A novel real-time crayfish weight grading method based on improved Swin Transformer

This study proposed a novel detection method for crayfish weight classification based on an improved Swin-Transformer model. The model demonstrated a Mean Intersection over Union (MIOU) of 90.36% on the crayfish dataset, outperforming the IC-Net, DeepLabV3, and U-Net models by 17.44%, 5.55%, and 1.01%, respectively. Furthermore, the segmentation accuracy of the Swin-Transformer model reached 99.0%, surpassing the aforementioned models by 1.25%, 1.73%, and 0.46%, respectively. To facilitate weight prediction of crayfish from segmented images, this study also investigated the correlation between the projected area and the weight of each crayfish part, and developed a multiple regression model with a correlation coefficient of 0.983 by comparing the total projected area and the relationship between the projected area and the actual weight of each crayfish part. To validate this model, a test set of 40 samples was employed, with the average prediction accuracy reaching 98.34% based on 10 representative data points. Finally, grading experiments were carried out on the crayfish weight grading system, and the experimental results showed that the grading accuracy could reach more than 86.5%, confirming the system's feasibility. The detection method not only predicts the weight based on the area but also incorporates the proportional relationship of the area of each part to improve the accuracy of the prediction further. This innovation makes up for the limitations of traditional inspection methods and shows higher potential for application. This study has important applications in industrial automation, especially for real-time high-precision weight grading in the aquatic processing industry, which can improve production efficiency and optimize quality control.