Abstract: The evolutionary characteristics of mining-induced fractures are one of the key bases for quantifying the dynamic manifestation characteristics of roadway surrounding rock. To reduce the interference of uneven illumination, noise, etc. on the imaging of surrounding rock boreholes, as well as the adverse effects of blurred edges and variable morphologies of mining-induced fractures in the boreholes on the identification of mining-induced fractures, an anti-interference identification method for mining-induced fractures in roadway surrounding rock boreholes based on Attention U²-Net is proposed. A self-developed holographic perception equipment for the situation of roadway surrounding rock is used to collect high-resolution images of mining-induced fractures in surrounding rock boreholes in real-time around the clock. Combined with various enhancement methods such as noise injection, histogram equalization adjustment, V-channel color disturbance in HSV, and 3D projection of fracture grayscale, the environmental generalization ability of image data under non-ideal imaging conditions is improved. By integrating mechanisms such as single-channel attention (SE, ECA), spatial attention (CBAM), global multi-channel attention (DANet), and combined attention (CBAM+ECA) into the benchmark model U²-Net, the ability to perceive and extract fractures under non-ideal acquisition environments such as low-visibility fractures is enhanced. In the training phase, a deep supervised composite loss function (Dice+BCE) is embedded into the 6 network output ends of the benchmark model U²-Net to promote the stable training and rapid convergence of the benchmark model U²-Net and the Attention U²-Net model, thereby alleviating the problems of gradient disappearance and discontinuity of small target fractures. The experimental results of anti-interference identification of mining-induced fractures in roadway surrounding rock boreholes show that: the IoU of the Attention U²-Net model is increased to 83.1%, the F₁ reaches 92.6%, and the E_MA is reduced to 0.052. Compared with the benchmark models U-Net and U²-Net, the convergence step in the training phase is advanced by 21 rounds and 10 rounds, and the F₁ is increased by 8.4% and 4.0% respectively. The Attention U²-Net model converges faster in training and has stronger capabilities in fracture edge detection, slender fracture extraction, and complex texture segmentation, providing reliable technical support for accurately analyzing the evolution characteristics of mining-induced fractures in surrounding rock boreholes and the dynamic manifestation characteristics of roadway surrounding rock.