Improved YOLOv5 based on deformable convolution and efficient decoupled head for pill surface defect detection

作者：Kang Liu, Zhongliang Lv, Kewen Xia, Chuande Zhou, Zhenyu Lu, Hailun Zuo, Zelun Li, Xuanlin Chen · 发表于：Other Conferences · 年份：2023 · DOI：10.1117/12.2688970 · 被引用次数：1 · 研究领域：Engineering

Aiming at the problems of low detection efficiency and poor detection accuracy caused by different shapes and sizes of pill defects in the existing market, an improved YOLOv5s model is proposed. In this study, deformable convolution is introduced into CBS structure and combined with C3 module to enable the model to adaptively adjust the receptive field and more effectively approach the shape and size of the defect target. In addition, the Efficient Decoupled Head (EDH) is introduced to replace the YOLOv5s detection head, separating the positioning task and the classification task, and improving the detection accuracy and target coverage. The experiment on the pill dataset shows that the model proposed in this paper effectively improved the detection accuracy of pill defects. The model improved the accuracy of round crack and elliptical crack defects with large differences in size and shape by 1.9% and 3.0%, respectively. Compared with YOLOv5s, mAP@0.5 and mAP@0.5:0.95 reached 97.7% and 72.3%, respectively, which increased by 1.2% and 2.6%. Additionally, the model has a smaller model volume and faster inference speed.