Computer Vision-Based Deep Learning Modeling for Salmon Part Segmentation and Defect Identification

Abstract

Accurate cutting of salmon parts and surface defect detection are the key steps to enhance the added value of its processing. At present, mainstream manual inspection methods have low accuracy and efficiency, making it difficult to meet the demands of industrialized production. A machine vision inspection method based on a two-stage fusion network is proposed in this paper, aiming to achieve accurate cutting of salmon parts and efficient recognition of defects. The fish body image is collected by building a visual inspection system, and the dataset is constructed by preprocessing and data enhancement. For the part cutting, the improved U-Net model that introduces the CBAM attention mechanism is used to strengthen the extraction ability of the fish body texture features. For defect detection, the two-stage fusion architecture is designed to quickly locate the defective region by adding the YOLOv5 of the P2 small target detection layer first, and then the cropped region is fed into the improved U-Net for accurate cutting. The experimental results demonstrate that the improved U-Net achieves a mean average precision (mAP) of 96.87% and a mean intersection over union (mIoU) of 94.33% in part cutting, representing improvements of 2.44% and 1.06%, respectively, over the base model. In defect detection, the fusion model attains an mAP of 94.28% with a processing speed of 7.30 fps, outperforming the single U-Net by 28.02% in accuracy and 236.4% in efficiency. This method provides a high-precision, high-efficiency solution for intelligent salmon processing, offering significant value for advancing automation in the aquatic product processing industry.