Metal surface defect detection based on grouped convolution and bidirectional feature fusion

In this paper, we propose a metal surface defect detection model based on grouped convolutions and bidirectional feature fusion to solve the problems of dense distribution, diverse types, and small contrast between object and background. The model uses a multi-branch structure based on grouped convolutions from ResNet residual blocks of the feature extraction network layer, which increases the cardinality and feature extraction ability of the network. In feature fusion, we introduce a path enhancement network and a bottom-up path in the network to form a bidirectional feature fusion module, which fuses information in shallower layers and enhances the detection ability for small objects. We add a regression scale to reduce the amount of computation on the ordinary small objects’ regression layer. The convolution layer is shared between the center-ness branch and the regression branch to better suppress prediction boxes far away from the object center. Experimental results on NEU-DET metal surface defect dataset show that the average accuracy of our network model reaches 90.2%, which is 8.3% higher than that of Fully Convolutional One Stage (FCOS), thus our model can obviously improve the detection rate of metal surface defects.