This paper proposes a deep convolutional neural network with an encoder-to-decoder structure and constrains the network's in-depth learning from the monocular image at both two-dimensional (2D) and three-dimensional (3D) levels. At the 2D image level, an attention mechanism of channels is introduced to connect encoder features with decoder features with weights at the same scale, so as to balance the shallow detail features and deep semantic features extracted by the network. In addition, a scale-invariant loss and a multi-scale edge loss based on image pyramids are designed to obtain a depth map with rich edge detail information. At the 3D geometric level, a global geometric constraint loss and a local geometric constraint loss of depth are designed based on the local and global geometric relationships of coordinate points in space, in a bid to enhance the geometric consistency between point clouds. Furthermore, the results obtained through the proposed method are quantitatively and qualitatively compared with that obtained through other methods from the NYU Depth-v2 dataset, and it is shown that the proposed method can estimate indoor scene depth with higher accuracy and detail representation, obtaining accurate and smooth 3D reconstruction results on a single image.