Parallel joint encoding for drone-view object detection under low-light conditions

Under low-light conditions, the accuracy of drone-view object detection algorithms is frequently compromised by noise and insufficient illumination. Herein, we propose a parallel neural network that concurrently performs image enhancement and object detection for drone-view object detection in nighttime environments. Our innovative coevolutionary framework establishes bidirectional gradient propagation pathways between network modules, improving the robustness of feature representations through the joint optimization of the photometric correction and detection objectives. The illumination enhancement network employs Zero-DCE++, which adaptively adjusts the brightness distribution without requiring paired training data. In our model, object detection is performed using a lightweight YOLOv5 architecture that exhibits good detection accuracy while maintaining real-time performance. To further optimize feature extraction, we introduce a spatially adaptive feature modulation module and a high- and low-frequency adaptive feature enhancement block. The former dynamically modulates the input features through multiscale feature fusion, enhancing the ability of the model to perceive local and global information. The latter module enhances semantic representation and edge details through the parallel processing of spatial contextual information and feature refinement. Experiments on the two data sets of VisDrone2019 (Night) and Drone Vehicle (Night) show that the proposed method improves 3.13 and 3.1% compared with the traditional YOLOv5 method mAP@0.5:0.95, and improves 6.3 and 2% in mAP@0.5, especially in the extreme low light and high noise environment. Thus, the proposed parallel model is an efficient and reliable solution for drone-based nighttime visual monitoring.