Abstract:
Underground target detection technology in coal mines is an indispensable component of constructing a smart mine, providing real-time monitoring and recognition capabilities. However, factors such as uneven illumination and significant obstruction underground lead to incomplete information for certain targets, greatly reducing the accuracy of target detection. To address this, an improved algorithm for multi-objective real-time detection of incomplete information in coal mine underground is proposed, based on enhancing YOLOv5s. Recognizing that incomplete targets can easily be confused with the underground background, this algorithm incorporates a CBAM (Convolutional Block Attention Module) into the Backbone of YOLOv5s. This inclusion strengthens the channels and spatial information in the feature map relevant to incomplete targets, thus enhancing the suppression of background interference. Furthermore, to effectively extract and enhance detailed features of small and occluded targets, the Weighted Bi-directional Feature Pyramid Network (BIFPN) is employed in place of the original PANet structure. Additionally, to better adapt to the shape variations of incomplete underground targets, an Enhanced Intersection over Union (EIOU) function is introduced, incorporating additional bounding box coordinate information to optimize the existing loss function. Finally, the proposed algorithm is validated using a custom-built underground dataset. Experimental results demonstrate that the improved target detection algorithm effectively addresses challenges posed by small target sizes, partial occlusion, and variations in texture and shape within the underground monitoring environment. The enhanced model achieves an average accuracy of 91.3%, an improvement of approximately 2.7% over the original model, and an F1-Score of 90.0%, an improvement of around 1.9% over the original model.