Abstract: To improve the automation level of the wood processing production line,a scratch defect detection model on the wood surface based on Faster RCNN was proposed to identify and locate scratch defects under different texture backgrounds.In the image preprocessing stage,an improved bilateral filtering algorithm was proposed to smooth the texture background while maintaining the details of the scratches.A gray-scale adaptive scratch generation method was proposed for data enhancement.The deformable convolution was introduced to enhance the feature extraction ability of the model,and the rotating bounding box was used and a new bounding box regression loss function was proposed to solve the problem that the proportion of scratch defects in the horizontal bounding box was much smaller than the texture background.The images collected by the actual wood board processing production line verified the effectiveness of the proposed model.The proposed model was compared with other defect detection methods,and the results proved the superiority of the proposed model.

Key words: scratch detection, Faster RCNN, deformable convolution, rotating bounding box, regression loss

摘要： 为提高木板加工生产线自动化水平,基于Faster RCNN提出一种木板表面刮痕缺陷检测模型,识别和定位不同纹理背景下的木板表面刮痕缺陷。图像预处理阶段提出改进双边滤波算法,在保持刮痕细节特征的同时对纹理背景进行平滑处理;提出灰度自适应刮痕生成方法进行数据增强处理。引入可形变卷积增强模型特征提取能力,使用旋转包围框标注并提出新的包围框回归损失函数,解决水平包围框中刮痕缺陷占比远小于纹理背景的问题。通过实际木板加工生产线采集的图像验证了提出模型的有效性,并将提出的模型与其他缺陷检测方法进行了对比测试,结果证明了所提模型的优越性。

关键词: 刮痕缺陷, Faster RCNN, 可形变卷积, 旋转包围框, 回归损失