Low-light images often undergo insufficient brightness, low contrast, and structural fidelity loss, severely affecting subsequent semantic vision tasks. To confront this difficulty, we formulate SFRNet, a deep enhancement architecture that integrates attention representation with feature fusion. Specifically, we design a Salient Feature Representation Module (SFRM) to adaptively highlight critical details while discarding redundant components, thereby enabling the pipeline to better record structural and contextual cues from degraded images. In addition, we introduce a Feature Integration Module (FIM) that concatenates the input features with the output of the SFRM, effectively enriching feature diversity and boosting representational strength. Empirical findings on varied benchmark datasets illustrate that SFRNet yields better visual outcomes and higher quantitative scores than cutting-edge counterparts.