Enhanced batch sorting and rapid sensory analysis of Mackerel products using YOLOv5s algorithm and CBAM: Validation through TPA, colorimeter, and PLSR analysis.

Abstract

This study employed the YOLOv5s algorithm to establish a rapid quality identification model for Pacific chub mackerel (S. japonicus) and Spanish mackerel (S. niphonius). Data augmentation was conducted using the copy-paste augmentation within the YOLOv5s network. Furthermore, a small object detection layer was integrated into the network structure's neck, while the convolutional block attention module (CBAM) was incorporated into the convolutional module to optimize the model. The model's accuracy was assessed through sensory evaluation, texture profile analysis, and colorimeter analysis. The findings indicated that the enhanced model achieved a mAP@0.5 score of 0.966, surpassing the original version's score of 0.953. Moreover, the improved model's params was only 7.848M, and an average detection time of 115ms/image (image resolution 2400×3200). Furthermore, sensory and physicochemical indicators are reliably distinguished between qualified and unqualified samples. The PLSR model exhibited R2X, R2Y, and Q2 values of 0.977, 0.956, and 0.663, respectively.