Grading Related Feature Extraction of Chinese Mitten Crab Based on Machine Vision

Abstract

The current grading of Chinese Mitten crab relies primarily on manual observation and weighing, resulting in high labor intensity, high cost, and low efficiency. These limitations no longer meet the requirements for the rapid development of crab industry. This study utilizes computer vision and deep learning to rapidly extract physiological features including gender, carapace length and width for grading. A YOLOv5- seg model was trained with 764 RGB images and manually measured physiological traits of Chinese mitten crabs. The performance of the constructed models in recognizing genders and predicting carapace length and width was evaluated. The results demonstrate an average accuracy rate of 100% for gender recognition. The average absolute percentage error was 1.5% for measuring the carapace length and width. The results of this study may facilitate the development of non-destructive high-precision crab grading systems and devices for the aquaculture industry.