Online recognition of magnetic tile defects based on UPM-DenseNet

Abstract

The detection of surface defects of magnetic tile has a crucial impact on the quality of the motor; however, owing to the large difference in the scale of the surface defects of magnetic tile and the complex background texture, it is difficult for the convolutional neural network to extract the characteristics of the defect itself, and the phenomenon that small defects easily lose features in forward propagation further increases the difficulty of extracting features from the network model. As a result, the accuracy and robustness of surface defect detection based on machine vision are low. Aiming at the above problems, this paper proposes a UPM-DenseNet two-stage detection model. First, the original image data are generated through the positioning network to generate areas where defects may exist, and the influence of complex backgrounds is eliminated to a feasible extent. Then, the obtained feature map is used as the input of the classification network. Finally, the online recognition of the magnetic tile defect is completed through the classification network. To enhance the neural network’s ability to recognize regions of interest for defects of different scales, a plug-and-play feature restoration module is further proposed. The experimental results on the defective set of magnetic tiles show that UPM-DenseNet was effective for different kinds of defects. Under normal, out-of-focus and, noise conditions, the sparse classification accuracy of this method was 96.385%, 95.107%, and 84.47%, respectively. Compared with mainstream network frameworks, it increased by 1.037%, 0.83%, and 1.531%, respectively. The average area under the curve of various defects reached 0.9863, and the average detection speed was 11.109 FPS, both of which meet the accuracy and speed requirements for the detection of small defects on the surface of the magnetic tile in industrial production.