基于卷积神经网络的多尺度图像语义分割

Multi-scale Image Semantic Segmentation Based on Convolutional Neural Network

针对现有的图像语义分割算法存在的因细节信息丢失导致分割效果不佳的问题,论文提出一种基于DeepLabV3+的改进算法。论文的深度学习网络分为编码器和解码器模块,在编码器模块使用改进的ResNet_101和空洞空间金字塔池化结构提取多尺度特征,在解码器模块结合多个输出,并且融合图像低层信息,解决目标细节丢失问题。为验证论文算法的有效性,在PASCAL VOC 2012数据集上进行实验,结果表明,改进后的算法在物体细节处理方面得到了良好效果,性能方面有所提升。

To solve the problem of poor segmentation results due to the loss of detailed information in existing image semantic segmentation algorithms,a multi-scale semantic segmentation model which based on improved DeepLabV3+is proposed.The deep learning network in this paper consists of encoder and decoder modules.In the encoder module,the improved ResNet_101 and the atrous spatial pyramid pooling are used to extract multi-scale features.The decoder module combines the outputs of multiple decod⁃ers and merges the lower layers of the picture to solve the problem of missing target details.The experimental results on PASCAL VOC 2012 data set show that the improved method has achieved better results in object detail processing compared to existing meth⁃ods.