针对人工检测缺陷模式或 YOLOv5 等深度学习算法对工业产品的缺陷检测存在识别准确率低、模型参数规模大等问题,提出一种对微小缺陷端到端检测的嵌入式算法 G-YOLO.该算法使用卷积核为 3 和卷积核为 1 的双层卷积 FConv模块,改善了原单层卷积带来的参数量较大的问题;改进的轻量级跨阶段 GSP 模块融合坐标注意力机制用于主干网络中能够利用冗余信息实现廉价的线性操作和聚焦缺陷信息来增强特征,以提高网络对缺陷特征的提取能力;去除原 YOLOv5 的颈部模块,减少网络的参数量和提升网络检测速度.结果表明,G-YOLO嵌入式算法减少了模型大小,改善了缺陷检测的效果,较好地满足轻量化嵌入式模型的要求.
An embedded algorithm G-YOLO for end-to-end detection of small defect was proposed to solving problems of low recognition accuracy and large model parameter scale during industrial product defect detection using manual defect detection or deep learning algorithms such as YOLOv5.Firstly,G-YOLO used a double-layer convolution FConv module with a convolution kernel of three and a convolution kernel of one,which improved the problem of large parameter quantities caused by the original single-layer convolution.Secondly,the improved lightweight cross-stage GSP module fused with coordinate attention(CA)mechanism used in the backbone network could enhance features by utilizing redundant information for inexpensive linear operations and concentrating the defect information.Consequently,the network's ability to extract defect features was improved.Finally,the neck of the original YOLOv5 module was removed,the amount of network parameters was reduced and the speed of network detection was improved.The results showed that the embedded algorithm G-YOLO compresses the size of models and improved the effectiveness of defect detection,which better met the requirements of lightweight embedded models.
石兰娟;张梦斯;刘文浩;周迪斌
杭州师范大学信息科学与技术学院,浙江 杭州 311121
计算机与自动化
嵌入式;特征融合;注意力机制;冗余信息;特征提取
embedded;feature fusion;attention mechanism;redundant information;feature extraction
《杭州师范大学学报（自然科学版）》 2024 (002)
201-208 / 8
10.19926/j.cnki.issn.1674-232X.2023.02.111
评论