Research on sand and gravel image segmentation technology based on deep learning
Li Chen, Libo Pan, Yifeng Li · 2024
Industrial sand and gravel extraction involved processes such as geological surveying, ore selection, subsequent processing, and transportation. Screening and washing were crucial steps in the subsequent processing phase, with screening serving as a prerequisite for reducing water consumption during gravel washing. Addressing the current industrial inefficiencies characterized by low mechanical screening efficiency, high maintenance costs, and the need for extensive manual inspection, this paper proposed an improved sand and gravel image recognition method based on the SeResUnet-50 model. This method replaced the backbone network with a pre-trained ResNet50 network for feature extraction in the encoding part of the network. Subsequently, the Senet attention mechanism was added to the upsampling feature fusion layer in the decoding part to extract more relevant features. Experimental results demonstrated that the sand and gravel image segmentation network model based on SeResUnet-50 achieved a pixel segmentation accuracy of 95.64% and an mean intersection over union (MIoU) of 91.75%. Compared to PSPNet, DeepLabv3+, and the original Unet model, it showed improved accuracy and MIoU, thereby achieving effective sand and gravel image segmentation and offering a novel approach to sand and gravel identification and classification in current industrial practices.