Fish Image Classification and Prediction Using Compact Convolutional Transformers

Abstract In this study, we propose a new method for fish image classification using Compact Convolutional Transformers (CCT). CCTs are a variation of the transformer architecture, which have been successful in natural language processing tasks. Our study begins with an in-depth background analysis, exploring the current state-of-the-art techniques in fish image classification and identifying potential gaps in the existing methodologies. We highlight the limitations of traditional convolutional neural networks (CNNs) in handling large-scale fish image datasets, such as variations in fish species. We introduce the Compact Convolutional Transformer, a fusion of Convolutional Neural Networks and Transformer architectures. We break down the methodology into distinct subsystems, encompassing a feature extraction module using CNNs, and a context modeling module employing the Transformer. By incorporating compact convolutional layers, CCTs are able to effectively capture local spatial information in images, while still maintaining the ability to model long-range dependencies. We assess the performance of our proposed method on a dataset of fish images and compare it to traditional convolutional neural networks and other state-of-the-art fish image classification methods. Our experiments show that the CCT model succeeds an accuracy of 98.6% on the test dataset. Our approach is a promising solution for fish image classification, and might be used to more associated tasks such as fish counting, fish identification and fish prediction.