镜面缺陷检测在精密制造中起着至关重要的作用。然而，传统的缺陷检测方法由于其镜面反射特性而不适用于镜面。镜面缺陷检测通常由检测人员执行，这使得缺陷检测成为一项耗时且不稳定的任务。偏转测量法已广泛用于结合机器学习的镜面表面缺陷检测。然而，传统的挠度测量方法使用基于展开相位的局部曲率偏差图，只能检测几何缺陷。此外，需要为每个特定任务定义手工制作的功能。我们提出了一种基于偏转测量和深度学习的方法。偏转测量为网络提供输入图像，深度学习网络完成缺陷的识别和定位。在偏转测量中，所提出的方法使用光强度对比图代替局部曲率图，可以检测几何和纹理缺陷。在传统网络的基础上，采用深度可分离卷积核来减少参数，利用残差卷积块来缓解梯度消失或爆炸。用于特征聚合的子网用于获取缺陷特征的多尺度信息。基于实验结果的性能评价证明了所提方法的有效性。在传统网络的基础上，采用深度可分离卷积核来减少参数，利用残差卷积块来缓解梯度消失或爆炸。用于特征聚合的子网用于获取缺陷特征的多尺度信息。基于实验结果的性能评价证明了所提方法的有效性。在传统网络的基础上，采用深度可分离卷积核来减少参数，利用残差卷积块来缓解梯度消失或爆炸。用于特征聚合的子网用于获取缺陷特征的多尺度信息。基于实验结果的性能评价证明了所提方法的有效性。
Defect detection for specular surfaces plays a vital role in precision manufacturing. However, traditional defect detection methods are unsuitable for specular surfaces because of their specular reflection property. The defect detection on specular surfaces is usually performed by inspectors, which makes the defect detection a time-consuming and unstable task. Deflectometry has been widely used in defect detection for specular surfaces combined with machine learning. Nevertheless, conventional deflectometry methods use the local curvature deviation map based on the unwrapped phase, which can only detect geometrical defects. Moreover, hand-crafted features need to be defined for each specific task. We present a method based on deflectometry and deep learning. Deflectometry provides the input images for the network, and the deep learning network completes the identification and location of defects. In deflectometry, the proposed method uses the light intensity contrast map to replace the local curvature map, which can detect both geometrical and textural defects. Based on conventional networks, depthwise separable convolution kernel is applied to reduce parameters, and residual convolution block is utilized to alleviate vanishing or exploding gradients. A subnet for feature aggregation is used to obtain multiscale information of defect features. Performance evaluation based on experiment results proved the effectiveness of the proposed method.
