Table 1.

Overview of preprocessing techniques and dataset balancing after mask application

Method	Formula	Description	Parameters
Brightness Adjustment		Adjusts brightness and contrast to enhance feature visibility.	α = 1.5, β = 15
Noise Removal [11]		Reduces salt-and-pepper noise while preserving edges using a median blur filter.	Kernel size k = 3
Contrast Enhancement [12]		Enhances local contrast using CLAHE, avoiding excessive noise amplification.	clipLimit = 2.0, tileGridSize = 3 × 3
Normalization		Scales pixel intensities to the range [0, 1] for faster neural network convergence.	None
Applying Mask		Binary mask M extracts regions of interest defined by bounding box annotations	None
Resizing		Resizes images to a uniform dimension to ensure consistency in input size for the models.	Final size = 224 × 224, nearest-neighbor interpolation
Random Downsampling (Balancing)	-	Randomly reduces the number of samples in each class to match the smallest class size after masking.	Based on class with minimum samples (894)