Degradation Type-Aware Image Restoration for Effective Object Detection in Adverse Weather.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Convolutional Neural Network (CNN)-based object detection is crucial but struggles in adverse weather.
• Current methods often fail to adapt to diverse weather conditions due to scenario-specific designs.

Purpose of the Study:

• To develop an object detection network resilient to various adverse weather conditions.
• To enhance object detection accuracy by integrating image restoration with degradation type awareness.

Main Methods:

• Proposed DTRDNet, featuring a shared feature encoder, object detection decoder, degradation discrimination image restoration decoder (DDIR), and degradation category predictor (DCP).
• Jointly trained the network on mixed datasets of clear and degraded images.
• Incorporated degradation type information into DDIR and enabled degradation category awareness in the shared feature encoder (SFE) via DCP.

Main Results:

• DTRDNet achieved an average mAP of 79.38% across clear, hazy, rainy, and snowy test sets.
• Demonstrated superior performance compared to advanced object detection algorithms in diverse weather scenarios.
• The DCP and DDIR modules can be removed during inference for real-time performance.

Conclusions:

• DTRDNet effectively addresses the challenge of object detection in adverse weather by leveraging degradation type-aware restoration.
• The network's adaptability to diverse weather conditions significantly enhances detection accuracy.
• The proposed architecture offers a flexible solution for real-time object detection in challenging environments.