To address the challenges in tactile paving defect detection, including elongated defect structures, complex background textures, and limited computational resources of edge inspection devices, a tactile paving defect detection algorithm named IDRS-YOLO was proposed by integrating strip convolution and re-parameterized feature enhancement. A heterogeneous feature extraction module,C3k2_IDSCB (C3k2-Inception depthwise shuffle convolution block), was designed by incorporating anisotropic strip convolution and a channel shuffle mechanism, through which the receptive field in specific directions was expanded while the number of model parameters was significantly reduced. An improved residual structure, RESTD (Rep-parameterized spatial-to-depth enhancement), was proposed to enhance the capture capability for subtle edge features by optimizing residual connections and spatial feature transmission paths. In addition,an anisotropic-aware decoupled detection head, StripHead (StripConv-head), was constructed to strengthen the localization accuracy and regression efficiency for elongated defect targets. Experimental results on the self-built tactile paving defect dataset showed that the mAP50 of IDRS-YOLO reached 80.3%, which was 2.2 percentage points higher than that of the baseline YOLO11n. The recall was improved by 2.9 percentage points to 71.5%, while the number of parameters and computational complexity were reduced by 14.8% and 9.5%, respectively. The results demonstrated that the improved algorithm maintained high real-time performance while achieving a good balance between detection accuracy and model compactness, verifying its engineering practical value for mobile inspection and edge computing platforms.