Abstract: Aiming at the problem that the quality of road image was destroyed under foggy conditions, which made the detection difficulty, an improved YOLOX detection algorithm DSACNet was proposed. YOLOX was used as the detection module in DSACNet and a reconstruction module similar to encoder-decoder was designed. By utilizing the feature reconstruction module to share the clean features generated by the reconstruction network with the detection network, the detection network could better learn the hidden features in foggy images, thereby helping DSACNet improve its detection capability under adverse weather conditions. In addition, the self_attention mechanism and self-calibration convolution were introduced to improve the feature extraction ability, and focal loss was added to solve the imbalance problem between positive and negative samples in the target detection task. The results show that the proposed DSACNet adopts an end-to-end training method, which can perform foggy image restoration and target detection at the same time and use a joint optimization strategy to combine the above two, so that the target detection network can obtain the hidden features explored by the recovery network, which is more conducive to road defect detection in foggy conditions. Compared to the original model YOLOX, the mAP of the proposed method reaches 93.5%, with an increase of 14%, and is superior to those of other mainstream target detection algorithms, which meets the accuracy requirements of road defect detection.

Key words: highway engineering; computer technology; road defect detection; self_attention mechanism

摘要： 针对在雾天条件下道路图像质量被破坏，使得检测困难的问题，提出了改进YOLOX的检测算法DSACNet。DSACNet采用YOLOX作为检测模块，设计了一个类似编码-解码（encoder-decoder）的重构模块，利用特征重构模块与检测网络共享重构网络产生的干净特征，使检测网络能够更好地学习到雾天图像中的隐藏特征，从而帮助DSACNet提高在恶劣天气条件下的检测能力；引入了自注意力机制、自校准卷积来提高特征提取能力，加入focal loss解决目标检测任务中正负样本之间的不平衡问题。结果表明：提出的DSACNet采用端对端的训练方式能够同时执行雾天图像恢复和目标检测，并采用联合优化的策略将二者进行联合，让目标检测网络能够获得恢复网络探索的隐藏特征，更利于雾天情况下的道路缺陷检测； 相较于原始模型YOLOX，平均精度均值达到93.5%，提高了14%，并且优于其他主流的目标检测算法，满足了道路表面缺陷检测对精度的要求。

关键词: 道路工程；计算机技术；道路缺陷检测；自注意机制