Finding accurate correspondences between two images is fundamental to many vision tasks. Recent research has addressed the limitation of Multi-Layer Perceptrons (MLPs) in extracting deep features, which lack local context, by employing CNN-based methods. However, simple CNNs tend to filter out some correct motion information as noise in complex scenes variations. Therefore, we propose a Spatial and Channel Feature Enhancement Network (SCFNet) to enhance the representation of motion vector features. This allows the network to better learn the motion patterns of inliers. Experiments show that compared to several existing methods, our proposed SCFNet achieves superior performance in the task of pose estimation.