一种多尺度CNN的图像语义分割算法

Semantic Segmentation with Multi-scale Convolutional Neural Network

针对目前多数图像语义分割方法需要人工设计图像特征的问题,借助卷积神经网络(Convolutional Neural Network,CNN)自动学习得到图像特征的优势,并综合考虑CNN的网络输入和物体上下文关系对图像语义分割结果的影响,以超像素为基本处理单元,结合多尺度技术和CNN网络设计了一种面向图像语义分割的多尺度CNN模型,并详细分析了该模型的结构以及模型推断。实验验证了所提出方法的有效性。

Most works of semantic segmentation still rely on human-crafted features.In contrast,convolutional neural network can automatically learn features at multiple levels from images,without depending completely on complicated humancrafted features and data reconstruction.By virtue of the advantages of CNN,an improved approach for semantic segmentation based on multi-CNN was presented,which considered the input of CNN and contextual information.Superpixels of the image were first produced by over-segmentation.Then,combined with the multi-scale technique and a CNN network,a multi-CNN model for semantic segmentation was designed.The model architecture and inference were analyzed.The experiments pertaining to real images demonstrated that the multi-scale technique and CNN is effective and the proposed approach improves the accuracy of semantic segmentation.