Detection-Friendly Dehazing: Object Detection in Real-World Hazy Scenes

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning object detection models suffer performance degradation in adverse weather.
• Image restoration techniques can enhance degraded images but integrating them with detection is challenging.
• Restoration labels are often unavailable in real-world scenarios.

Purpose of the Study:

• To propose an end-to-end framework, BAD-Net, that jointly performs image dehazing and object detection.
• To address the technical challenge of correlating image restoration and object detection tasks.
• To improve the robustness and accuracy of object detection in hazy conditions.

Main Methods:

• Developed BAD-Net, a union architecture connecting dehazing and detection modules.
• Implemented a two-branch structure with an attention fusion module for feature integration.
• Introduced a self-supervised haze robust loss and an interval iterative data refinement training strategy for weak supervision.

Main Results:

• BAD-Net demonstrated improved detection performance compared to state-of-the-art methods on RTTS and VOChaze datasets.
• The proposed architecture mitigates performance degradation caused by poor dehazing module performance.
• The framework achieves detection-friendly dehazing, enhancing overall accuracy.

Conclusions:

• BAD-Net offers a robust detection framework that bridges low-level image dehazing and high-level object detection.
• The end-to-end approach and novel training strategies enable effective dehazing and detection in challenging conditions.
• This work advances the integration of image restoration and object detection for real-world applications.