Abstract
The Chinese mitten crab (Eriocheir sinensis) is a high-value seafood. Efficient quality-grading methods are needed to meet rapid increases in demand. The current grading system for crabs primarily relies on manual observations and weights; it is thus inefficient, requires large amounts of labor, is costly, and no longer meets the requirements for the market. Here, we employed computer vision techniques combined with deep learning modeling to efficiently quantify key physiological traits, such as sex identification, carapace dimensions (length and width), and fatness assessment for quality classification. To this end, a YOLOv5-seg integrated with an SE attention model was developed using 2282 RGB images and manual measurements of the physiological traits of 300 crabs. The RGB dataset was further augmented by rotating and resizing. The results revealed that the accuracy of sex recognition was 100%, and the mAP for carapace segmentation was 0.995, which was superior to YOLOv8-seg and other variants. In addition, we proposed an improved conditional factor K to evaluate the fatness of crabs and classify their quality based on fatness. The consistency between the grading method proposed in this article and manual grading was 100%. This study could aid in developing precise and non-destructive grading systems for the aquaculture and food industries.