Existing roughness measurement approaches based on machine vision cannot accurately measure irregular components with complex shapes, such as helical gears. Owing to the occlusion of relative positions between teeth, it is not possible to directly obtain an image that only contains the target surface, which decreases the accuracy and efficiency of the measurement model. This paper proposes a novel visual approach for the roughness measurement of helical gears. First, a region of interest (ROI) extraction method is designed to filter the interference information in the original image and extract the effective region. Then, a convolutional neural network (CNN) is applied to evaluate the roughness with the ROI processed image as input. The machine vision-based roughness values calculated before and after ROI extraction are compared with the stylus device-based roughness values. The accuracy and generality of the proposed approach are proved by two cases of helical gear and leadscrew roughness measurements.

现有的基于机器视觉的粗糙度测量方法无法准确测量具有复杂形状的不规则部件，例如斜齿轮。由于牙齿之间相对位置的遮挡，无法直接获得仅包含目标表面的图像，降低了测量模型的准确性和效率。本文提出了一种用于斜齿轮粗糙度测量的新视觉方法。首先，设计了一种感兴趣区域（ROI）提取方法来过滤原始图像中的干扰信息并提取有效区域。然后，应用卷积神经网络 (CNN) 以 ROI 处理后的图像作为输入来评估粗糙度。将 ROI 提取前后计算的基于机器视觉的粗糙度值与基于触控笔设备的粗糙度值进行比较。通过斜齿轮和丝杠粗糙度测量的两种情况证明了所提出方法的准确性和通用性。