Improving the Segmentation Method of Overburden Dump Wall in Open-Pit Mines Using HRNetV2

In order to reduce the safety hazards of vehicles in open-pit mine soil yards and improve the accuracy of semantic segmentation for retaining wall images, this paper proposes an improved network model based on HRNetV2. By increasing the width parameters in the backbone network branches, the network capacity and feature representation ability are enhanced. Additionally, a global contextual attention module is employed to capture the dependencies and correlations among images at different scales, thus enhancing the model’s performance in segmenting image edges and details. Experimental results demonstrate the effectiveness of the proposed semantic segmentation model, achieving an average intersection-over-union (IoU) of 88.10%, category average pixel accuracy of 92.86%, precision of 93.75%, and recall of 93.14%. The difficulty in semantic segmentation of retaining wall images lies in the complexity of the scene background and the diversity of retaining wall structures. The improved HRNetV2 network proposed in this paper fully leverages its advantage in multi-resolution feature fusion, enhancing the semantic understanding and segmentation accuracy of retaining wall images. It holds significant practical value due to its effectiveness in addressing the challenges posed by complex scenes and diverse retaining wall structures.