Automated Underwater Fish Species Recognition using Deep Learning-based Techniques

Knowledge about fish species with continuous monitoring play a dominant role in determining short and long-term effects on ecosystems and generating ways to manage the problem through specific treatment. Categorization and identification of fish species from underwater film are critical for maintaining resources of fisheries and environmental balance. It is difficult to get an accurate result from actual underwater films owing to poor resolution and object motion. Deep Learning (DL) technology is effectively used in a variety of sectors including aquaculture. In order to identify and recognize fishes in the presence of marine background noise, this research offers architecture for fish detection system that includes an updated training model and a multi-scale fish detection module using enhanced Convolutional Neural Network (CNN). The training technique entails gradual training of enhanced CNN by initially focusing on learning tough species followed by gradual learning of new species by incrementally employing knowledge distillation loss while maintaining high performance of previously learned species. This detection model employs a DL strategy based on CNN to improve detection performance on small-sized objects. As a consequence, the issues prevalent in identifying objects underwater are reduced, and performance of the framework is improved. In contrast to current identification methods, the proposed techniques offer better accuracy.