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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Drone-based object detection accuracy degrades significantly under low-light and noisy conditions.
• Existing algorithms struggle with insufficient illumination, compromising surveillance capabilities.

Purpose of the Study:

• To develop an efficient and robust parallel neural network for drone-view object detection in nighttime environments.
• To concurrently enhance image quality and improve object detection accuracy under adverse lighting.

Main Methods:

• A coevolutionary framework with bidirectional gradient propagation between image enhancement and object detection modules.
• Integration of Zero-DCE++ for adaptive illumination adjustment and a lightweight YOLOv5 for real-time detection.
• Introduction of spatially adaptive feature modulation and high/low-frequency adaptive feature enhancement blocks for optimized feature extraction.

Main Results:

• The proposed method achieved significant improvements in mean Average Precision (mAP) on VisDrone2019 (Night) and Drone Vehicle (Night) datasets compared to traditional YOLOv5.
• Demonstrated enhanced performance in extreme low-light and high-noise scenarios, with mAP@0.5:0.95 improvements of 3.13% and 3.1%, and mAP@0.5 improvements of 6.3% and 2% respectively.
• The parallel model proved effective in improving feature representation robustness and detection accuracy.

Conclusions:

• The developed parallel neural network offers an efficient and reliable solution for nighttime drone-based visual monitoring.
• The joint optimization of image enhancement and object detection significantly boosts performance in challenging low-light conditions.
• The model's architecture enhances feature perception and semantic representation for improved drone surveillance.