Improvement of YOLOv7 model for intelligent recognition of mining surface cracks

Abstract

The surface cracks caused by mining have a significant impact on the safety production of coal mines. Utilizing unmanned aerial vehicle (UAV) image recognition to identify the development characteristics of cracks in coal mining subsidence areas holds practical value. In this paper, an improved YOLOv7 model for intelligent recognition of mining-induced cracks is proposed. It incorporates the GAM global attention mechanism into the backbone network of the model to reduce the loss of feature information during the image convolution process. Additionally, depth-wise separable convolutions are introduced to replace the ordinary convolutions in the backbone network, which reduces the size of the image feature map without causing feature loss during the convolution process. Furthermore, GIOU_Loss is further used to improve the network boundary frame loss function, which improves the small target identification ability of the model. For the experiments, a training dataset for mining-induced crack recognition was created using the LabelImg image annotation software, containing 2 143 images with a resolution of 5 472×3 648 pixels. After initial preprocessing of the UAV images in the subsidence area to construct a digital orthophoto model (DOM), the DOM was further cropped using a sliding window of size 640×640 for feature recognition. The experimental results demonstrate that the improved YOLOv7 model exhibits significant improvements in the recognition performance and missed detections of dense, small-sized, and vegetation-covered crack images. The average precision (mAP) reaches 0.84.

Keywords

• uav imagery
• mining-induced cracks
• feature extraction
• yolov7
• global attention mechanism