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.