| 王旭， 管声启*， 刘通， 张理博， 王静国， 于资江.基于改进YOLOv5的陶瓷环缺陷检测算法[J].轻工机械,2023,41(3):66-71 |
| 基于改进YOLOv5的陶瓷环缺陷检测算法 |
| 基于改进YOLOv5的陶瓷环缺陷检测算法 |
| |
| DOI：10.3969/j.issn.1005 2895.2023.03.010 |
| 中文关键词: 陶瓷环 缺陷检测 改进YOLOv5 注意力机制 CARAFE上采样 |
| 英文关键词:ceramic ring defect detection improved YOLOv5 attention mechanism CARAFE upsampling |
| 基金项目:西安市创新能力强基计划：人工智能技术攻关项目（21RGZN0021）。 |
|
| 摘要点击次数: 598 |
| 全文下载次数: 231 |
| 中文摘要: |
| 针对陶瓷环缺陷信息微弱且特征各异、种类繁多导致特征难以提取的问题，课题组提出了一种基于改进YOLOv5的陶瓷环缺陷检测算法。首先，在YOLOv5中的Backbone网络添加CBAM注意力机制模块，通过学习的方式在空间和通道上对特征图像进行权重分配，有效地提升模型对于不同类型缺陷通道特征和空间特征信息的提取能力；然后，采用CARAFE算子替换原YOLOv5中的最近邻上采样算子，该模块基于输入特征自适应生成上采样内核，有效的增加模型的感受域；最后，添加一个新的特征融合层，提取较低的空间特征并将其与深层的特征进行融合生成新的特征图，提升模型对小目标缺陷的检测能力。实验结果表明课题组提出的陶瓷环缺陷检测算法检测所有缺陷种类平均精度均值可以达到90.7%，能够实现陶瓷环缺陷的检测。 |
| 英文摘要: |
| Aiming at the problem of weak defect information and diverse features of ceramic ring defects, which makes it difficult to extract detection features，a ceramic ring defect detection algorithm based on improved YOLOv5 was proposesed. First, a CBAM attention mechanism module was added to Backbone in YOLOv5 to assign weights to feature images in space and channels by learning, which effectively improved the ability of model to extract channel features and spatial feature information for different types of defects; Then the nearest neighbor upsampling operator in original YOLOv5 was replaced by the CARAFE operator. This module adaptively generated an upsampling kernel based on input features, effectively increased the perceptual domain of the model; Finally, a new feature fusion layer was added to extract the lower spatial features and fuse them with the deeper features to generate a new feature map, and improved the ability of model to detect small target defects. The experimental results show that the mean average precision of the proposed ceramic ring defect detection algorithm for detecting all defect types can reach 90.7%, which can effectively achieve the detection of ceramic ring defects. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
