Evaluating the Effectiveness of Image Resizing Algorithms for CNN-Based Fish Disease Classification

The rapid growth of aquaculture has made the accurate diagnosis of fish diseases a critical challenge, with the potential for significant economic losses. This study evaluates the effectiveness of various image resizing algorithms in the context of convolutional neural network (CNN)-based fish disease classification. The performance of six resizing methods, including bilinear, Gaussian, and Hamming filters, applied in both one-step and two-step approaches, was examined across 30 experiments. The analysis focused on prediction accuracy, error rates, and variance. The results indicate that for bilinear and Gaussian algorithms, the two-step resizing methods generally outperformed the one-step approaches. In contrast, the Hamming resizing algorithm showed superior results with the one-step method. Notably, the one-step Hamming approach achieved the highest average prediction accuracy at 93.4% across 30 trials with varying random initializations. These findings suggest that selecting the appropriate image resizing technique is crucial for enhancing the accuracy of CNN-based models in fish disease classification.