Image Enhancement Guided Object Detection in Visually Degraded Scenes.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Object detection accuracy significantly decreases in visually degraded environments like underwater, hazy, or low-light conditions.
• Separating image enhancement and object detection tasks leads to suboptimal performance.
• Existing methods often fail to effectively bridge the gap between image quality and detection accuracy.

Purpose of the Study:

• To develop an integrated, end-to-end method for enhancing object detection in visually degraded scenes.
• To improve the robustness and accuracy of object detection models by incorporating image quality awareness.
• To provide a computationally efficient solution for real-world applications.

Main Methods:

• Propose an image enhancement guided object detection method with parallel enhancement and detection branches.
• Introduce a feature guided module to align features from enhanced and original images.
• Employ a frozen enhancement branch during training to guide the object detection learning process.

Main Results:

• The proposed method significantly improves detection performance on underwater, hazy, and low-light datasets.
• Integration enhances popular object detection networks like YOLO v3, Faster R-CNN, and DetectoRS.
• No additional computational cost is incurred during the testing phase.

Conclusions:

• The end-to-end approach effectively guides object detection learning by leveraging image enhancement features.
• The method makes detection networks aware of both image quality and object detection tasks.
• This strategy offers a significant performance boost for object detection in challenging visual conditions.