A Survey on Fish Detection and Species Recognition

Automated fish identification is essential to overcome the challenges and time constraints associated with manual identification processes. Various techniques are explored, evaluating their performance based on factors like pre-processing methods, significant characteristics, and accuracy. Fish classification is a well-studied problem with applications in diverse fields, including target marketing and government initiatives to manage fish supply and maintain ecological balance. These automated methods play a pivotal role in sectors like commercial fisheries, agriculture, marine science, and the broader industrial fish market, supporting industries such as nutrition and canning factories. One approach involves enhancing fish recognition algorithms based on models like AlexNet, IDNet, SAFNet. Methods such as item-based soft attention mechanisms, reduced structural complexity, and transfer learning have been employed to improve accuracy and reduce training time. Similarly, researchers have applied deep learning algorithms like Mask R-CNN, MobileNet, and Fast RCNN to tackle species identification, length estimation, and object detection in marine environments. These techniques have shown promising results, especially when dealing with overlapping fish or elongated objects.Furthermore, lightweight models have been developed for underwater object detection, considering issues like low visibility, colour distortion, and small target size. Techniques like attentional feature fusion, modified feature pyramid generation, and adaptive anchor generators have been proposed to optimize performance, achieving a balance between accuracy and processing speed.