To overcome the inherent limitations of existing unsupervised monocular depth estimation frameworks and enhance the network's generalizability across various scenarios, a lightweight unsupervised monocular depth estimation method that combines convolutional neural networks, attention mechanisms, and speeded-up robust features (SURF) is proposed. First, a residual block with a linear self-attention mechanism (CCT-Block) and a residual block with a coordinate attention mechanism (CA-Block) were designed. These residual blocks were alternately used within the residual network framework to construct a multiscale encoder capable of capturing rich contextual information and mapping the relationship between the depth and image features while reducing the requirements for parameter computation and storage. In addition, the reprojection error of SURF was introduced to mitigate ambiguities that may arise in depth and pose estimation networks. Finally, evaluations were conducted on multiple datasets, including KITTI, Make3D, NYUDepth-v2, and Cityscapes. The experimental results show that the proposed method achieves an absolute relative error of 0.107 and a root mean square error of 4.674 on the KITTI dataset using only 4.9×10⁶ model parameters. Furthermore, the proposed method exhibits strong generalizability across different datasets.