Road Disease Detection based on Improved RTMDet

In order to better maintain road diseases, efficient and accurate detection is required. It is feasible to use computer vision instead of manual and sensor for automatic road diseases detection accompanied by the development of deep learning. The actual road background is complex, and the types of diseases are diverse, and the detection and classification of road diseases still face great challenges. We propose a road diseases detection algorithm based on RTMDet to resolve these issues, and it meets real-time and high precision requirements. First, we reconstructed the detector head, combined with an efficient attention mechanism, and reduced the computational load and model complexity of the algorithm. Then we redesigned the algorithm's Loss function, and used the TAL label assign strategy, Varifocal loss, and Distribution Focal Loss to improve the association for classification and regression tasks, and effectively enhance the precision of the algorithm's classification and bounding box.