Deformable Feature Pyramid Network for Aluminum Profile Surface Defect Detection

Abstract

Surface defect detection is an important part in the process of aluminum production. The goal of a complete defect detection task is to realize the specific category and precise location of each defect in the image. Due to the similarity between surface defects and background, as well as the large diversity and difference in the appearance of the defects, this task is still challenging for applying this task in practice. In order to solve these problems, a deep learning based aluminum profile defect detection network is proposed. In order to realize strong classification ability, the deep convolutional neural network is used to generate feature graphs at each stage. On this basis, the variable feature pyramid module (DFP) is proposed, and deformable convolution is added to the feature layer to make the feature layer have the ability to adapt to defect deformation. Based on these multi-layer features, the size of anchor box was customized by Kmeans clustering algorithm, and then the region extraction network (RPN) was used to generate the region of interest (ROIs). For each ROI, a detector, consisting of a classifier and a bounding box regressor, produces the final detection results. Finally, we used a defect detection dataset of aluminum profiles to train and evaluate our approach. Through two sets of ablation experiments, the effectiveness of the introduced module is proved. Finally, we reached 76.92 mAP in aluminum profile data set.