Lightweight Road Defect Detection Method Based on Dynamic Deformable Convolution

To enhance the accuracy of road defect detection while addressing the multifaceted and irregular nature of these defects, a novel lightweight road defect detection method based on dynamic deformable convolution has been proposed. This method follows an encoder-decoder network architecture, utilizing ResNet18 as the backbone network within the residual framework. Innovatively, it incorporates a Multi-Path Coordinate Attention Mechanism (MPCA), which is deeply integrated with the Deformable Convolutional Module (DCN), resulting in a more advanced dynamic deformable convolution module. This design allows for the creation of a residual network capable of adapting its receptive field to effectively handle road defects of varying sizes and shapes. Moreover, within the Adaptive Intra-Scale Feature Interaction module (AIFI), a deformable attention mechanism is introduced to bolster the ability to capture and extract critical target information from images. Additionally, lightweight modules have been integrated into the Cross-Scale Feature Fusion Module (CCFM) to facilitate the lightweight design of the neck network. The model was benchmarked against eight other methods, including YOLOv5-m, on the Global Road Defect Detection Challenge (GRDDC2020) dataset. The results demonstrated that this model not only provides superior detection performance but also exhibits improved metrics in terms of model parameters and computational complexity. Specifically, the model achieved a mean Average Precision (mAP) of 61.1%, had 19.6 million parameters (Params), 50.2 billion floating-point operations per second (GFLOPs), and operated at a frame rate of 42.3 milliseconds (FPS). These attributes collectively attest to its effectiveness in accurately detecting road defects and provide valuable insights for pavement maintenance tasks.