Complex dark environment-oriented object detection method based on YOLO-AS.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Object detection in complex dark environments suffers from low accuracy, false positives, and missed detections.
• Incomplete context and missing information hinder the effectiveness of existing methods.
• This necessitates advanced techniques for reliable object detection under adverse lighting.

Purpose of the Study:

• To propose an enhanced object detection method specifically for complex dark environments.
• To improve detection accuracy and robustness in low-light conditions.
• To address the limitations of current object detection models in challenging scenarios.

Main Methods:

• Developed a Zero-DCES image enhancement module for adaptive contrast enhancement in dark images.
• Constructed a YOLO-AS detection model integrating ECA-ASPP and SK attention mechanism.
• Utilized dilated convolution to expand receptive fields and channel attention for dynamic feature detection and multiscale expression.

Main Results:

• The proposed method achieved 78.39% map@50 on the ExDark dataset, a 5.78% improvement over the benchmark.
• Maintained comparable detection speed to existing mainstream models.
• Demonstrated significantly improved detection accuracy in complex dark environments.

Conclusions:

• The YOLO-AS based object detection method effectively enhances image quality and detection performance in dark conditions.
• The integration of Zero-DCES, ECA-ASPP, and SK attention mechanisms improves multiscale feature expression and detection accuracy.
• This approach offers a promising solution for reliable object detection in challenging low-light environments.