基于改进U型神经网络的路面裂缝检测方法

针对传统的裂缝分割算法难以识别狭窄裂缝且分割边缘不精准，从而造成识别精度较低的问题，研究了基于改进U型神经网络（Unet）的路面裂缝检测方法。由于传统Unet特征提取网络是层次较浅的浅层神经网络，难以提取更复杂的裂缝特征信息，故本文以牛津大学视觉几何组网络（VGG16）作为传统Unet的特征提取网络，提高网络的裂缝特征提取能力；为抑制高低阶特征融合时产生的无用特征，本文在模型解码部分添加压缩与激励单元（SE block），构建裂缝注意力单元，使得网络可以关注不同通道下的裂缝特征，建立了基于SE block和VGG16的改进Unet网络（SE-VUnet）。研究采用迁移学习的方法，将在ImageNet上预训练好的VGG16网络权重迁移到裂缝检测中。通过挑选Crack500数据集，并使用摄像头采集图片构建1 600张路面裂缝数据集，再次训练SE-VUnet模型，获得裂缝区域分割结果。以查准率（precision）与查全率（recall）的加权调和平均值F1和雅卡尔（Jaccard）相似系数作为量化评价指标。将SE-VUnet分别与Unet、SOLO v2、Mask R-CNN以及Deeplabv3+进行分割效果和实时性对比。研究结果表明:SE-VUnet模型的综合F1和雅卡尔系数分别为0.840 3和0.722 1，相比于Unet分别高出了1.04%和1.51%，且均高于其他3种对比模型；SE-VUnet的单帧图片预测时间为89 ms，在分割效果提升明显的情况下仅比Unet慢5 ms，优于其他模型。 

A Detection Method for Pavement Cracks Based on an Improved U-Shaped Network

Due to the traditional crack segmentation algorithm is difficult to identify narrow cracks and the segmentation edge is not accurate. This paper proposes a pavement crack detection method based on improved U-Shaped Network (Unet) to increase detection accuracy. Since traditional Unet is a type of"shallow"neural network, it is not good for extracting complex crack features. The Oxford University Visual Geometry Group Network (VGG16) is therefore used for feature extraction, in order to improve the accuracy of crack feature extraction. In addition, the fusion of high- and low-order features generate several useless features. The compression and excitation unit (SE block) is added to the decoding part of the model to develop a crack attention unit which allows the network to focus on the crack features under different channels. Moreover, an improved Unet is proposed by combining SE block with VGG16 (SE-VUnet). In addition, a transfer learning method is used to transfer the pre-trained VGG16 network weight on ImageNet for crack detection. By selecting the Crack500 data set and using the camera to collect images to develop1600 pavement crack data sets, the SE-VUnet model is trained again to obtain the crack segmentation results. The weighted harmonic mean F1 of Precision and Recall and Jaccard similarity coefficient are used as quantitative evaluation indicators. The segmentation effect and real-time performance of SE-VUnet are compared with Unet and three other representative models. Study results show that the comprehensive F1 and the Jaccard coefficient of SE-VUnet model is 0.840 3 and 0.722 1, which is 1.04% and 1.51% higher than Unet respectively, as well as other three comparison models. The time for the SE-VUnet to screen a single-frame image is 89 ms, which is only 5ms slower than the Unet but with a significant improvement over the crack segmentation and detection process.