YOLO-SFT: Road Damage Detection Algorithm Based on Feature Diffusion

Abstract

Pavement damage detection is an important research area in road maintenance and traffic safety, but traditional detection methods have shortcomings such as low accuracy and poor real-time performance. A pavement damage detection algorithm, You Only Look Once-spatial feature transformation (YOLO-SFT), based on feature diffusion is proposed in this paper to improve the detection accuracy and efficiency. First, a new module, StarNet-context anchor attention (Star-CAA), is designed to replace the C2f of the backbone part, which enhances the ability of feature extraction and gradient flow and optimizes the detection performance and generalization ability of the model. After that, a new pyramid network, focusing diffusion cross stage (FDCS), is independently developed to optimize the neck part. Through the unique feature-focusing diffusion mechanism, features with rich contextual information are diffused to various detection scales. Finally, the detection head part is redesigned to propose a new efficient detection head, task align dynamic (TAD), which obtains joint features by learning task interaction features from multiple convolutional layers. It strengthens the accuracy and real-time performance of pavement damage detection. Experimental results show that the F1 score of YOLO-SFT is improved by 4.0% and the mean average precision (mAP) is enhanced by 5.6%. In addition, the computational parameters of YOLO-SFT are reduced by 16.9%, the model size is reduced by 14.9%, and the running speed reaches 64.5 frames per second (FPS). The proposed algorithm has good application prospects and provides an effective solution for pavement distress detection in intelligent transportation systems.