基于空洞空间金字塔池化的雾天图像语义分割

针对雾天图像语义分割中分割精度不高的问题,基于空洞空间金字塔池化、Xception模块和残差网络,提出一种雾天图像语义分割算法.该算法一方面使用空洞空间金字塔池化和残差网络,以多个采样率的并行卷积以及卷积核大小为1×1的卷积对输入图像进行多尺度的上下文特征提取;另一方面,采用解码器结构使用预训练后的Xception模块对提取的特征进行分类,得到每一个像素的预测结果,用于增强分割边界的细化得到边界精细的分割结果.实验表明,所提算法在foggy cityscapes数据集上的平均交并比为73.03％、73.81％与74.50％,分割性能良好.

Semantic segmentation of foggy images based on atrous spatial pyramid pooling

Aiming at the problem of low segmentation accuracy in semantic segmentation of foggy images, this paper proposed a semantic segmentation method of foggy images that based on atrous space pyramid pooling, Xception module and residual network. In this algorithm, on the one hand, it used the atrous space pyramid pooling and residual network to extract the multi-scale context features of the input image by the parallel convolution of multiple sampling rates and the convolution kernel size of 1×1. On the other hand, it used the decoder structure to classify the extracted features with Xception module after pretraining, and the prediction results of each pixel enhanced the refinement of the segmentation boundary and got the segmentation result with fine boundary. Experimental results show that the average intersection and union ratio of the proposed algorithm on foggy cityscapes dataset are 73.03%, 73.81% and 74.50%, and the segmentation performance is good.