A Dynamic Learning Network for Steel Surface Inspection Based on Structural Re-parameterization

Steel surface quality inspection is a necessary process to ensure industrial product quality. With the development of the industrial internet of things (IIoT), deep learning methods started to detect defects for high efficiency and accuracy. It is challenging to improve the model's performance when considering the different characteristics of the device in the IIoT scenario. There is a tremendous amount of computing resources in the cloud and limited computing power on edge devices. A lightweight deep neural detection network with a Dynamic structural re-parameterization strategy, Multi-task Enhanced Module, and Dual weight label assignment method (DN-DMD) is thus proposed for steel surface detection in the IIoT. To utilize the computing power in the cloud, a dynamic learning strategy is first designed to expand the backbone's structure with high significance to extract more information during training. The model can be losslessly compressed into a lightweight model by structural re-parameterization during inferencing on edge devices. Secondly, a lightweight Multi-task Enhanced Module is embedded to improve scale, spatial, and task perception features. Finally, an anchor-free label assignment method called dual weight assignment is applied to develop more information from positive and negative samples to improve performance without inference cost. This DN-DMD has been proven to achieve much better mean average precision (MAP) and mean average recall (MAR) than traditional methods.