An improved deep CNN-based freshwater fish classification with cascaded bio-inspired networks

Abstract

Bangladesh has plentiful water, which is essential to its freshwater fish traditions. Environmentalconcerns and other causes have reduced the country’s water resources, threatening many nativefreshwater fish species. Thus, the younger generation deficiencies recognition of local freshwa-ter fish and struggles to recognize them. Traditional methods are very insufficient to overcomethese issues. To address these research gaps, the research proposes an automatic system for cat-egorizing Bangladesh’s freshwater fish. The proposed methodology involves several key steps,including building a comprehensive dataset, extracting features from fish images using pre-trained Convolutional Neural Network (CNN) models, and employing typical ML approaches.Initially comprising eight classes, the dataset undergoes feature extraction using CNN algo-rithms, followed by the utilization of various feature selection methods such as Support VectorClassifier, Principal Component Analysis, Linear Discriminant Analysis, and optimization modelslike Particle Swarm Optimization, Bacterial Foraging Optimization, and Cat Swarm Optimiza-tion. In the final phase, seven conventional ML techniques are applied to classify the images oflocal fishes. Empirical measurements are gathered and analyzed to assess the proposed frame-work’s performance. Particularly, the present approach achieves the highest accuracy of 98.71%through the utilization of the ML mechanism Logistic Regression with Resnet50, SVC, and CSOmodels