Adaptive Block Relative Edge Density (ABRED): An Interpretable Feature for Tunnel Surrounding Rock Fracture Identification

Fracture density estimation from tunnel face imagery is a critical step in rock mass classification and quantification of fragmentation severity. However, traditional texture descriptors have three inherent drawbacks: fixed spatial partitioning, sensitivity to illumination, and lack of modeling of local contrast variations. To overcome these problems, we propose the Adaptive Block Relative Edge Density (ABRED) feature as a physically motivated, interpretable design for geological image interpretation. A quadtree decomposition guided by local gradient variance dynamically adjusts region granulation to be finer in areas with denser fracture and coarser otherwise. ABRED directly calculates relative edge density, where relative edge density is defined as the ratio of a block's normalized edge count to the mean edge count of its eight neighboring blocks, demonstrating empirical robustness to illumination variations in our tests. Multi-scale representation is performed by aggregating Gaussian blur edge density maps on gradually blurring scales. Compared to fixed-grid competitors and handcrafted features, including LBP, HOG, and Gabor, ABRED achieves an overall classification performance of 87.6% on an operational, real-world dataset of 1200 tunnel surrounding rock images annotated with three unique fracture severity classes, significantly outperforming baseline fixed-grid methods and handcrafted feature approaches. Ablation studies confirm each component contributes individually, while PCA-based embedding results in well-separated, class-discriminative clusters. Importantly, by incorporating domain-specific physical information directly into its feature engineering-the opposite of data-driven black box optimization-ABRED achieves performance comparable to recent deep learning approaches on our specific dataset (87.6% vs. 88-91% reported for CNN-based methods under similar conditions), while requiring substantially lower computational overhead and offering full interpretability. We note, however, that deep learning models may achieve superior accuracy given substantially larger training datasets. The advantage of ABRED lies in its efficiency, transparency, and suitability for CPU-based field deployment where resources are constrained.