中图分类号：TP183 文献标志码：A DOI: 10.16157/j.issn.0258-7998.256454
中文引用格式： 韩德强，张洪瑞，杨淇善. 基于改进UNet的沥青道路缺陷检测系统的研究与实现[J]. 电子技术应用，2025，51(11)：1-9.
英文引用格式： Han Deqiang，Zhang Hongrui，Yang Qishan. Research and implementation of an asphalt road defect detection system based on improved UNet[J]. Application of Electronic Technique，2025，51(11)：1-9.

Research and implementation of an asphalt road defect detection system based on improved UNet

Abstract： Aiming at the problems such as the low generalization ability of traditional methods in road defect detection, which are vulnerable to environmental interference, and the high power consumption and low speed when deploying deep learning models on computing platforms, an acceleration and deployment strategy for the semantic segmentation model based on a low-power FPGA (Field-Programmable Gate Array) platform is proposed. Firstly, a multi-source dataset containing road cracks and potholes is constructed, and data augmentation techniques are used to balance the sample distribution. Secondly, channel pruning is carried out separately for the feature extraction network and the upsampling network of the UNet model. Combined with the quantization technique, the model weights are compressed from FP32 (32-bit floating-point) to INT8 (8-bit integer), further reducing the computational load. Finally, the Vitis AI toolchain is utilized to complete the model quantization and compilation, and the model is deployed to the FPGA platform to fully exert its parallel computing capability. The experimental results show that, on the premise of ensuring that the loss of the mean intersection over union (MIoU) is less than 5%, the inference speed of the optimized model reaches 17 ms. The number of model parameters and the computational load are significantly reduced, and the power consumption is remarkably decreased. This method achieves efficient and low-power road defect detection at the edge side, providing a feasible solution for the automated maintenance evaluation of asphalt roads.

Key words : road defect detection；semantic segmentation；model compression；FPGA model deployment