电子科技大学学报2025，Vol.54Issue(6)：935-944,10.DOI:10.12178/1001-0548.2024132
ACDet:强化自我注意力机制的药品包装轮廓检测方法
ACDet:Enhanced self-attention mechanism for pharmaceutical packaging contour detection
摘要
Abstract
This article proposes ACDet(self-attention and concatenation based detector),an object vector detection and recognition method based on convolutional neural networks.This method aims to efficiently detect dense and unordered pharmaceutical packaging contours under varying lighting conditions.By employing combined image enhancement techniques,the method enhances the model's ability to learn the appearance features of objects.The C2F-A(coarse-to-fine with attention)computational module utilizes multiple gradient flows for multidimensional self-attention enhancement,encompassing both feature and spatial dimensions.The WConcat(weighted concatenation)module facilitates weighted concatenation of various levels of feature maps,capturing more critical features,thereby enhancing the network's cognitive ability.In experiments on the CPPD(cancer pathological and pharmaceutical dataset)for pharmaceutical cases,ACDet achieved 81.0%mAP(mean average precision)and 79.5%SmoothmAP,outperforming other YOLO(you only look once)architecture models by an average of 5.5%to 16.6%,and leading by 0.7%to 6.9%on public datasets.Additionally,zero-shot testing achieved a review success rate of 99.9%.The research results suggest that the proposed ACDet can overcome complex detection scenarios,enhance network robustness,and support intelligent industrial production.关键词
YOLO/药品包装轮廓/动态照度/视觉检测Key words
YOLO/pharmaceutical packaging contour/dynamic lighting/visual detection分类
计算机与自动化引用本文复制引用
陈路,李阳,周昊昱,王钧慷,钱伟中,张新昱,陈丽竹,高勇..ACDet:强化自我注意力机制的药品包装轮廓检测方法[J].电子科技大学学报,2025,54(6):935-944,10.基金项目
四川省科技计划(24SYSX0210) （24SYSX0210）
跨域飞行交叉技术实验室项目(2024-KF03004) （2024-KF03004）
