Automated road defect detection is a critical technology for ensuring traffic safety and infrastructure maintenance. To address the limitations of existing algorithms in identifying slender cracks, locating irregular defects, and balancing real-time performance with accuracy, this paper proposes DSL-YOLO (Deformation-Sensitive Learning, You Only Look Once), a road defect detection model based on deformation-sensitive learning. First, a Dynamic Snake Convolution module (DySnakeConv) is designed, which adaptively adjusts sampling points along crack morphologies through asymmetric convolutional kernels combined with learnable offsets. Second, a Large Separable Kernel Attention mechanism (LSKA) is introduced, employing a one-dimensional decomposition strategy to expand the receptive field and enhance the perception of complex defects. Third, a Location Quality Estimator (LQE) is proposed, which dynamically evaluates localization quality by analyzing the statistical characteristics of bounding box regression distributions. On a self-constructed dataset of 3,321 images, DSL-YOLO achieves a mean Average Precision (mAP@0.5) of 71.3%, representing a 7.6 percentage point improvement over the YOLOv8m baseline, with a precision of 73.2%, a recall of 68.4%, and a real-time detection speed of 86 Frames Per Second (FPS). Ablation studies confirm the synergistic effects of the three proposed modules, and heatmap analysis reveals the mechanism by which the model accurately focuses on defect regions. This study provides a solution for road defect detection that effectively balances accuracy and efficiency.