基于改进SegNet的沥青路面病害提取与分类方法

针对现有SegNet算法难以精确区分裂缝和灌封裂缝等具有相似特征的沥青路面病害的问题，提出了基于改进SegNet网络的沥青路面病害提取方法。针对道路标线和光照不均匀等导致路面病害图像质量差异化的因素，本研究在去除道路标线的基础上，运用带色彩恢复的多尺度视网膜增强算法，降低道路标线和光照对图像质量的影响以及增强路面病害图像的对比度、色调和亮度，提高病害的识别精度；为了充分利用图像的上下文信息，解决SegNet网络对细微病害分割效果不佳的问题，引入残差神经网络（ResNet）作为编码器，并对解码器中每个上采样产生的特征图拼接2个分别由卷积层（1×1的卷积核）和空洞卷积层从对应的编码器中获取的尺度相同的特征图；运用形态学闭运算连接分割结果中不连续的裂缝。为了验证改进算法的有效性，将其与典型的语义分割方法（SegNet和BiSeNet）在测试集上进行测试和性能对比。研究结果表明，3种方法的平均交并比（MIoU）和F1分数（F1-score）分别为（82.4%，98.9%），（69.4%，94.0%），（80.5%，98.1%）；利用这3种方法对甘肃省部分路段路面病害的提取效果进行对比测试，提出方法的裂缝漏检率和误检率分别为2.91%，1.94%，优于SegNet（10.68%，14.56%）和BiSeNet（6.80%，12.62%）。本研究所提方法能够更精确地提取和区分沥青路面裂缝和灌封裂缝。 

A Method for Detecting and Differentiating Asphalt Pavement Distress Based on an Improved SegNet Algorithm

The existing SegNet methods are unable to accurately differentiate asphalt pavement distress with similar characteristics, such as cracks and sealed cracks. To solve this problem, a method of detecting asphalt pavement distress based on an improved SegNet network is proposed. In order to remove the negative impact of road markings and uneven illumination onto image quality of road surface and subsequent failure detection, a multi-scale Retinex algorithm with color restoration(MSRCR)is used to reduce the impact of road marking and uneven lighting on image quality. Through enhancing the contrast, hue and brightness of the images for road surface, the accuracy of distress recognition is improved. In order to fully use of the contextual information of the image, and overcome the issue with the SegNet network of being ineffective in segmenting and identifying subtle diseases, a residual neural network(ResNet)is introduced as the encoder, and two feature maps with a same scale obtained by a convolutional layer with a 1×1 kernel and a dilated convolutional layer with different dilation rates are fused for each feature map, generated by up-sampling in the decoder. And a closed, morphological operation is used to connect discontinuous cracks. To verify the effectiveness of the improved algorithm, it is compared with the classic semantic segmentation methods(such as SegNet and BiSeNet)over the test sets. The average intersection over Union(MIoU)and F1 score are(82.4%, 98.9%), (69.4%, 94.0%)and(80.5%, 98.1%), respectively. The three methods are compared in terms of their extraction efficiency in identifying pavement diseases using the pavement images collected at several freeway sections in the Gansu Province. The misdetection rate and false detection rate of cracks of the proposed method are 2.91%, 1.94%, respectively, which are much better than those of the SegNet(10.68%, 14.56%)and BiSeNet(6.80%, 12.62%). The above results show that the proposed method can be used to extract and identify asphalt pavement cracks and sealed cracks with a higher accuracy.