Segmentation Method of Machine-Made Sand and Stone Powder Based on Residual Adaptive Attention U-Net

To solve the cumbersome steps and low accuracy problems in the detection method of stone powder content in machine-made sand, this paper proposes a segmentation method for machine-made sand and stone powder based on the Residual Adaptive Attention U-Net (RA2-U-Net). The image of machine-made sand particles is first enhanced and denoised through contrast enhancement, look-up table algorithm, and low-pass filtering. The U-Net network is then improved using residual connections, adaptive convolution, and attention mechanism modules to build the RA2-U-Net model. The results show that the RA2-U-Net achieves IOU, recall rate, and precision of 83.71%, 93.83%, and 88.79%, respectively. Thus, the segmentation effect is substantially improved.