基于注意力机制的轻量级RGB-D图像语义分割网络

目的针对卷积神经网络在RGB-D(彩色-深度)图像中进行语义分割任务时模型参数量大且分割精度不高的问题,提出一种融合高效通道注意力机制的轻量级语义分割网络。方法文中网络基于RefineNet,利用深度可分离卷积(Depthwiseseparableconvolution)来轻量化网络模型,并在编码网络和解码网络中分别融合高效的通道注意力机制。首先RGB-D图像通过带有通道注意力机制的编码器网络,分别对RGB图像和深度图像进行特征提取;然后经过融合模块将2种特征进行多维度融合;最后融合特征经过轻量化的解码器网络得到分割结果,并与RefineNet等6种网络的分割结果进行对比分析。结果对提出的算法在语义分割网络常用公开数据集上进行了实验,实验结果显示文中网络模型参数为90.41 MB,且平均交并比(mIoU)比RefineNet网络提高了1.7%,达到了45.3%。结论实验结果表明,文中网络在参数量大幅减少的情况下还能提高了语义分割精度。

Lightweight Semantic Segmentation Network for RGB-D Image Based on Attention Mechanism

The work aims to propose a lightweight semantic segmentation network incorporating efficient channel attention mechanism to solve the problem of large number of model parameters and low segmentation accuracy when Convolutional Neural Network performs semantic segmentation in RGB-D images. Based on RefineNet, the network model was lightened by Depthwise Separable Convolution. In addition, an efficient channel attention mechanism was applied to the encoding network and the decoding network. Firstly, the features of RGB image and depth image were extracted by the encoder network with channel attention mechanism. Secondly, the two features were fused in multiple dimensions by the fusion module. Finally, the segmentation results were obtained by the lightweight decoder network and compared with the segmentation results of 6 networks such as RefineNet. The proposed algorithm was tested on public datasets commonly used in semantic segmentation networks. The experimental results showed that the parameters of the proposed network model were only 90.41 MB, and the mIoU was 1.7% higher than that of RefineNet network, reaching 45.3%. The experimental results show that the proposed network can improve the precision of semantic segmentation even when the number of parameters is greatly reduced.