基于上下文情景结构的图像语义分割

Semantic segmentation of images based on contextual structure

语义分割的目标是对图像进行像素级分类并分割为表示不同语义的区域,以便实现对更复杂场景的分析及理解。考虑到更大的卷积核对目标的定位及分割都有促进作用,故提出的网络里使用了一种基于上下文情景结构层(contextual structure)的卷积神经网络,既增大了感受野的范围,同时解决了增大传统卷积核带来大量参数的问题。自然图像中的目标往往具有不同的尺度和纵横比,提取丰富的细节信息和上下文信息对于像素级的预测至关重要。提出的网络同时实现了多尺度特征的提取,从规模较小区域到规模较大区域,再到完整目标,可以有效提取局部信息和全局信息,达到分割多尺度目标的效果。实验中使用PASCAL VOC 2012数据集对提出的方法进行评测,在综合考虑算法复杂度以及运行时间效率的基础上,提出算法取得了更好的实验结果。

The goal of semantic segmentation is to classify pixel-level images and divide pixels into regions which represent different semantics in order to implement analysis and understanding of complex scene.Considering that the large convolutional kernel has a promoting effect on positioning and segmentation of targets,the proposed network uses a contextual structural convolutional neural network which increases the receptive fields,and at the same time solves the problems that traditional convolution kernels bring a large number of parameters.Targets in natural images often have different scales and aspect ratios.It is very important to extract detailed and contextual information for pixel-level prediction.The proposed network also extracts multi-scale features,from the smaller area to larger area,and then to the complete target.The extracted information enables the network to achieve the multi-scale segmentation.In the experiment,PASCAL VOC 2012 dataset was used to evaluate the method and get better results on the basis of the structural complexity and time efficiency.