A method for low-resolution vehicle detection based on perceptual enhancement and multi-scale fusion

To address the issues of low detection accuracy and high false negative rates caused by insufficient feature information in low-resolution objects, this paper proposes a low-resolution vehicle detection method based on perceptual enhancement and multi-scale fusion. First, a Spatial Local Feature-based Perceptual Enhancement Backbone Network (SLFPB-ST) is designed. By integrating dilated convolutions and residual connections, it enhances the perception of local image context, reducing information loss during feature extraction for low-resolution objects. Second, a Multi-Scale Feature Integration Network (MSIFN) is constructed, incorporating a weight distribution mechanism to aggregate detail information across different scales. A High-resolution objects Suppression Block (HOSB) is integrated within MSIFN to suppress feature responses from high-resolution objects, thereby focusing on the feature representation of low-resolution objects. Finally, an anchor-free detection mechanism is adopted to avoid matching errors between predefined anchors and low-resolution objects, further reducing false negatives. Experiments on the UA-DETRAC and Vehicle datasets demonstrate that compared to the Swin Transformer baseline, the proposed method achieves improvements of 5.15%, 9.35%, and 4.35% in mAP, AP₅₀, and AP₇₅ metrics, respectively. This is achieved with only a 14MB increase in model parameters and a 0.4 fps decrease in detection speed (FPS). Experimental results validate the proposed method's superior robustness and practicality for low-resolution vehicle detection tasks.