Adaptive convolutional neural network for aluminum surface defect detection

Abstract

The aluminum surface defect detection is not trivial for the high computational cost and labor-intensive data annotation. Particularly, the characteristics of lots of very tiny objects, sample sparsity, and variations, limit the detection performance. In this paper, we demonstrate a sophisticated and efficient object detection model based on hierarchical attention and contextual information for aluminum surface defect detection. Specially, we first use a deep residual learning strategy to obtain the defect feature maps. Secondly, we add corresponding weight matrices to the defect feature maps by fusing attention mechanism and adaptive deformable convolution to achieve the fine feature. Thirdly, we construct a feature pyramid structure to achieve the fusion of multi-scale feature information. Finally, we use the obtained contextual feature information for class prediction and bounding box regression. The comprehensive experiments on the surface defect data set of aluminum and the surface defect data set of copper foil and aluminum foil respectively show that our method compared to state-of-the-art object detectors. Code is available at https://github.com/yunsheng-Wei/DFA-FRCNN.git.