Research on Surface Defect Detection of Aluminum Based on Improved Cascade R-CNN

Aluminum profiles are widely used in many industries with good characteristics. Surface defects of aluminum profiles will affect the quality, reliability and safety of products. In recent years, deep learning has been applied in aluminum profile surface defect detection. However, there are still some problems unsolved. The tiny and narrow defects are easily ignored in feature extraction. The length-width ratio of different aluminum surface defects varies widely, but the fixed anchor boxes in traditional deep learning algorithms will easily miss defects. Due to the lack of training on background images, some backgrounds are easily misidentified as defects. To address these problems, a novel Cascade R-CNN network with deformable convolution, guided anchoring and sample augmentation (GAE-Cascade R-CNN model) is proposed. The deformable convolution enhances the feature extraction ability of the network. The guided anchoring reduces the missed detection by automatically generating anchors to match narrow defects. The sample augmentation effectively reduces missed defect detection by training a large number of background images. The experimental results show that the proposed GAE-Cascade R-CNN model can achieve accuracy of 98.85% for identification and mean average precision (mAP) of 80.55% for surface defect detection of aluminum profiles. The performance of the proposed network outperforms other deep learning methods in terms of both missed detection rate and false detection rate.