基于RGB-D图像的室内场景语义分割网络优化

Optimization of Indoor Scene Semantic Segmentation Network Based on RGB-D Image

针对卷积神经网络在室内场景的图像语义分割中难以取得较高的分割精度,提出一种基于RGB-D图像的室内场景语义分割网络。该网络采用分别训练逐渐融合的方式对原始数据进行处理,并在解码阶段加入强化监督模块,有效提高语义分割的准确率;同时引入反残差的解码方法和跳跃结构降低信息损失。实验结果表明:REDNet的像素精度达80.9%,平均精度达58.4%,区域交集精度达46.9%,这些分割精度均高于FCN-32s,FCN-16s,SegNet,Context-CRF,FuseNet,RefineNet等常用语义分割网络。

In recent years,convolution neural network has been widely used in image semantic segmentation and achieved great success.We propose a semantic segmentation network for indoor scenes based on RGB-D images:REDNet.This network model uses separate training and gradual fusion to process the original data,and adds an forced supervision module in the decoding phase,which effectively improves the accuracy of semantic segmentation.At the same time,the anti-residual decoding method and jump structure are introduced to reduce the information loss.The experimental results show that REDNet has 80.9%pixel accuracy,58.4%mean accuracy and 46.9%IoU,which are higher than FCN-32s,FCN-16s,SegNet,Context-CRF,FuseNet,RefineNet and other semantic segmentation networks.