基于深度CRF网络的单目红外场景深度估计

Depth Estimation of Monocular Infrared Scene Based on Deep CRF Network

对单目红外图像进行深度估计,不仅有利于3D场景理解,而且有助于进一步推广和开发夜间视觉应用。针对红外图像无颜色、纹理不丰富、轮廓不清晰等缺点,本文提出一种新颖的深度条件随机场网络学习模型(deep conditional random field network,DCRFN)来估计红外图像的深度。首先,与传统条件随机场(conditional random field,CRF)模型不同,DCRFN不需预设成对特征,可通过一个浅层网络架构提取和优化模型的成对特征。其次,将传统单目图像深度回归问题转换为分类问题,在损失函数中考虑不同标签的有序信息,不仅加快了网络的收敛速度,而且有助于获得更优的解。最后,本文在DCRFN损失函数层计算不同空间尺度的成对项,使得预测深度图的景物轮廓信息相比于无尺度约束模型更加丰富。实验结果表明,本文提出的方法在红外数据集上优于现有的深度估计方法,在局部场景变化的预测中更加平滑。

Depth estimation from monocular infrared images is required for understanding 3D scenes;moreover,it could be used to develop and promote night-vision applications.Owing to the shortcomings of infrared images,such as a lack of colors,poor textures,and unclear outlines,a novel deep conditional random field network(DCRFN)is proposed for estimating depth from infrared images.First,in contrast with the traditional CRF(conditional random field)model,DCRFN does not need to preset pairwise features.It can extract and optimize pairwise features through a shallow network architecture.Second,conventional monocular-image-based depth regression is replaced with multi-class classification,wherein the loss function considers information regarding the order of various labels.This conversion not only accelerates the convergence speed of the network but also yields a better solution.Finally,in the loss function layer of the DCRFN,pairwise terms of different spatial scales are computed;this makes the scene contour information in the depth map more abundant than that in the case of the scale-free model.The experimental results show that the proposed method outperforms other depth estimation methods with regard to the prediction of local scene changes.