A low illumination target detection method based on a dynamic gradient gain allocation strategy

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current target detection methods struggle with feature extraction in low illumination, leading to missed or false detections.
• Effective target recognition in adverse lighting conditions remains a significant challenge in computer vision.

Purpose of the Study:

• To develop an efficient and accurate target detection method specifically designed for low illumination environments.
• To overcome the limitations of existing methods in handling feature extraction and localization under poor lighting.

Main Methods:

• Introduced DimNet, incorporating a novel neck structure for efficient multi-scale feature fusion.
• Designed a feature aggregation module to fuse channel, spatial, local, and global information for improved network representation.
• Developed a new detection head using reparameterization and parameter sharing for enhanced accuracy and computational efficiency.
• Implemented a new loss function focusing on target center and employing dynamic gradient gain for improved localization accuracy.

Main Results:

• DimNet achieved a mean Average Precision (mAP50) of 75.60% on the ExDark dataset.
• Demonstrated a 3.77% improvement over the baseline model and a 2.25% improvement over state-of-the-art (SOTA) models.
• Outperformed previous and current SOTA methods in detection accuracy and overall performance.

Conclusions:

• DimNet offers a significant advancement in target detection for low illumination scenarios.
• The proposed enhancements in feature fusion, aggregation, detection head, and loss function contribute to superior performance.
• DimNet provides a robust and efficient solution for real-world applications requiring reliable target detection under challenging lighting.