现代电子技术

针对路面裂缝图像背景噪声复杂、裂缝形态复杂和误分割严重的问题，文中提出一种基于U型网络改进的路面裂缝分割算法（DCM⁃Net）。DCM⁃Net采用双编码器设计，新增加的支路减轻了由于一条支路简单堆叠卷积池化造成的信息丢失；在原有的跳跃连接中增加CoTAttention，旨在加强低级语义信息中的重要特征，减轻由于背景噪声以及车道线和井盖等杂物产生的影响，增强有用信息的特征表达能力；对原编码器中的卷积模块进行重新设计，引入膨胀卷积增大感受野，采取多维特征提取的策略，提高模型在不同裂缝形态下的特征提取能力。对比实验结果表明，在自建数据集CrackNew上，DCM⁃Net在Dice、平均交并比、准确率、召回率和F1 上相较于UNet分别提升了6.3%、5.7%、5.4%、1.8%、5.3%。同时，优于其他主流分割模型，在Crack500和Gaps384两个公开数据集上各个指标仍保持领先，在DeepCrack数据集上进行了消融实验，证明了各模块的有效性。对比其他分割模型，DCM⁃Net提高了路面裂缝的分割精度，该模型可适用于复杂环境下的道路裂缝分割。

An improved U⁃Net⁃based pavement crack segmentation algorithm named DCM⁃Net is proposed in response to the challenges posed by complex background noise, intricate crack patterns, and severe mis⁃segmentation in pavement crack images. A dual⁃encoder design is adopted in the DCM⁃Net, and the additional branch mitigates information loss caused by the simple stacking of convolution and pooling in a single branch. CoTAttention mechanism is incorporated into the original skip connections to enhance important features within low⁃level semantic information and mitigate the impact of background noise, lane markings, manhole covers, and other obstructions, so as to enhance the feature representation of useful information. The convolution module in the original encoder is redesigned. The dilated convolution is introduced to increase the receptive field. The multi⁃dimensional feature extraction strategy is adopted to improve the feature extraction ability of the model across various crack morphologies. The comparative experimental results show that on the self ⁃ built dataset CrackNew, the Dice, mean intersection over union (mIoU), precision, recall rate and F1 of the DCM⁃Net are improved by 6.3%, 5.7%, 5.4%, 1.8% and 5.3%, respectively, in comparison with those of the UNet. Meanwhile, it is superior to the other mainstream segmentation models. On the publicly available datasets Crack500 and Gaps384, the DCM ⁃ Net maintains leading performance across all metrics. Ablation experiments conducted on the dataset DeepCrack confirm the effectiveness of each module of the DCM ⁃ Net. In comparison with the other segmentation models, the DCM ⁃ Net enhances the segmentation precision for pavement cracks significantly. To sum up, the model can be applied to road crack segmentation in complex environment.