Analysis of Effect of Image Augmentation with Image Enhancement on Fish Image Classification Using Convolutional Neural Network

Developing a fish species classification model using a Convolutional Neural Network (CNN) requires a diverse and abundant training dataset. In addition, the quality of the images in the dataset also affects the model's performance. This research aims to investigate the effect of image enhancement through augmentation on fish image classification using CNN. The Fish-gres dataset is used in this study, and two image enhancement techniques, Histogram Equalization (HE) and Contrast Limited Adaptive Histogram Equalization (CLAHE) are applied to the training dataset. The experiment involved a non-pre-trained CNN and three pre-trained CNN models (ResNet50, Xception, and VGG16) trained using three different datasets, i.e., the original training data, training data augmented with HE, and training data augmented with CLAHE. We also experimented using different learning rates. The accuracies of each model are compared and evaluated. The experiment results showed that all models except Xception augmented with HE and CLAHE produced higher accuracy performance than those without augmentation. The best model for the Fish-gres dataset is CNN with ResNet50 pre-trained model, HE-augmented training data, and a learning rate of 10^-3 with 100% accuracy.