Fish Classification System Using Customized Deep Residual Neural Networks on Small-Scale Underwater Images

Abstract

Recent improvements in marine science research have increased the importance of underwater fish species identification. Using technology to automate fish species identification would positively impact marine biology. Since deep learning techniques, image classification problems have become increasingly popular. Wild natural habitats make it harder to identify fish species because of the complex background and noise in the raw images. Some of the most advanced approaches for categorizing fish species in their natural habitats have been developed in the previous decade. This paper demonstrated an automated approach for classifying fish species based on deep residual networks. Existing transfer learning models do a good job working with smaller datasets, but not so effectively. A novel RESNET model (SmallerRESNET) is developed to reduce the overfitting generated by the standard pre-trained RESNET model. Convolutional and fully linked layers are used in the more straightforward form of the RESNET model. We evaluated and compared six different versions of the RESNET model. In addition to the number of convolutional and fully connected layers, the number of iterations required to achieve 80.56% accuracy on training data, batch size, and the dropout layer is examined. Compared to the original RESNET model, the proposed and modified RESNET model with fewer layers obtained 90.26% testing accuracy with a validation loss of 0.0916 on an untrained benchmark fish dataset. The inclusion of a dropout layer enhanced our proposed model's overall performance. It is more efficient with less memory, fewer training photos, and less computing complexity than its predecessor.