Single-Scale Workpiece Defect Detection Based on Deep Learning

As Mask R-CNN network can effectively detect objects, it can be used to detect product defects on the images from the production line in more intelligent manner. However, in order to speed up and enhance the recall rate for specific defects, there are many parts in the architecture can be optimized and improved. In this paper, the anchor box with specific scales are designed to adapt to specific defects to avoid redundant multiple anchors, and the shared feature extraction layers of Mask R-CNN is improved for more efficient process. The improved network is applied to detection of gears with tiny defects. Experimental results demonstrate that the improved networks aimed at specific defects of gears achieve a notable improvement in terms of both recall index and detection speed. This method can be applied to other specific defects such as button scratches, printed circuit defect,etc.