EP-Net

To address this issue, we propose EP-Net, a lightweight degradation-aware detection network, built upon a staged division of labor that emphasizes shallow representation enhancement and deep degradation suppression.

Specifically, at the shallow stage, we build DBMConv, which enhances the preservation of object structures and details via explicit low-high-frequency decomposition and dual path convolutions, and further reduces computational overhead at inference through re-parameterization. At the deep stage, we build DP-Former, which serially couples degradation-aware spatial modeling (DSM) and degradation-aware channel modeling (DACGM) to selectively suppress noise propagation and unreliable channel responses at semantically stronger levels.