Blasthole Detection on the Tunnel Face under Low-Light Environments

Abstract

Because of factors such as the complex surrounding rock backgrounds and the inability of underground artificial light sources to produce even illumination, blasthole detection on tunnel faces is a challenging problem. To address the performance degradation of existing detection methods for blastholes on tunnel faces, this paper proposes the blasthole cascaded detection method. Firstly, a preprocessing module with trainable parameters is designed. Through the parameter prediction network, the required gamma, color transformation, and sharpening parameter values are obtained and applied to the image, improving the image contrast, enhancing the color features of the blasthole regions, and enhancing the edge information and texture information of the blastholes. The blasthole regions are made more prominent, and the shapes and outlines of the blastholes are clearer. Then, the accurate and efficient you only look once (AE-YOLO) detection network is proposed based on the characteristic that blastholes in the image appears as extremely small objects. By adding a high-resolution detection head, and attention feature fusion module, and an efficient vision transformer block, the AE-YOLO detection network enables the detection results to effectively extract the shape features of the blastholes and the different features between the blastholes and shadows, reducing the loss of blasthole features and increasing semantic information. Finally, to evaluate the effectiveness of the proposed method, the CUMTB-BI dataset is established, which includes 5440 blasthole images taken under different complex conditions. Experimental results show that the proposed method has a good blasthole image detection performance, exhibits good generalization performance in the tunnel face scenarios, and reduces false detection and missed detection.