We proposed a road defect detection method that applies deformable convolution and context enhancement algo- rithms. Our method focuses on texture feature extraction and fusion to enhance the defect detection capability of the model. Our method uses deformable convolution to effectively adapt the considerable variations in shape and size among various types of road defects. By obtaining more efficent features, our model avoids the region of interest deviating from the ground to appear in the sky. An enhanced contextual module is introduced to facilitate the more efficient fusion of multi-scale texture features. This adaptation tackles the challenge of varying defect sample scales arising from the fluctuating distance between the camera and the ground. Additionally, Our Method also incorporates the CBAM (Convolutional Block Attention Module) to obtain supe- rior feature representation and higher level of critical information perception by considering both spatial positional information and channel-related details. The experimental results show that the mAP is increased to 64.7%, and the number of parameters is reduced to 13.5M. This method not only successfully obtains stronger defect feature expression ability, but also improves the fusion ability of multi-scale features. The proposed model is a high-performance and low-cost road defect detection model.
