Surface Defect Detection Method Based on Improved Faster-RCNN
Xinhong Tong, Yixin Huang, Linjunhao Xiao, Xiao Chen, Runjie Shen
Abstract
Traditional physical inspection technology is limited by the pipeline shape, material and other factors while detecting pipeline surface defects. Therefore, the current industrial field is gradually adopting computer-aided machine vision technology to achieve defect detection. However, targets such as cracks and corrosion have great difference in size and irregular shapes, which cause the weak detection effect of the original Faster-RCNN network. In order to improve the feature extraction ability of Faster-RCNN network, we propose to add feature fusion and enhancement module, deformable convolution module and context pooling module. The experimental results show that the improved Faster-RCNN network can effectively extract the features of different scale targets by using multi-resolution feature maps and convolution kernel offsets, which improves the accuracy and recall rate of the model. For crack detection, the accuracy is 89.2% and the recall rate is 96.4%. Compared with the original fast RCNN with 80.4 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> accuracy and 92.6% recall rate, it has a significant improvement.