FSDJL-Net v1

Vehicle detection under adverse weather conditions often suffers from low contrast, blur, and sparse textures, accompanied by noise and scattering interference, which makes it difficult for detectors to stably capture critical structural cues. Existing methods usually emphasize either frequency-domain or spatial-domain cues, or directly fuse them at shallow stages, which may introduce noise and fail to fully exploit the complementarity between the two domains. To this end, we propose FSDJL-Net, a Frequency-Spatial Domain Joint Learning network for vehicle detection. Specifically, we design a Joint Domain Feature Downsampling (JDFD) module to suppress noise during downsampling while constructing complementary representations; propose a Frequency Domain Feature Extraction (FDFE) module to strengthen the modeling of high-frequency cues (e.g., details and edges) while maintaining inference efficiency; and further build a Joint Domain Learning (JDL) module, which performs large-receptive-field spatial modeling and enables more effective cross-domain interaction and fusion to obtain more robust joint representations.