A Multilevel Information Fusion-Based Deep Learning Method for Vision-Based Defect Recognition

摘要

Vision-based defect recognition is an important technology to guarantee quality in modern manufacturing systems. Deep learning (DL) becomes a research hotspot in vision-based defect recognition due to outstanding performances. However, most of the DL methods require a large sample to learn the defect information. While in some real-world cases, it is difficult and costly for data collecting, and only a small sample is available. Generally, a small sample contains less information, which may mislead the DL models so that they cannot work as expected. Therefore, this requirement impedes the wide applications of DL in vision-based defect recognition. To overcome this problem, this article proposes a multilevel information fusion-based DL method for vision-based defect recognition. In the proposed method, a three-level Gaussian pyramid is introduced to generate multilevel information of the defect so that more information is available for model training. After the Gaussian pyramid, three VGG16 networks are built to learn the information and the outputs are fused for the final recognition result. The experimental results show that the proposed method can extract more useful information and achieve better performances on small-sample tasks, compared with the conventional DL methods and defect recognition methods. Furthermore, the analysis results of the robustness and response time also indicate that the proposed method is robust for the noise input, and it is fast for defect recognition, which takes 13.74 ms to handle a defect image.