A Hierarchical Feature Fusion-based Method for Defect Recognition with a Small Sample

As one of the breakthroughs in modern manufacturing, deep learning (DL) performs large-scale network architectures and achieves some outstanding performances in vision-based defect recognition. However, most of these large-scale networks require a large sample for training, and a small sample might cause the networks overfitting and collapse. Since the defect often occurs with a low probability, it is costly to collect large-scale samples. To overcome this problem, a hierarchical feature fusion-based method is introduced for defect recognition with a small sample. The proposed method divides a pretrained VGG16 network into different blocks, and learns the hierarchical features from the low- and high- level blocks. The results are better than the other methods. This result manifests the proposed method suits problem, and the defect recognition could be deployed earlier with the proposed method.