基于注意力机制与图卷积神经网络的单目红外图像深度估计

对场景中的物体进行深度估计是无人驾驶领域中的关键问题，红外图像有利于在光线不佳的情况下解决深度估计问题。针对红外图像纹理不清晰与边缘信息不丰富的特点，提出了将注意力机制与图卷积神经网络相结合来解决单目红外图像深度估计问题。首先，在深度估计问题中，图像中每个像素点的深度信息不仅与其周围像素点的深度信息相关，还需考虑更大范围的其他像素点的深度信息，采用注意力机制可以针对这一点有效提取图像的像素级全局深度信息关联。其次，基于深度信息关联得到的特征可以考虑为非欧数据，进一步使用图卷积神经网络（graph convolutional neural network, GCN）来进行推理。最后，在训练阶段将连续的深度估计回归问题转化成分类问题，使训练过程更稳定，降低了网络的学习难度。实验结果表明，该方法在红外数据集NUST-SR上获得了良好的效果，在阈值指标小于1.253时，准确率提升了1.2%，相较其他方法更具优势。

Depth estimation of monocular infrared images based on attention mechanism and graph convolutional neural network

The depth estimation of objects in the scene is a key issue in the field of the unmanned driving.The infrared images are helpful to solve the depth estimation problem under poor light conditions.Aiming at characteristics of unclear infrared images texture and insufficient edge information,a combination of attention mechanism and graph convolutional neural network was proposed to solve the problem of monocular infrared images depth estimation.First of all,in the depth estimation problem,the depth information of each pixel in the image was not only related to the depth information of its surrounding pixels,but also needed to consider the depth information of a larger range of other pixels.The attention mechanism could be effectively extract the pixel-level global depth information association of images.Secondly,the features obtained based on the depth information association could be considered as non-Euclidean data,and the graph convolutional neural network(GCN)was further used for reasoning.Finally,in the training phase,the continuous depth estimation regression problem was transformed into the classification problem,which made the training process more stable and reduced the learning difficulty of the network.The experimental results show that the proposed method has obtained good results on the infrared data set NUST-SR.When the threshold index is less than 1.25³,the accuracy rate is improved by 1.2%,which is more advantageous than other methods.