基于级联解码与边缘感知的RGB-D图像语义分割算法

RGB-D Semantic Segmentation Based on Cascade Decoder and Boundary Perception

随着深度传感器技术的发展,由于其采样得到的深度图像含有场景的空间几何信息,有助于提升语义分割精度,因此RGB-D图像的语义分割受到了广泛关注,然而现有的基于编码器-解码器架构的双流RGB-D语义分割网络仍存在分割不够精细和边缘区域分割精度低的问题。针对上述问题,论文提出了一种基于级联解码与边缘感知的RGB-D语义分割方法。首先设计了级联解码器,采用自顶向下的解码方式逐级整合多尺度信息,细化分割结果;随后将边缘检测任务以多任务学习的方式引入RGB-D语义分割中,增大卷积神经网络所提特征对边缘像素的关注度,提升分割边缘的精度。在NYUDv2数据集上的实验结果表明,所提出的语义分割方法优于其他对比方法。

Depth information is proved to be useful cues in the semantic segmentation of RGB-D images for providing a geometric counterpart to the RGB representation.However,the existing dual-stream encoder-decoder RGB-D semantic segmentation methods still have the problem of unsatisfactory segmentation detail and low boundary accuracy.Aiming at these problems,a RGB-D semantic segmentation method based on cascade decoder and boundary perception is proposed.First of all,a cascade decoder is designed to integrate multi-scale information by using a top down decode method,which refines the segmentation details.Then,a boundary detection sub-network is designed in the way of multi-task learning,which makes the network pay more attention to pixels that located on boundaries,thereby improving the accuracy of boundaries.The experimental results on NYUDv2 dataset show that the proposed algorithm achieves better results than other methods.