부산대학교 도서관

Fast and accurate extraction of segmented images of coal-rock interfaces is one of the most important technologies to realize intelligent operation of comprehensive mining face. In this paper, we propose a multi-mechanism fusion U 2 Net model for recognizing coal-rock interfaces with complex backgrounds and diverse morphologies. We construct the basic network module by replacing the ordinary convolution in RSU with depth-divisible convolution and expansion convolution; add the coordinate attention mechanism in the last layer of the sampling stage of the improved RSU module to construct a new residual structure; and combine this residual structure with the U 2 Net model to construct a new U 2 Net model to realize the automatic learning of segmented images of coal-rock interfaces. Among them, the expansion convolution can obtain a larger sensory field without reducing the resolution; the depth separable convolution helps to simplify the model; the coordinate attention mechanism can suppress the interference of each residual block in the encoding and decoding process, and improve the model performance at a small cost. Compared with the semantic segmentation models commonly used in deep learning, this method improves the accuracy of extracting features from different stages of coal rock images, reduces the model training cost, accelerates the convergence speed of the model, and improves the model's anti-interference. Experiments were conducted to compare U 2 Net-CA-DC with four semantic segmentation networks (FCN, SegNet, UNet, and U 2 Net) to verify the effectiveness of the improved model. The accuracy of the method is 95.31% and the mIOU is 93.46% on real coal rock image datasets collected from actual underground wells. The experimental results show that the method provides a new solution for coal rock image segmentation detection and can provide data support for coal rock image segmentation.