Authors

Abstract

Intelligent and accurate shelling technology is essential for improving the quality of clam products. To enable the rapid and precise localization of clam meat in Ruditapes philippinarum (with half-shell) on an automated processing line, an improved clam meat detection algorithm - EET-YOLOv5, based on YOLOv5s - is proposed. This algorithm enables real-time detection and localization of clam meat on the production line. It integrates the Efficient Local Attention (ELA) mechanism to enhance target localization, adopts the EIoU loss function to reduce bounding box regression error, and replaces the original detection head with a TSCODE decoupled head to improve detection accuracy. The algorithm achieved a Precision of 93.03%, Recall of 97.03%, and mean Average Precision (mAP) of 93.55%, with a detection speed of 13.3 ms. Compared to YOLOv4, Faster R-CNN, SSD, and the standard YOLOv5 series, EET-YOLOv5 demonstrated superior performance. It was deployed on a test workbench for positioning experiments, achieving an average response time of 1.8 seconds and a positioning success rate of 92.7%, indicating its suitability for automated clam shell-meat separation production lines.

Abstract in Chinese

智能化、精确化的脱壳技术是提高蛤肉产品质量的关键。为了实现自动化加工生产线上对开半壳的菲律宾蛤仔中蛤肉的快速准确识别，提出了一种基于YOLOv5s改进而来的蛤肉检测算法（EET-YOLOv5），可以对生产线上的蛤肉进行实时定位识别。该算法融合Efficient Local Attention（ELA）注意力机制，能够有效捕捉目标位置；采用EIOU损失函数，减少边界框回归损失；使用TSCODE解耦头替换原有检测头，提高检测准确率。该算法检测蛤肉的精确率、召回率和平均精度均值分别达到93.03%，97.03%，93.55%，检测速度达到13.3ms。将其与YOLOv4、Faster-RCNN、SSD和YOLOv5系列等算法比较具有明显优势，适用于自动化蛤仔壳肉分离生产线。