DSNet: Joint Semantic Learning for Object Detection in Inclement Weather Conditions

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Object detection using convolutional neural networks is prevalent in computer vision.
• Detecting objects in adverse weather, particularly fog, remains a significant challenge due to reduced visibility.

Purpose of the Study:

• To address the limitations of current object detection methods in foggy conditions.
• To develop a novel network for simultaneous visibility enhancement and object detection.

Main Methods:

• Introduction of a Dual-Subnet Network (DSNet) trained end-to-end.
• DSNet jointly learns visibility enhancement, object classification, and localization.
• Utilizes RetinaNet as the detection subnet and incorporates a feature recovery module in the restoration subnet.

Main Results:

• DSNet achieved 50.84% mAP on a synthetic foggy dataset.
• DSNet achieved 41.91% mAP on the Foggy Driving dataset.
• Outperformed state-of-the-art object detectors and dehazing-detection combination models.

Conclusions:

• DSNet offers a significant improvement for object detection in foggy environments.
• The proposed method effectively enhances visibility and maintains high detection speed.
• DSNet presents a promising approach for real-world applications requiring robust object detection in adverse weather.