al.‘s²⁷ application of CNN algorithms with Keras, TensorFlow, and Python for predicting the health status of bitter gourd. Li et al.²⁸ improved Faster R-CNN and incorporated a Feature Pyramid Network (FPN), enhancing the mAP for bitter gourd disease detection to 86.39%, with a 16.56% improvement in the detection accuracy of gray spot disease. These studies show that current deep learning algorithms have relatively limited applications in the automated extraction and analysis of key phenotypic traits directly related to bitter gourd breeding. Therefore, to address the lack of rapid and automated phenotypic acquisition methods in bitter gourd breeding, this study focuses on the fruit shape and surface tubercle characteristics of bitter gourd. The YOLOv8 model is improved to enhance the accuracy of fruit feature detection. The main contributions are as follows:

1. First, a detailed morphological classification and data annotation of bitter gourd fruit based on shape and tubercle characteristics is conducted, and a bitter gourd image dataset is established.

2. ConvNeXt V2 is integrated into the backbone network, and the introduction of a Fully Convolutional Mask Autoencoder (FCMAE) and Global Response Normalization (GRN) layers enhances the feature competition ability between channels and improves the model’s feature capturing capability.

3. The EMA attention mechanism is embedded before the Spatial Pyramid Pooling-Fast(SPPF) layer to simultaneously capture local and global contextual information, effectively aggregating short-range and long-range dependencies. This improves the learning of attention relationships between network channels while maintaining computational efficiency and reducing false negatives and false positives.

4. A dynamic focal weighting mechanism from Focal Loss is introduced into the CIoU loss function to address the challenge of class imbalance when identifying difficult samples. By combining Focal Loss with CIoU, the Focal-CIoU loss function is introduced to effectively mitigate the impact of sample imbalance on detection results, enhancing the overall performance of the model.

Materials and methods

Data collection and dataset construction

Data collection

Images of bitter gourd used in this study were collected from the vegetable fields of the New Variety Demonstration and Evaluation Base (Langqi) of the Fujian Seed Industry Innovation Center, as well as from the Minhou breeding farm of Fuzhou Tianmei Seedling Technology Co., Ltd. To ensure broad applicability of the experiments, we selected multiple bitter gourd cultivars with rich phenotypic diversity in fruit traits, including ‘Yuchuan 2’, ‘Qisheng 308’, and ‘Tianmei 20’. The collected samples cover a wide range of variation in both fruit shape and tubercle characteristics. Fruit shapes include conical (enlarged at the apex and gradually tapering toward the base), ovoid (swollen in the middle with an overall nearly spherical appearance), fusiform (narrow at both ends and enlarged in the middle), and rod-shaped (approximately cylindrical). Tubercles range from granular tubercles and ridge-like tubercles to mixed patterns such as many tubercles with few ridges and few tubercles with many ridges. Representative combinations include ovoid fruits with granular tubercles (Fig. 1a), fusiform fruits with few tubercles and many ridges (Fig. 1b), fusiform fruits with few tubercles and many ridges (Fig. 1c), and rod-shaped fruits with ridge-like tubercles (Fig. 1d).

Fig. 1.

Bitter gourd images with different phenotypic features. (a) Ovoid shape + granular tubercles; (b) Fusiform shape + few granules, many strips; (c) Conical shape + granular tubercles; (d) Rod-shaped + strip-like tubercles.

All images were acquired under natural light between 9:00–12:00 and 14:00–17:00, covering various weather conditions such as sunny and cloudy days to enhance data diversity. Image acquisition was performed using multiple devices, including an iPhone 13, iPhone 15, and HUAWEI Mate60 Pro+. During shooting, camera parameters were harmonized across devices to minimize the impact of hardware differences on image quality. A random shooting strategy was adopted without constraining the shooting angle or illumination level, and the shooting distance was maintained between 0.3 m and 1 m. In total, 1,716 original bitter gourd images were obtained.

To ensure the robustness of the model in complex field scenarios, the image acquisition and data augmentation strategies in this study were specifically designed to account for issues commonly encountered during real-world deployment: (1) bitter gourd fruits partially occluded by vines (Fig. 2a); (2) images with naturally overlapping fruits (Fig. 2b); (3) slight motion blur caused by hand-held image capture (Fig. 2c); and (4) highly variable illumination conditions in the field (Fig. 2d). These challenging scenarios are ubiquitous under natural cultivation conditions and were deliberately incorporated into both the dataset and augmentation pipeline so that the trained model can maintain stable recognition performance under real-world conditions.

Fig. 2.

Representative bitter gourd images under typical conditions encountered in real-world deployment: (a) occlusion by vines; (b) natural overlap of fruits; (c) motion blur caused by hand-held imaging; and (d) variable illumination in the field.

Dataset construction

Data augmentation is a key technique for improving model performance. By increasing the diversity of training data, it enhances the model’s generalization ability, leading to a significant optimization of the overall model performance²⁹. Considering the potential for overfitting during the model training process, and to improve the model’s generalization ability and robustness, various techniques such as mirroring, brightness and contrast adjustment, Gaussian blur, random translation, and image stitching were randomly combined^30–33. These techniques were applied to expand the original dataset to a total of 5190 images.

In this study, based on the differences in bitter gourd fruit shape and tubercle characteristics, and after thoroughly analyzing the collected images, 16 categories were defined. These categories include label names, specific descriptions, and sample counts, as shown in Table 1. However, dueto the limited sample sizes of classes LR, LWMRL, LWLRM, and FR, these classeswere excluded from subsequent model training and testing. Bitter gourd images were manually annotated using the Labeling tool (v1.8.1). The annotation was independently performed by two horticultural specialists who had received standardized training in label definitions and discrimination criteria. To ensure labeling consistency, we implemented a cross-check quality control strategy whereby, after completing their annotations independently, the two annotators thoroughly reviewed each other’s labels to minimize individual subjective bias. All inconsistencies were resolved through expert consultation to reach a final consensus. During the annotation process, the “single target, single box” principle was strictly followed, ensuring the complete bounding of target areas. A total of 13,163 high-quality target samples were obtained. The annotated images were randomly shuffled, and the dataset was then split into training, validation, and test sets in a 7:2:1 ratio. This ensured that both the training and test sets contained a balanced distribution of various sample types.

Table 1.

Description of bitter gourd detection labels.

ID	Labels name	Descriptions
0	YW	Conical shape + Granular tubercles.
1	YR	Conical shape + Strip-like tubercles.
2	YWMRL	Conical shape + Many granular and few strip-like tubercles.
3	YWLRM	Conical shape + Many strip-like and few granular tubercles.
4	LW	Ovoid shape + Granular tubercles
5	LR	Ovoid shape + Strip-like tubercles
6	LWMRL	Ovoid shape + Many granular and few strip-like tubercles.
7	LWLRM	Ovoid shape + Many strip-like and few granular tubercles.
8	FW	Fusiform shape + Granular tubercles
9	FR	Fusiform shape + Strip-like tubercles.
10	FWMRL	Fusiform shape + Many granular and few strip-like tubercles.
11	FWLRM	Fusiform shape + Many strip-like and few granular tubercles.
12	BW	Rod-shaped shape + Granular tubercles
13	BR	Rod-shaped shape + Strip-like tubercles.
14	BWMRL	Rod-shaped shape + Many granular and few strip-like tubercles
15	BWLRM	Rod-shaped shape + Many strip-like and few granular tubercles.

To quantify the degree of class imbalance, we counted the number of samples for each class in the training set (Fig. 3). we computed the imbalance ratio (IR), defined as the ratio between the largest and smallest class sizes in each split³⁴. For our training, validation, and test sets, the IR values are 1.71, 1.63, and 1.88, respectively, all below 2. According to commonly used taxonomies, datasets with 1.15 < IR ≤ 3.5 are considered mildly imbalanced, whereas only IR > 3.5–9 is regarded as highly or extremely imbalanced³⁵. Therefore, our dataset can be regarded as approximately balanced with only mild class skew rather than a severely imbalanced setting. This level of skew is not sufficient to substantially interfere with model training or to cause instability in the optimization process.

Fig. 3.

Training Samples per Class. Note: Class0 stands for YW; Class1 stands for YR; Class2 stands for YWMRL; Class3 stands for YWLRM; Class4 stands for LW; Class5 stands for FW; Class6 stands for FWMRL; Class7 stands for FWLRM; Class8 stands for BW; Class9 stands for BR; Class10 stands for BWMRL; Class11 stands for BWLRM.

YOLOv8-CEFC model

To efficiently detect the fruit shape and tubercle characteristics of bitter gourd and improve the accuracy and robustness of phenotypic feature recognition, this study proposes an improved YOLOv8-CEFC network structure based on YOLOv8. In the improved backbone network, the ConvNeXt V2 module is introduced. By combining the Fully Convolutional Mask Autoencoder (FCMAE) framework and Global Response Normalization (GRN) layers, feature competition between channels is enhanced, further improving the model’s performance in various recognition tasks. Additionally, an Efficient Multiscale Attention (EMA) module is cascaded before the Pyramid Pooling (SPPF) layer. This dual-path fusion mechanism, which involves cross-channel interaction and spatial attention, significantly strengthens the feature response to tubercle textures. In the loss function improvement, Focal-CIoU is used to replace the original CIoU loss function. The dynamic focal weighting mechanism reduces the impact of class imbalance on model accuracy. The improved network structure is shown in Fig. 4.

Fig. 4.

Improved network structure diagram.

ConvNeXt V2 backbone structure

To enable the backbone network to capture more refined bitter gourd features and improve the accuracy and robustness of phenotypic detection, this study introduces the ConvNeXt V2 module proposed by Sanghyun Woo et al. ConvNeXt V2 combines self-supervised learning frameworks with architectural improvements. By incorporating the Fully Convolutional Mask Autoencoder (FCMAE) and Global Response Normalization (GRN) layers, it enhances feature competition between channels³⁶. As shown in Fig. 5, ConvNeXt V2 applies random masking to specific regions of the bitter gourd image and utilizes sparse convolution techniques to process the masked areas and predict the missing portions. This approach effectively captures detailed information from the bitter gourd images, significantly enhancing the model’s ability to learn and understand visual data³⁷.

Fig. 5.

FCMAE: fully convolutional masked autoencoder.

Furthermore, ConvNeXt V2 removes the LayerScale layer from ConvNeXt V1 and adds a Global Response Normalization (GRN) layer (Fig. 6). The GRN layer strengthens the competition between feature channels, promoting feature diversity across different channels and preventing feature collapse³⁸. This improvement enables the model to better differentiate subtle differences between bitter gourd varieties when dealing with complex backgrounds, such as varying lighting and occlusion, significantly enhancing the model’s recognition accuracy and generalization ability.

Fig. 6.

ConvNeXt block designs.

The computation process of the GRN layer is divided into three steps. First, global aggregation is performed on the features of each channel. Second, the aggregated features of each channel are normalized, ensuring that each channel’s features have relative importance across the entire feature map, which enhances competition between channels. Finally, the normalized results are used to adjust the original feature responses, helping the network retain more feature diversity. Assuming the input features are denoted as, where C is the number of channels and H and W are the spatial dimensions of the feature map, the specific computation formula is as follows:

1

2

3

In the formula, represents the global feature aggregation result of the i-th channel; is the normalized feature response, indicating the relative importance of the i-th channel; is the bias term, used to further adjust the feature response.

EMA attention mechanism

To maintain computational efficiency while enhancing the model’s ability to capture key target features in complex backgrounds and reduce the occurrence of false negatives and false positives, this study introduces the EMA (Efficient Multi-Scale Attention) mechanism proposed by Daliang Ouyang et al.³⁹. By preserving both global and local contextual information, it precisely models channel and spatial information, thereby increasing the network’s focus on target channel features, highlighting the positional information of the target, and reducing the interference of background noise. This allows the model to more accurately detect various details of the bitter gourd phenotype. The structure of the EMA attention mechanism is shown in Fig. 7.

Fig. 7.

Efficient Multi-Scale attention.

Where c is the number of input feature channels, and h and w are the dimensions of the input features in the vertical and horizontal directions, respectively. First, the input channels are divided into g groups, with each group responsible for learning different semantic information. Then, feature extraction is performed through three branches: two parallel branches perform average pooling in the vertical and horizontal directions, respectively, and the resulting feature maps are concatenated. A single 1 × 1 convolution kernel is shared, and matrix multiplication is used to aggregate the two channel attention maps within each group, enabling cross-dimensional feature interaction. The third branch uses a 3 × 3 convolution to capture local cross-channel interactions and expand the feature space. Finally, EMA extracts the dependencies between the three paths through a cross-space information aggregation method. Specifically, EMA introduces two tensors, the outputs of the 1 × 1 and 3 × 3 convolution branches. Global average pooling is applied to encode the global spatial information of the 1 × 1 branch output. The specific formula for this pooling operation is as follows:

4

In the formula: represents the pooled output feature map; is the value of the feature element at position on the C-th channel.

Focal-CIoU loss function

The YOLOv8 object detection model uses the CIoU Loss as the bounding box regression loss function. CIoU Loss provides superior performance over traditional IoU and GIoU in terms of bounding box regression accuracy by considering factors such as the distance between the center points, overlap area, and aspect ratio of the predicted and ground truth boxes. However, CIoU Loss assigns equal weight to all samples when processing training data, neglecting differences in sample quality. This leads to the model struggling to focus on and optimize learning from difficult samples. To address the issue of distinguishing sample quality in CIoU Loss, this study introduces the Focal Loss⁴⁰, which is an improvement based on standard Cross-Entropy. Focal Loss dynamically suppresses the loss contribution of easy samples while amplifying the loss weight of hard samples, thereby addressing the class imbalance issue that leads to model training bias. The specific formula is as follows:

5

In the formula, is used to balance the weights of positive and negative samples; is the model’s predicted probability for the true class, ranging from [0,1]; γ ≥ 0 is the focusing parameter that adjusts the loss contribution of easy-to-classify samples. The larger the value of γ, the lower the loss contribution from easy samples, allowing the model to focus more on hard-to-classify samples.

By combining CIoU loss with Focal loss, the Focal-CIoU Loss is proposed. Focal-CIoU Loss adjusts the form of the loss function from a gradient perspective, applying different degrees of penalty to high-quality and low-quality anchor boxes⁴¹. In this way, even with a large number of low-quality anchor boxes, their impact on the overall training process can be effectively mitigated, while high-quality anchor boxes receive more attention and optimization. Through this approach, Focal-CIoU Loss better balances the impact of anchor boxes of varying quality, thereby improving the performance of object detection. The formula for the Focal-CIoU Loss function is as follows:

6

7

8

9

In the formula, v is used to measure the consistency of the aspect ratio between the predicted box and the target box; α is a balancing parameter that dynamically adjusts the influence of v in the loss function; x and y are the horizontal and vertical coordinates of the predicted box center, respectively; w and h are the width and height of the predicted box, respectively; and are the horizontal and vertical coordinates of the ground truth box center, respectively; and are the width and height of the ground truth box, respectively; wg and hg are the width and height of the minimal enclosing rectangle of the predicted and ground truth boxes, respectively; γ is used to control the weight between the classification loss and the regression loss.

Evaluation metrics

This study uses Precision, Recall, F1-Score, mAP50, and mAP50-95 as evaluation metrics for the model. The model’s Precision is the proportion of correct predictions among all the predictions made by the model; Recall is the proportion of correctly predicted positive samples among all positive samples; F1-Score is the harmonic mean of Precision and Recall, providing a comprehensive evaluation of the model’s balance between positive and negative samples; mAP50 is the mean Average Precision (mAP) calculated at an IoU threshold of 0.5, while mAP50-95 is the average of multiple mAPs computed across IoU thresholds from 0.5 to 0.95. The specific formulas are as follows:

10

11

12

13

14

Where TP (True Positive) represents the number of actual positive samples that were predicted as positive; FN (False Negative) represents the number of actual positive samples that were predicted as negative; K is the total number of classes; AP represents the area under the Precision-Recall curve; and mAP refers to the mean Average Precision.

To assess the statistical reliability of the performance improvement of the proposed model over the baseline, we employed a non-parametric bootstrap procedure on the test set. Specifically, in each bootstrap iteration, we performed sampling with replacement on the test set to construct a resampled test subset. Both the baseline model and the improved model were then evaluated on this subset, and the corresponding mAP50 and mAP50–95 metrics were computed. For each iteration, we calculated the performance difference Δ between the improved model and the baseline model.

15

Based on the empirical distribution of Δ, we report its mean, standard deviation, and the 95% bootstrap confidence interval defined by the 2.5th and 97.5th percentiles. When the entire confidence interval lies above 0, the performance gain can be regarded as statistically significant.

Experimental environment and results analysis

Experimental environment

All training and testing in this study were conducted on the same server. The hardware and software environment is shown in Table 2.

Table 2.

Model training Environment.

Configure	Parameters
CPU	Intel Coro i7-13700 K CPU
Random access memory (RAM)	32GB
GPUs	GeForce RTX 4070 Ti
Display memory	12GB
Training environment	CUDA 11.3
Operating system	Windows10(64-bit)
Development environment (computer)	PyTorch 1.2.0 Python 3.8.10

In this experiment, the input image resolution was set to 640 × 640. The initial learning rate was set to 0.001, with a momentum decay strategy and a momentum value of 0.937. Additionally, the weight decay factor was set to 0.0005, and the batch size for each training session was 32.

Model training results

Performance of the model before and after improvement

This study conducted a comprehensive comparison and analysis of the overall performance of the improved YOLOv8-CEFC model. The results show a significant performance difference between the model before and after the improvement. As shown in Fig. 8, the YOLOv8-CEFC model achieved Precision, Recall, mAP50, mAP50-95, and F1-Score values of 93.9%, 94.4%, 96.3%, 93.6%, and 94.15%, respectively. Compared to the original YOLOv8n model, the improvements in each metric were 2.0%, 3.5%, 1.1%, 3.4%, and 2.75%. The improved YOLOv8-CEFC significantly enhances detection accuracy and generalization ability.

Fig. 8.

Performance of the model before and after improvement.

Confusion matrix analysis

The confusion matrix is a tool used to evaluate the performance of classification models, allowing for the assessment of the model’s performance across different categories⁴². In this study, a performance comparison of the models before and after the improvement was made using the confusion matrix (Fig. 9). It can be observed that both the pre-improvement and post-improvement models performed excellently in recognizing basic fruit shapes (conical/ovoid/fusiform/rod-shaped) and single tubercle textures (pure granular/pure strip-like). However, when faced with complex textures that involve a mixture of granular tubercles and strip-like tubercles with proportional differences, the recognition accuracy of the pre-improvement model significantly decreased. In contrast, the improved YOLOv8-CEFC model showed significant advantages in recognizing bitter gourd phenotype features. Specifically, the model maintained more than 95% recognition accuracy for the YW, YR, LW, FW, BW, and BR metrics, and performed significantly better than the pre-improvement model in recognizing the complex features of YWLRM, FWMRL, and BWLRM, with improvements of 1%, 2%, and 4%, respectively. Notably, the recognition accuracy for BWMRL increased from 76 to 82%. The improved model greatly reduced misclassification between fruit shape categories and provided more accurate recognition of subtle features such as the density and distribution of tubercles.

Fig. 9.

Confusion Matrix Comparison of Models Before and After Improvement.

Although the YOLOv8-CEFC model exhibits strong overall performance, we further analyzed its failure cases on the test set. The main sources of error include: (1) severe occlusion: when more than 50% of the bitter gourd fruit is covered by dense leaves or vines, the model is prone to missed detections; (2) extreme close-up imaging: when the shooting distance is excessively short, local features of the fruit are overly magnified while global contour information is lost, which can lead to misjudgment of fruit shape; and (3) extreme illumination conditions: in a small number of images captured under strong backlighting or deep shadows, the contrast of tubercle textures is reduced, resulting in incorrect tubercle-type classification. These cases highlight directions for future model optimization, such as incorporating more robust occlusion-handling modules and enhancing data augmentation strategies targeting extreme lighting conditions.

Precision–recall (PR) curve analysis

To further evaluate the performance of the improved model, this study plotted the PR curves of the models before and after the improvement (as shown in Fig. 10) to directly observe the model’s optimization effect. The PR curve (Precision-Recall Curve) is a graph where Precision is on the vertical axis and Recall is on the horizontal axis. The larger the area under the PR curve, the better the model’s overall performance⁴³. The PR curve of the improved YOLOv8-CEFC model shows a noticeable change, with a larger area covered by the curve. This indicates that, compared to the pre-improvement model, the YOLOv8-CEFC model has higher Precision and Recall. Additionally, the mAP50 of the YOLOv8-CEFC model improved from 95.2 to 96.3%, demonstrating enhanced feature extraction capability that significantly boosts performance in tasks like identifying bitter gourd’s tubercles.

Fig. 10.

The PR curves of the model before and after improvement.

Statistical validation of performance gains

Using 30 bootstrap resamples of the test set, we obtained 30 pairs of measurements. The baseline YOLOv8 and the improved YOLOv8-CEFC achieved mean mAP50 scores of 0.9574 ± 0.0044 and 0.9675 ± 0.0053, respectively, and mean mAP50–95 scores of 0.9206 ± 0.0047 and 0.9433 ± 0.0050, respectively. The corresponding average performance gains (CEFC − Baseline) were 0.0101 for mAP50 and 0.0227 for mAP50–95, with 95% bootstrap confidence intervals of [0.0018, 0.0155] and [0.0162, 0.0270], respectively (Table 3). Since the lower bounds of both confidence intervals are greater than 0, the performance improvements of YOLOv8-CEFC over the baseline YOLOv8 can be regarded as statistically significant.

Table 3.

Bootstrap-based statistical comparison between baseline YOLOv8 and YOLOv8-CEFC (30 resamples).

Metric	Baseline YOLOv8 (mean ± std.)	YOLOv8-CEFC (mean ± std.)	Mean Δ (CEFC − Baseline)	95% CI of Δ
mAP50	0.9574 ± 0.0044	0.9675 ± 0.0053	0.0101	[0.0018, 0.0155]
mAP50–95	0.9206 ± 0.0047	0.9433 ± 0.0050	0.0227	[0.0162, 0.0270]

Ablation study

To verify the performance of different improvement modules, this study conducted an ablation study by testing various combinations of the improvement modules while keeping the dataset split unchanged. The experimental results are shown in Table 4, where “-” indicates that the original structure remains unchanged.

Table 4.

Model ablation study Results.

ID	Backbone	Attention	Loss	Precision	Recall	mAP50	mAP50-95	F1-score
1	–	–	–	91.90%	90.90%	95.20%	90.20%	91.40%
2	ConvNeXt V2	–	–	93.50%	93.20%	96.30%	93.70%	93.35%
3	–	EMA-	–	91.00%	92.20%	95.50%	92.20%	91.60%
4	–	–	Focal-CIoU	91.00%	92.10%	95.30%	91.00%	91.55%
5	–	EMA	Focal-CIoU	91.80%	92.40%	95.40%	91.90%	92.10%
6	ConvNeXt V2	–	Focal-CIoU	93.20%	93.80%	96.60%	94.00%	93.50%
7	ConvNeXt V2	EMA	Focal-CIoU	93.90%	94.40%	96.30%	93.60%	94.15%

The specific ablation experiment analysis is as follows: Group 1: A baseline experiment was conducted using the original YOLOv8n model. Group 2: The ConvNeXt V2 module was introduced into the backbone network, resulting in a 1.6% increase in Precision, 2.3% in Recall, 1.1% in mAP50, and 1.95% in F1-Score. Notably, mAP50-95 increased by 3.5%, demonstrating the module’s adaptability to challenging samples and improving the model’s robustness in high-precision localization and complex environments. Group 3: After adding the EMA attention mechanism alone, Recall improved by 1.3%, but Precision slightly decreased. This indicates enhanced robustness in occluded scenes, but the mechanism still requires integration with other modules to reduce false positives. Group 4: Replacing the original CIoU Loss with Focal-CIoU Loss, which dynamically weights the focus, resulted in a 1.2% increase in Recall, effectively mitigating the issue of sample imbalance. Group 5: The combination of EMA attention and Focal-CIoU Loss led to improvements in Precision compared to using EMA or Focal-CIoU alone, with the most significant increase in Recall. This indicates that these two components contribute to more accurate target localization, reducing localization errors and effectively decreasing false negatives. Group 6: Combining the ConvNeXt V2 backbone network with the Focal-CIoU Loss function resulted in a Precision of 93.2%, Recall of 93.8%, mAP50 of 96.6%, and mAP50-95 of 94.0%. This shows that adding both the ConvNeXt V2 backbone and Focal-CIoU Loss significantly improved the overall detection performance of the model. Group 7: Finally, combining EMA attention with the ConvNeXt V2 backbone network and Focal-CIoU Loss exhibited the best performance, achieving a Precision of 93.9%, Recall of 94.4%, mAP50 of 96.3%, mAP50-95 of 93.6%, and an F1 score of 94.15%. The improvements in both Precision and Recall were particularly remarkable, indicating that this combination not only accurately detects the target but also ensures high precision, drastically reducing the false positive rate. In conclusion, through a series of ablation experiments, the improved YOLOv8-CEFC model proposed in this study effectively reduces both false negatives and false positives, significantly improving recognition accuracy for bitter gourd’s tubercles and their distribution variations.

Comparative experiments

A comparative experiment was conducted to evaluate the overall performance of various algorithms in the bitter gourd phenotype detection task. The models compared include YOLOv8n, YOLOv11, YOLOv3-tiny, YOLOv3, YOLOv5, YOLOv5s, YOLOv10n, YOLOv10s, YOLOv6s, and YOLOv8-CEFC, to further verify the performance of YOLOv8-CEFC. As shown in the experimental results in Table 5, the original YOLOv8n model has the smallest number of parameters and the fastest processing speed at 553.55 FPS, but its Precision and mAP50-95 are relatively low, indicating that it sacrifices some precision in high-speed detection, potentially missing detailed features. The YOLOv3-tiny model has a similar parameter count to the improved model, but its Precision and Recall are both lower than those of YOLOv8-CEFC, particularly in mAP50-95, where the difference is 3.4%. This suggests that its feature extraction capability is insufficient and that it struggles with adapting to complex scenes. In contrast, YOLOv8-CEFC performs exceptionally well across several key metrics, with its F1-Score reaching 94.15%, an improvement of 2.75% over the original YOLOv8n model. This indicates that the model achieves a better balance between Precision and Recall, offering the best overall detection performance and making it well-suited for practical applications that require high reliability and accuracy.

Table 5.

Comparative experiment Results.

Models	Precision	Recall	mAP50	mAP50-95	F1-Score	Parameters	GFLOPs	FPS
YOLOv8n	91.90%	90.90%	95.20%	90.20%	91.40%	2,686,708	6.8	553.55
YOLOv11	91.90%	92.60%	95.80%	89.50%	92.25%	2,584,492	6.3	462.06
YOLOv3-tiny	91.70%	91.50%	94.80%	90.20%	91.60%	9,525,192	14.3	378.79
YOLOv3	92.90%	92.40%	94.80%	91.30%	92.45%	98,456,212	261.8	46. 95
YOLOv5	92.20%	90.70%	95.40%	91.30%	91.44%	2,184,004	5.8	507.8
YOLOv6	88.6%	89.7%	93.5%	90.1%	89.15%	4,156,212	11.5	408.32
YOLOv10n	93.00%	89.30%	95.10%	89.30%	91.11%	2,699,096	8.2	494.63
YOLOv5s	92.70%	92.20%	96.30%	93.50%	92.45%	7,818,260	18.8	318.89
YOLOv6s	93.10%	90.30%	95.00%	90.60%	91.67%	15,978,484	42.8	294.12
YOLOv8-CEFC	93.90%	94.40%	96.30%	93.60%	94.15%	9,488,884	21.7	235. 76

In addition, to evaluate the computational efficiency and real-time potential of the models, we measured the inference speed of all comparison models under the same experimental conditions, with the results summarized in Table 5. As shown in the table, YOLOv8-CEFC achieves an inference speed of 235.76 frames per second (FPS). Although its number of parameters (9.49 M) and computational complexity (21.7 GFLOPs) are higher than those of lightweight models such as YOLOv8n (2.69 M, 6.8 GFLOPs), leading to a decrease in FPS, a speed of 235.76 FPS still indicates excellent real-time processing capability on the current computing platform and far exceeds the basic requirement for real-time detection (typically 30 FPS). The increase in model complexity represents a necessary trade-off to obtain higher detection precision (an improvement of 2.0%) and recall (an improvement of 3.5%).

In summary, YOLOv8-CEFC achieves a favorable balance among detection performance, computational efficiency, and model complexity. The model embodies an application-oriented and purposefully designed trade-off: through appropriate architectural enhancements, it can realize high-precision, high-throughput detection of complex bitter gourd phenotypic traits at an acceptable computational cost, providing a practical and effective solution for field phenotyping in bitter gourd breeding.

To clearly demonstrate the effectiveness of the YOLOv8-CEFC model, this study collected bitter gourd images with varying fruit shapes and tubercle characteristics for inference results analysis. The results show that YOLOv5s and YOLOv3-tiny performed poorly when recognizing the complex texture of mixed granular tubercles and strip-like tubercles, resulting in a high number of recognition errors. Additionally, the YOLOv8n and YOLOv11 models misidentified leaves and vines as bitter gourds during the detection process, exhibiting a high false positive rate. The YOLOv10n model, on the other hand, displayed a relatively higher false negative rate, as some results, shown in Fig. 11, were missed. In contrast, the YOLOv8-CEFC model demonstrated significant performance advantages under the same testing conditions. It accurately recognized different bitter gourd phenotype features and exhibited higher robustness in handling complex backgrounds and interference from similar features. These findings indicate that the YOLOv8-CEFC model has a remarkable advantage in capturing subtle features of bitter gourd, particularly excelling in detecting complex tubercle textures, while effectively reducing both false positives and false negatives.

Fig. 11.

Detection performance of some comparative experiments. (a) Original, (b) YOLOv8-CEFC, (c) YOLOv5s, (d) YOLOv8n, (e) YOLOv3-tiny, (f) YOLOv11.

Discussion

This study proposes an improved YOLOv8-based high-throughput bitter gourd phenotype detection model (YOLOv8-CEFC), which integrates the ConvNeXt V2 module, EMA attention mechanism, and Focal-CIoU loss function. The experimental results demonstrate that the YOLOv8-CEFC model outperforms other comparative models in key metrics such as Precision, Recall, mAP50, and F1-Score, achieving 93.9%, 94.4%, 96.3%, and 94.15%, respectively. The model successfully achieves high-precision detection of key bitter gourd features, including fruit shape (conical, ovoid, fusiform and rod-shaped), tubercle characteristics (granular tubercles and strip-like tubercles), and the distribution ratio between them. This effectively addresses the challenges of low efficiency and subjectivity in traditional manual phenotype analysis.

However, despite the promising results, there are still some limitations in this study. First, the size and diversity of the dataset are limited. Although a large number of bitter gourd images were collected from two bases, the dataset is still insufficient compared to larger and more diverse datasets, which may limit the model’s generalization ability when dealing with bitter gourd phenotype features from different regions, environments, and growing conditions⁴⁴. Second, the model’s adaptability to extreme environmental conditions needs further validation. In practical agricultural production, complex environmental factors such as rain-soaked fruit surfaces, dew reflections, and strong sunlight can affect the stability and detection accuracy of the model⁴⁵. Lastly, the computational cost and complexity of the model are relatively high, which presents challenges in deploying the YOLOv8-CEFC model in resource-constrained environments and limits its use on low-power devices⁴⁶.

To address these limitations, future research will focus on the following aspects: First, expanding the dataset. We will collect more bitter gourd images from different regions and countries, especially images with extreme environmental interference, to build a larger dataset. This will help enhance the model’s generalization ability and adaptability, enabling it to better handle diverse agricultural environments. Second, optimizing the model structure. We plan to use techniques such as “knowledge distillation” to compress the model, reducing computational costs and improving inference speed while maintaining model performance. This will allow the model to run efficiently on a wider range of devices and improve its adaptability and deployment efficiency⁴⁷. Third, we aim to expand the application range of the model. We will further explore cross-crop adaptive detection frameworks and attempt to apply this model to phenotype recognition in other crops to test its generalizability and scalability.

In summary, the YOLOv8-CEFC model proposed in this study has made significant progress in detecting bitter gourd fruit shape and tubercle characteristics, providing an efficient and accurate tool for phenotype analysis in bitter gourd breeding. We believe that with the continuous development and improvement of related technologies, deep learning-based crop phenotype detection technologies will play an increasingly important role in smart agriculture, injecting new momentum into the intelligent development of the agricultural industry.

Conclusion

To address the issues of low efficiency and high subjectivity in manual phenotype detection for bitter gourd breeding, this study proposes the YOLOv8-CEFC model. It enhances the recognition ability of fruit shape contours in complex environments using the ConvNeXt V2 module, reduces false positives caused by leaf interference with the EMA attention mechanism, and minimizes boundary box localization errors due to mixed granular and strip-like tubercles using the Focal-CIoU loss function. Experimental results show that the model achieves breakthrough performance on the bitter gourd dataset, with a precision of 93.9%, recall of 94.4%, and mAP50 of 96.3%. Compared to the original YOLOv8n model, the model’s performance improves by 2.0%, 3.5%, 1.1%, and 2.75%, respectively, in these metrics. The performance gain of the model was further examined using the bootstrap method, which confirmed that the improvement is statistically significant. At the same time, the model achieves a balance between precision and inference efficiency with a reasonable model size and computational complexity. The core value of this study lies in significantly improving breeding efficiency, enabling millisecond-level detection of bitter gourd phenotype features, reducing manual evaluation costs, and shortening breeding cycles. It provides reliable technical support for parent selection in bitter gourd breeding, promoting the transition of breeding decisions from experience-based judgment to scientific and standardized practices.

Author contributions

Conceptualization, S.L. and Y.X.C.; methodology, S.L. and Y.X.C.; investigation, S.L., Y.X.C. and H.B.X.; validation, S.L., Y.X.C., H.B.X. and Y.D.; formal analysis, S.L., Y.X.C., H.B.X., Y.D. and X.H.Z.; data curation, S.L., Y.X.C., H.B.X., Y.D., X.H.Z. and J.T.; writing-original draft preparation, S.L., Y.X.C. and H.B.X.; writing-review and editing, X.H.Z., J.T., F.J.B., J.L., Z.Q.L., F.X.Z., M.H.R., H.L.L., Z.D.H., C.H.Z. and F.L.Z.; visualization, S.L., Y.X.C., H.B.X., Y.D., X.H.Z., J.T., F.J.B., J.L., Z.Q.L., F.X.Z., M.H.R., H.L.L., Z.D.H., C.H.Z. and F.L.Z.; software, H.B.X., Y.D., X.H.Z., J.T., F.J.B., J.L., Z.Q.L., F.X.Z., M.H.R., H.L.L. and Z.D.H.; resources, Y.D., X.H.Z., J.T., F.J.B., J.L., Z.Q.L., F.X.Z., M.H.R., H.L.L. and Z.D.H.; supervision, C.H.Z. and F.L.Z.; project administration, C.H.Z. and F.L.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the following grants: Facility Solanaceous Vegetable Breeding and Industrialization Development Sub-topics (2023Fjnk04009); Seed Industry Innovation and Industrialization Project in Fujian Province (zycxny2021009); Fujian Modern Agricultural Vegetable Industry System Construction Project (2019-897).

Data availability

The datasets analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.