Rock borehole panoramic image recognition based on improved MobileViT

Abstract Panoramic imaging and state recognition technology for borehole environments in rock strata, as an intuitive method for investigating the stratigraphic distribution of target regions, offers advantages such as intuitive detection results, high accuracy in stratum state identification, and reliable detection technology. Research on this technology holds significant academic value and engineering importance. In this study, based on a dual-camera module combining a fisheye lens and a high-resolution lens, a compact and lightweight borehole imaging system integrating wide-field-of-view and high-resolution imaging was developed through structural optimization research. Subsequently, by analyzing the characteristics of the borehole inner wall in a simulated rock stratum environment, panoramic unwrapping of wide-field-of-view imaging results was achieved through image registration and coordinate mapping. Furthermore, an improved MobileViT network was proposed, which incorporates an efficient multi-scale attention mechanism to enhance multi-scale feature extraction and a cosine classifier to improve inter-class discriminability, and was combined with sliding window prediction to realize intelligent recognition of seven types of rock borehole images. Experimental results demonstrate that the vision system can achieve close-range, high-resolution panoramic imaging of long boreholes while enabling intelligent state assessment of rock materials. The recognition accuracy on the test set reached 92.12%. This study confirms that the proposed system is capable of comprehensive intelligent evaluation of geotechnical borehole conditions, providing a reliable and efficient technical means for geological investigation.