ELCNN: A Deep Neural Network for Small Object Defect Detection of Magnetic Tile

Weijian Liang, Yaoru Sun

Abstract: Automatic detection of magnetic tile surface defects is an important and challenging task, especially for small objects such as cracks. To address this problem, we propose a novel framework called feature enhancement and loop-shaped fusion convolutional neural network (ELCNN), which enhances shallow features and fuses features with a loop-shaped feature pyramid structure. First, our network focuses on shallow features and maximizes their extraction and enhancement. We believe that if the shallow features cannot be completely extracted and effectively propagated, any further improvements will be futile. Second, inspired by the mechanism of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">looking and thinking twice</i> , we propose a loop-shaped feature pyramid network (LFPN), which not only applies top-down and bottom-up multiscale feature fusion, but also connects features recursively to the left and right, creatively constructing a loop-shaped feature pyramid structure. The structure enables features to be propagated and fused in all directions, making each layer more directly connected. Just like the shortcut in ResNet, the loop structure is like a loop shortcut in the feature pyramid, which can make feature fusion more efficient and comprehensive. Also, the loop-shortcut can improve the flow of information and gradients throughout the network, making them easy to train. The proposed ELCNN is validated on the production line, and the results demonstrate that our network can significantly improve detection accuracy and meet real-time requirements. We also built a new large-scale dataset magnetic tile-small object dataset (MT-SOD) for the detection of small defects in magnetic tiles and proved its effectiveness in practical applications.