特征增强U形网络的图像语义分割

复杂场景语义分割任务是对场景图像逐像素进行分类并标记.图像中目标种类多,尺度多样的特点给分割任务增加了难度,提出了特征增强U形卷积神经网络(feature enhanced U shape networks,FEUNet)是一种改进的编码器加解码器的结构,编码阶段引入局部特征增强模块(local feature enhanced,LFE)提取局部感知特征来改善非显著目标的分割效果;考虑到神经网络深层和浅层之间特征表达的差异,在解码阶段利用全局池化方法(global pooling)设计全局特征增强模块(global feature enhanced,GFE),实现选择性地从深层特征图提取上下文信息作为对浅层特征图的指导,改善深层和浅层特征图的融合,保证同类像素预测的一致性.采用CamVid和Cityscapes数据集进行试验,模型mIOU测评值分别达到64.5％和73.2％,对比其他主流语义分割算法,该方法在分割性能和模型体积上具有一定竞争力.

Semantic segmentation of images based on feature enhanced U-shape convolutional network

The complex scene semantic segmentation task is to classify and label the scene image pixel by pixel. The large number of targets with diverse scales in the image make the task difficult. This method proposes a feature enhanced U-shaped convolutional neural network (FEUNet), which is an improved encoder and decoder structure. The local feature enhancement module (LFE) is introduced in encoder to extract local perceptual features to improve the segmentation of non-significant targets. Considering the difference in feature expression between the deep and shallow layers of the network, the global Pooling is used to design the global feature enhancement module (GFE) in decoder, to selectively extract context information from deep feature maps as a guidance to shallow feature maps. GFE model can improve the fusion of deep and shallow feature maps by ensuring the consistency of pixel prediction. Finally, experiments are performed on the datasets CamVid and Cityscapes to prove the effectiveness of the modules, and our FEUNet achieves a good segmentation result.