Abstract: This research focuses on fish species classification using deep learning models, specifically ResNet-50 and MobileNetV2. The ResNet-50 model, a 50-layer deep convolutional neural network, is employed for its proven excellence in image classification tasks. The dataset comprises 20 freshwater fish species, with 57 samples per species captured using a Samsung Galaxy M30s mobile phone. Data preprocessing involves image conversion and offline augmentation to address the small dataset size. The ResNet-50 architecture consists of 5 stages with convolution and identity blocks, employing batch normalization and ReLU functions. The model is trained on 224x224x3 image inputs and utilizes over 23 million trainable parameters. The bottleneck design is incorporated into the residual units for enhanced performance. The system's effectiveness is evaluated through species detection, where the model predicts the fish species from input images. In contrast, the MobileNetV2 model is introduced as a lightweight convolutional neural network tailored for mobile and embedded devices. It employs depth wise separable convolutions and an inverted residual structure for efficiency. The network architecture includes Conv 1x1 and Dwise 3x3 layers, showcasing its ability to reduce computational cost and improve information flow. The effectiveness of both models is assessed in the context of fish species classification, providing insights into their performance and suitability for the given task. The research contributes to the understanding of deep learning models in the domain of fish classification, with implications for applications in aquatic biodiversity monitoring. Our experiments reveal that MobileNetV2 consistently outperforms ResNet50 in terms of accuracy. MobileNetV2 achieved an impressive accuracy of 99.44%, while ResNet50 achieved 98.45%. The higher accuracy of MobileNetV2 suggests its efficacy in capturing intricate features within images, even with its more compact architecture.

Keywords: MobileNetV2, ResNet50, Convolutional neural network (CNN), Data Preprocessing, Fish species detection.

| DOI: 10.17148/IARJSET.2024.11563