MC-MobileFishNet: An efficient algorithm for phenotypic analysis in heat-resistant breeding of Micropterus salmoides using keypoint detection

Abstract

The breeding of heat-resistant fish holds significant economic value in aquaculture, and phenotypic analysis reflects the growth performance of fish in high-temperature environments, making it a key factor in selecting heat-resistant varieties. To efficiently conduct phenotypic analysis for the breeding of heat-resistant Micropterus salmoides, we designed a phenotypic analysis system specifically for the selection of heat-resistant Micropterus salmoides. Utilizing a stereo keypoint detection method with two vertically aligned cameras and curve fitting analysis, the system accurately analyzes fish phenotypes. Our approach inherently avoids the pixel matching issues found in global stereo algorithms and effectively handles occlusion and overlap problems. The proposed model integrates the MobileNetV3 module and Squeeze-and-Excitation attention mechanism into the YOLOv8 backbone network to enhance feature extraction. Additionally, the neck section introduces a novel multidimensional feature concatenation structure called MultiConcat. This structure fuses and concatenates features across multiple dimensions through weighted integration, suppressing irrelevant features while emphasizing the effective ones. The improved model demonstrates significant performance enhancements, with mAP${}_{\text{50:95}}$ increasing by 3.7%, while model parameters and FLOPs decreased by 0.41 MB and 0.692G compared to the baseline model. The inference speed also increased, with FPS rising from 33.78 to 40.82. Experimental results demonstrate average accuracies of 97.65% for total length, 98.49% for body length, 93.34% for head length, and 96.20% for body height. Moreover, comparison experiments in aquaculture ponds demonstrated the model’s strong generalization, making it highly efficient and accurate in real fish ecology monitoring.