Advanced Image Enhancement and Transfer Learning Optimization for Freshwater Fish Classification

Abstract

This study proposed an enhanced computer vision framework for fish image classification to support potential fishing zones identification. The research workflow consisted of literature review, data collection, data augmentation, image enhancement, transfer learning experiments, hyperparameter optimization, and comprehensive evaluation. Image data were collected for five freshwater fish species, namely Clarias sp., Pangasius sp., Oreochromis niloticus, Osphronemus goramy, and Cyprinus carpio. Four pretrained convolutional neural network architectures, i.e. ResNet50, MobileNetV2, DenseNet, and EfficientNet. In addition, four image enhancement pipelines were evaluated, namely bilateral filtering (BF), bilateral filtering with color constancy (BFCC), bilateral filtering with colour constancy and dark channel prior (BFC2DCP), and the proposed BFC2DCP-SRGAN method. Further experiments focused on optimizing the ResNet50 model using combinations of Dropout and L2 regularization with Adam, SGD, and RMSprop optimizers. Experimental results demonstrated that the BFC2DCP-SRGAN approach achieved the best performance, yielding training, validation, and test accuracies of 95.75%, 94.54%, and 89.54%, respectively.