基于通道注意力机制的RGB-D图像语义分割网络

RGB-D semantic segmentation using channel attention mechanism

针对RGB-D图像的语义分割问题,提出了一种结合通道注意力机制的RefineNet网络。考虑到网络特征图中各个通道重要性的不同,将通道注意力机制分别引入基本RefineNet的编码器和解码器模块,以增强网络对重要特征的学习和关注;同时,使用focal loss函数代替传统的交叉熵损失函数,以处理多类语义分割任务中的类别数量不平衡和难分样本问题。在SUNRGBD和NYUv2数据集上的实验结果表明,相比于最新的主流语义分割网络如Depth-aware、RDFNet和Refinenet,本文网络在保持相近网络参数量和计算量的同时,有效地提高了分割精度,其mIOU分别达到45.7%和49.4%。

This paper present an improved RefineNet network with channel attention mechanism for RGB-D image segmentation.In the proposed semantic segmentation network,considering the different importance of each channel in feature graphs,channel attention mechanism is introduced into the basic RefineNet encoder and decoder,which enforces the network to focus on important features.Moreover,the focal loss function is used to replace the traditional cross-entropy loss function to deal with the imbalance of the category number and the hard samples in multi-class semantic segmentation tasks.The experimental results on SUNRGBD and NYUv2 datasets show,compared to the latest segmentation networks such as Depth-aware,RDFNet and Refinenet,the proposed network can effectively improve the segmentation accuracy while maintaining similar network parameters and computational load,and its mIOU on the two datasets are 45.7%and 49.4%m,respectively.