【摘要】 传统深度学习语义分割算法中存在特征细节信息丢失以及像素关联度降低的问题，导致图像边缘出现误分割、分割边界不连续。由此在复用底层特征以及捕获多尺度语义信息的基础上提出一种改进型DeepLabv3图像语义分割算法。依据底层特征有助于还原图像细节，构建轻量级并行卷积神经网络(CNN)处理残差网络浅层输出的特征图，获取更多图像空间信息；采用像素级特征融合将获取的图像空间信息与深层网络特征进行融合，并利用金字塔池化模块获取融合特征图的多尺度语义信息；然后利用双线性插值获取图像分割结果。实验结果表明，所提方法可以有效改善语义分割中出现的误分割、分割边界不连续的现象。更多还原

【Abstract】 Traditional deep learning semantic segmentation algorithms have the problems of loss of feature detail information and reduced pixel relevance, resulting in false segmentation of image edges and discontinuous segmentation boundaries. Therefore, an improved DeepLabv3 image semantic segmentation algorithm is proposed on the basis of reusing the underlying features and capturing multi-scale semantic information. According to the fact that the underlying features help restore the image details, a lightweight parallel convolutional neural network is constructed to process the feature map of the shallow output of the residual network to obtain more image spatial information. Pixel-level feature fusion is used to fuse the acquired image spatial information with deep network features, and the pyramid pooling module is used to obtain the multi-scale semantic information of the fused feature map, then the bilinear interpolation is used to obtain the image segmentation result. Experimental results show that the proposed method can effectively improve the phenomenon of false segmentation and discontinuity of segmentation boundaries in semantic segmentation.更多还原