基于FRRN注意力监督的沥青路面积水区域分割

为实现不同光照与气候条件下的沥青路面积水区域自动分割，在现有全分辨率残差网络(FRRN)的上采样层设计了一种注意力监督机制；加入积水区域注意力模块(PAAM)与深度监督模块，建立了一种含积水区域注意力监督模型(PAAM-FRRN)；利用最大池化与上采样构建编码器-解码器结构，在全局尺度下捕获了积水的视觉特征，在上采样层引入了注意力监督机制，针对积水区域进行上采样并融合不同网络层之间的特征，最小化网络各层损失函数，优化网络总体最终损失；采集不同光照与气候条件下的1 770个(弱光750个、强光740个、雨天280个)沥青路面积水图像进行五折交叉验证试验，获得了积水区域分割结果；以人工标注结果作为真值，利用Dice相似系数(DSC)、Jaccard相似系数(JSC)、精确率、召回率与豪斯多夫距离(HD95)作为量化评价指标，将提出的模型分别与FRRN和其他7种有代表性的传统模型进行了分割效果对比。研究结果表明:提出的模型在DSC、JSC、精确率与召回率所取得的均值分别为0.91、0.86、0.92和0.93，相比FRRN分别提高了3.41%、6.17%、2.22%和4.49%，且均高于传统7种对比模型；在DSC、JSC、精确率以及召回率所取得的标准差分别为0.12、0.15、0.11和0.12，与FRRN相比分别降低了20.00%、16.67%、21.43%和25.00%，且均低于传统7种对比模型；其HD95为38.56，均低于其他对比模型，提出的模型能够实现对不同光照与气候条件下的沥青路面积水区域的准确、有效分割。 

Puddle area segmentation of asphalt pavements based on FRRN attention and supervision

To realize the automatic segmentation of puddle area of asphalt roads under various brightness levels and climate conditions, an attentional supervision mechanism was developed on the upsampling layer of existing full resolution residual network (FRRN). A puddle area attention module based on the full-resolution residual network (PAAM-FRRN) model was established by adding the puddle area attention module (PAAM) and the deep supervision module. An encoder-decoder structure was constructed by max-pooling and upsampling to capture the visual characteristics of puddle in a global scale. An attentional supervision mechanism was introduced in the upsampling layer to conduct the upsampling for the puddle area and fuse the features of different network layers to minimize the loss function of each network layer and optimize the overall final loss of the network. A total of 1 770 images (750 under low light, 740 under strong light and 280 under rainy weather) of asphalt pavement with puddle were collected for the five-fold cross-validation test, and the segmentation results of the puddle area were obtained. Manual annotation results were used as truth values, and the dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), precision, recall, and Hausdorff distance (HD95) were used as quantitative evaluation indexes. The segmentation effect of the proposed model was compared with the FRRN and seven other representative traditional models. Research results show that the mean values of DSC, JSC, precision, and recall obtained by the proposed model are 0.91, 0.86, 0.92, and 0.93, respectively, which are 3.41%, 6.17%, 2.22%, and 4.49% higher than those obtained by the FRRN, and are all higher than those obtained by the seven traditional comparison models. Standard deviation values of DSC, JSC, precision, and recall as computed by the proposed model are 0.12, 0.15, 0.11, and 0.12, respectively, which are 20.00%, 16.67%, 21.43%, and 25.00% lower than those of FRRN, and are all lower than those of the seven traditional comparison models. The HD95 of the proposed model is 38.56, which is lower than those of the other comparison models. Therefore, the proposed model can achieve accurate and effective segmentation of puddle areas of asphalt pavements under different brightness levels and climate conditions. 6 tabs, 11 figs, 40 refs.