| 摘要: |
| 目的 铝合金铸件表面缺陷检测是工业中的一个重要应用,正确且快速地检测出铸件表面的缺陷可以大大
提高产量和质量。 针对图像中缺陷目标较小,缺陷类别易混淆,定位不精准等问题,提出了一种在一级检测器基础
上改进的 SAD-YOLOv5 模型。 方法 针对一般卷积神经网络中由于跨步卷积和池化层导致网络训练过程中信息丢
失的问题,通过引入空间到深度( space-to-depth,SPD) 模块避免细粒度信息的丢失,提高对小目标的特征学习能
力;为进一步提升网络模型精度, 在 网 络 的 Head 中 引 入 自 适 应 空 间 特 征 融 合 ( adaptively spatial feature fusion,
ASFF) 和 Decoupled Head,其中 ASFF 通过实现不同特征之间的自适应融合,抑制了不同尺度特征之间的不一致
性,保留更有鉴别性的信息,从而提升网络学习能力;使用 Decoupled Head 替换原先的耦合头,将分类和回归进行
解耦,使分类更加关注纹理信息,回归更加关注边缘信息,二者各司其职,进一步提升网络判断能力。 结果 在自己
拍摄的铸件缺陷检测数据集中的测试结果表明,SAD-YOLOv5 的 mAP@ 0. 5 和 mAP@ 0. 5:0. 95 分别为 95. 1% 和
68%,较基线模型( YOLOv5) 分别提升了 1%和 3. 3%。 结论 SAD-YOLO5 能更准确地完成铝合金铸件的表面缺陷
检测任务。 |
| 关键词: 铝合金铸件 表面缺陷检测 YOLOv5 SPD ASFF Decoupled Head |
| DOI： |
| 分类号: |
| 基金项目: |
|
| SAD-YOLOv5 Aluminum Alloy Surface Defect Detection Method Based on YOLOv5 |
|
YUAN Junsen, LING Liuyi
|
|
School of Artificial Intelligence Anhui University of Science and Technology Anhui Huainan 232000 China
|
| Abstract: |
| Objective Surface defect detection of aluminum alloy castings is a critical application in industry. Accurate
and rapid detection of defects on the surface of castings can significantly improve production and quality. In response to
challenges such as small defect targets easily confused defect categories and imprecise localization in images an
improved SAD-YOLOv5 model based on a primary detector was proposed. Methods Addressing the problem of
information loss during network training caused by strided convolutions and pooling layers in general convolutional neural
networks space-to-depth SPD module was introduced to avoid the loss of fine-grained information and enhance the
feature learning capability for small targets. To further improve the model accuracy adaptively spatial feature fusion
ASFF and Decoupled Head were introduced in the network?? s Head. ASFF achieved adaptive fusion between different
features suppressing inconsistencies among features of different scales to retain more discriminative information and
enhance network learning capability. Decoupled Head replaced the original coupled head to decouple the classification and
regression allowing classification to focus more on texture information and regression to focus more on edge information.
This division of responsibilities further enhances the network?? s decision-making capability. Results Testing on a self- captured dataset for casting defect detection showed that SAD-YOLOv5 achieved mAP @ 0. 5 and mAP @ 0. 5 0. 95 of
95. 1% and 68% respectively. This represented a 1% and 3. 3% improvement over the baseline model YOLOv5 .
Conclusion SAD-YOLOv5 demonstrates the ability to more accurately perform surface defect detection tasks on aluminum
alloy castings. |
| Key words: aluminum alloy casting surface defect detection YOLOv5 SPD ASFF Decoupled Head |
