The current mainstream feature matchers primarily focus on enhancing the feature interaction between image pairs for 3D reconstruction or SLAM tasks. However, these methods mainly focus on the matching stage itself while neglecting the impact of shallow features on subsequent matching. To bridge this gap, we propose an improved feature matching framework based on JamMa, termed SFE-JamMa, which mainly consists of the Shallow Feature Enhancement (SFE) and ResMixer modules. Specifically, the SFE module is composed of the proposed Pinwheel-shaped Mask Convolution and Central Difference Convolution. This module enhances shallow feature representation during extraction, effectively preserving fine-grained details to improve subsequent matching accuracy. The ResMixer module is designed to strengthen the interaction between features in the fine matching stage. Moreover, the computational pressure caused by module SFE has been alleviated. Experimental results on the MegaDepth dataset show that our proposed SFE-JamMa surpasses most current methods regarding Relative Pose Estimation accuracy and visualization results and achieves the state-of-the-art performance among sparse and semi-dense matchers.