Abstract
Fish capture usually requires classification of fish species, and the cost of manual classification is relatively high. Recently, deep learning has been widely applied in the fishery field. Transfer learning was conducted on ResNet18, ShuffleNet, EfficientNet, MobileNetV3, and YOLOv8. Through analysis of the influence of the law of learning rate on accuracy during the network learning process, a variable-step learning rate optimization strategy was proposed. Experimental results indicate that the optimal learning rates for fish classification utilizing this strategy were determined to be 0.01, 0.015, 0.001, 0.001, and 0.006 for ResNet18, ShuffleNet, EfficientNet, MobileNetV3, and YOLOv8, respectively. The recognition accuracy rates on the sample set reach 96.33%, 96.74%, 97.50%, 86.73%, 88.49%, respectively, and the average recognition accuracy rate between the sample set and other multi-species interfering fish reaches 93.13%, 93.44%, 96.13%, 95.21%, and 92.16%, respectively. This enables high-precision and rapid sorting of the target fish and other multi-species interfering fish. Compared with global optimization, the number of optimizations can be reduced by more than 97.1%; and compared with the same number of optimizations, the accuracy can be improved by more than 34.21%, which improves the efficiency and accuracy of network training and provides a theoretical reference for the setting of learning rate during model training in the field of deep learning.