Optimizing Fish Classification with a Hybrid SAE-SVM Model: Performance and Evaluation on Fish-gres and Fish4-Knowledge Datasets

Accurate classification of fishes is thus very important in marine biology for species identification, biodiversity monitoring, and ecological studies. Most typical methods used are manual feature extraction techniques and expert analyses, which generally are time-consuming and unreliable. These are, however, surmounted in effectiveness and scalability by automated classification systems, especially those based on picture data. In the context of the present research, we focus our attention on the development of an image-based fish classification system supported by deep learning technologies. It could potentially enhance both the accuracy and speed of species identification much further. We propose one of its kind classification systems that employs a Convolutional Neural Network (CNN). The feature extraction and analysis process are done automatically through cascadable convolutional layers from the first stage to the last. Convolution neural networks, owing to their multilayer composition, manage to extract high-level features, but this extraction normally occurs at the partialcost of some essential low-level information which may reside in further layers. The hybrid approach proposed here handles this limitation because CNN is combined with a Stacked Autoencoder for dimensionality reduction and hence Support Vector Machine for classification. This has aimed to reduce the impact of the mentioned constraint. It also helps the SVM to enhance the process of classification, while the autoencoder contributes to maintaining low-level information. The findings of our experiments reveal that our hybrid model performs better compared to state-of-the-art approaches, proving its usefulness in the classification of fish species.