Identifying half-holes during tunnel drill-and-blast construction often suffers from low accuracy, robustness, and automation. To address these issues, a deep learning-based model for intelligent tunnel blasting half-hole recognition, called deformable Transformer-illumination-adaptive-Transformer (IAT)-Hough (DIH), is proposed. This model leverages the geometric principles of the Hough transform and employs an IAT to enhance illumination adaptively, thereby improving image quality. A ResNet-50 backbone extracts low-level features of half-holes, including edges, textures, and brightness gradients. In addition, a deformable Hough transform method is introduced to enhance linear feature representations by incorporating the geometric characteristics of half-holes. The fused multi-scale features and positional encoding are then processed within a Transformer-based global modeling framework, where a self-attention mechanism enables precise localization and recognition of half-hole regions. Results show that the DIH model effectively suppresses noise interference from water stains, rock fractures, and varying illumination in complex rock mass environments, significantly enhancing the accuracy and stability of half-hole recognition. Comparative experiments reveal that DIH outperforms mainstream detection algorithms in key metrics such as mAP, Recall, and Precision. Notably, the mAP@0.5:0.95 achieves 59.4%, representing an 8.1% improvement over the original Transformer baseline, verifying the model′s robustness and generalization capabilities in complex tunnel scenarios. Using the recognition results, an automatic half-hole ratio calculation system for tunnel engineering is further developed.