| 邹凯鑫,张自嘉,孙伟,付锦燚.改进U型网络的路面缺陷图像分割算法[J].电子测量与仪器学报,2024,38(8):15-25 |
| 改进U型网络的路面缺陷图像分割算法 |
| Improving the road defect image segmentation algorithm of U-Net |
| |
| DOI： |
| 中文关键词: 深度学习 语义分割 路面缺陷 U-Net 注意力机制 |
| 英文关键词:deep learning semantic segmentation pavement cracks U-Net attention mechanism |
| 基金项目:国家自然科学基金（62376128）项目资助 |
|
|
| Author | Institution |
| Zou Kaixin | School of Automation, Nanjing University of Information Science and Technology, Nanjing 210044, China |
| Zhang Zijia | 1.School of Automation, Nanjing University of Information Science and Technology, Nanjing 210044, China;
2.Jiangsu Collaborative Innovation Center on Atmospheric Environment and Equipment Technology,
Nanjing University of Information Science and Technology, Nanjing 210044, China |
| Sun Wei | 1.School of Automation, Nanjing University of Information Science and Technology, Nanjing 210044, China;
2.Jiangsu Collaborative Innovation Center on Atmospheric Environment and Equipment Technology,
Nanjing University of Information Science and Technology, Nanjing 210044, China |
| Fu Jinyi | School of Automation, Nanjing University of Information Science and Technology, Nanjing 210044, China |
|
| 摘要点击次数: 71 |
| 全文下载次数: 105 |
| 中文摘要: |
| 针对路面缺陷图像存在对比低以及其复杂的拓扑结构,现提出的大部分分割算法,在捕获感受野和提取路面缺陷特征时还存在很大的不足。因此本文提出改进U Net的路面缺陷图像分割算法。首先,在经典的U-Net的卷积块中提出SN-Disout残差块,增强模型对过拟合的鲁棒性;其次,在编码器与解码器之间引入循环十字交叉模块,增强模型在特征图中不同位置之间捕获特征的能力,对缺陷的边界进行更准确的建模;最后,在解码器中引入空间通道挤压与激励模块,这使得网络能够更加专注于重要的特征,同时减少不相关或噪声特征的依赖;并将位置感知多头注意力加入模型的颈处,进一步有助于模型更好地理解和利用输入数据的内在关系,从而提高模型的性能和表现能力,并且使用混合损失函数Dice+BCE取代单一的损失函数。该算法在Crack500图像数据集上的交并比与F1分别达到了60.13%和75.22%,均超过U-Net、DeepLabV3+、PSPNet、TransU-Net、UNet++等主流语义分割网络。实验结果表明,该算法能有效改善网络的预测精度与细小目标的分割结果,在保证分割精度的情况下也满足了实时性的要求。 |
| 英文摘要: |
| In view of the low contrast and complex topological structure of pavement defect images, most of the currently proposed segmentation algorithms still have great shortcomings in capturing the receptive fields and extracting pavement defect features. Therefore, this article proposes an improved U-Net road defect image segmentation algorithm. First, the SN-Disout residual block is proposed in the classic U-Net convolution block to enhance the model’s robustness against overfitting. Secondly, a criss-cross module is introduced between the encoder and the decoder to enhance the model’s ability to capture features between different positions in the feature map and more accurately model the boundaries of defects. Finally, the spatial channel squeeze and excitation module is introduced in the decoder, which enables the network to focus more on important features while reducing the dependence on irrelevant or noisy features; position-aware multi-head attention is added to the neck of the model to further helps the model better understanding and utilizing the internal relationship of the input data, thereby improving the performance and performance capabilities of the model, and using the hybrid loss function Dice+BCE to replace a single loss function. The intersection ratio and F1 of this algorithm on the Crack500 image data set reached 60.13% and 75.22% respectively, both exceeding mainstream semantic segmentation networks such as U-Net, DeepLabV3+, PSPNet, TransU-Net, and UNet++. Experimental results show that this algorithm can effectively improve the prediction accuracy of the network and the segmentation results of small targets, and it also meets the real-time requirements while ensuring segmentation accuracy. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
