Abstract: To enhance the precision of road defect detection, a lightweight, real-time Transformer-based detection network incorporating dynamic deformable convolutions has been proposed. Initially, a multi-path coordinate attention mechanism module is developed and seamlessly integrated with deformable convolution modules. This integration creates a dynamic deformable convolution module with variable receptive fields, capable of handling road defects of diverse shapes. Subsequently, a deformable attention mechanism is introduced into the intra-scale feature interaction module, enhancing the network's ability to capture and extract critical information from images. Finally, to achieve an overall reduction in network complexity, lightweight modules are employed in the cross-scale feature fusion stage. Experimental evaluations conducted on the Global Road Defect Detection Challenge (GRDDC2020) dataset compared this network against eight others, including YOLOv5-m. The results indicate that the proposed network not only delivers superior detection performance but also exhibits better efficiency in terms of its parameter count and computational requirements. Specifically, the network achieves a mean Average Precision (mAP) of 61.1%, with 19.6 million parameters and 50.2 GOPS (billion floating-point operations per second). Additionally, it operates at a detection speed of 42.3 frames per second. This makes it effective for detecting road defects and providing essential data for road maintenance efforts.