PenPy-DETR: A Penta-Pyramid Framework for Small Target Perception in Low-Light

Low-light object detection remains challenging due to low brightness, high noise, and small targets. Existing low-light enhancement methods often optimize for visual quality, causing feature misalignment with detection tasks. We propose Penta-Pyramid DETR (PenPy-DETR), an end-to-end framework that combines multi-scale feature enhancement with detection. It introduces two key components: a dual-pyramid enhancement network (DPE-Net) that serves as a front-end to enhance input features specifically for the detection task, and a CNN-based Enhanced Feature Fusion (CEFF) module that refines the detector's internal feature pyramid to better capture details of small targets. Experiments on the VisDrone dataset, degraded by a realistic Image Signal Processing (ISP) pipeline to simulate low-light noise and color artifacts, show that PenPy-DETR improves mAP₅₀ from 36.3% to 46.8% over the RT-DETRv2 baseline, while maintaining a moderate model size of 22.9M parameters.