RCPU-Net: A multi-scale multi-object segmentation model for coal and gangue under uneven lighting based on improved U-Net

Image segmentation is an important prerequisite for coal gangue separation. However, in actual separation scenarios, the morphological differences between coal and gangue are small, and the image illumination is uneven, resulting in background adhesion and semi-occlusion. Existing segmentation methods may lose detailed information and lead to missed separation. To address these issues, we propose RCPU-Net, a multi-scale segmentation model specifically designed for coal and gangue, building upon the U-Net framework. Firstly, ResNet50 serves as the backbone network, and wavelet convolution is introduced instead of standard convolution in the auxiliary feature extraction module. Secondly, the residual structure of the original network is optimized, and the design criteria of dilated and mixed dilated convolution are combined to construct a dilated residual module to expand the receptive field. At the same time, the serial CP attention mechanism is introduced into the dilated residual module to enhance the deep feature extraction. Finally, a linear interpolation fusion module is designed in the upsampling to better restore semantic features. The Dice loss function is used to focus on the target boundary. Experimental results show that the improved U-Net coal and gangue segmentation model achieves mIoU and mPA of 91.84% and 96.20% on the coal gangue dataset, respectively, which are 1.44% and 1.24% higher than the baseline model and 45.41% higher than the original U-Net model on the mIoU index. This indicates that the network can optimize the boundary of the segmented target objects in the coal gangue separation scenario and improve the missed detection problem. • Collected and self-produced multi-objective coal gangue datasets in real scenarios. • Proposed an auxiliary network to enhance extraction of complex edge features. • Designed an expansion convolution residual module to retain semantic information. • Proposed a linear interpolation fusion for better multi-scale feature fusion.