Welding Surface Inspection of Armatures via CNN and Image Comparison

This paper proposes a new method for detecting defects on the welding surface of armature based on image comparison and Convolutional Neural Network (CNN). General classification methods based on CNN need strict boundaries on the classification of categories to prevent the duplication of data distribution between classes. And in order to achieve a good recognition effect, it is necessary to ensure that all data distributions in the class are learned by the network. For industrial data sets, the data distribution of negative samples is diverse, and the boundaries of its subdivision are not clear. If all the negative samples are divided into one category, it will have a serious negative impact on the network’s learning. The method based on CNN and image comparison we proposed is aimed at industrial data sets and can solve the problem above to a certain extent. First of all, the way of image comparison allows the network to learn the difference between positive and negative samples without worrying about the specific data distribution of negative samples. Secondly, we filter the same features and extract different features of image pairs by Spatial Correlation Attention M echanism we designed, which provides an effective attention mechanism for learning the differences between images. Through experiments, our proposed detection method obtains better results than armature classification and image comparison directly by feature vector.