Attention-Guided Low-Light Object Detection Based on Receptive Field With Adaptive Depth Convolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Object detection models struggle in low-light environments, leading to performance degradation.
• Existing methods often fail to adequately restore image quality and enhance features in dim conditions.

Purpose of the Study:

• To develop an attention-guided object detection model for improved performance in low-light scenarios.
• To enhance image quality and accentuate target features for more accurate detection.

Main Methods:

• Proposed a novel framework, ZRF-YOLOv9, integrating an image enhancement module (Zero-DiDCE++) with an improved YOLOv9 network.
• Introduced Feature Self-Attention (FSA) and Sparse Receptive Field Adaptive Depth Convolution (SRDConv) modules to refine feature representation.
• Employed adaptive depthwise convolution for better feature extraction in varying illumination.

Main Results:

• Achieved improved low-light detection accuracy on ExDark, Dark Face, and Tiny-Person datasets.
• Demonstrated accuracy improvements of 1.1%, 2.0%, and 0.5% in mAP$_{50}$ on the respective datasets.
• The model effectively enhances image visibility and target feature detection in low-light scenes.

Conclusions:

• The ZRF-YOLOv9 model offers a significant advancement for object detection in low-light conditions.
• The integration of image enhancement and attention-guided modules proves effective in boosting detection accuracy.
• This approach provides a robust solution for real-world applications requiring reliable object detection under poor lighting.