SFA-Net: A Selective Features Absorption Network for Object Detection in Rainy Weather Conditions

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) excel in object detection for clear images.
• Visibility reduction in rainy conditions significantly hinders CNN-based object detection performance.

Purpose of the Study:

• To develop a robust object detection model, SFA-Net, that performs effectively in both normal and adverse rainy weather.
• To introduce a large-scale dataset, srRain, for training and evaluating object detection models under various rain conditions.

Main Methods:

• Proposed a novel Selective Features Absorption Network (SFA-Net) comprising three subnetworks: feature selection, feature absorption, and object detection.
• Developed and utilized the srRain dataset, containing 25,900 synthetic and real-world rainy images with 181,164 annotated instances across five categories.

Main Results:

• SFA-Net achieved state-of-the-art mean average precision (mAP): 77.53% (normal), 62.52% (synthetic rain), 37.34% (natural rain), and 32.86% (real rain).
• The model outperformed existing object detectors and combined deraining-detection approaches.
• SFA-Net maintained high detection speed.

Conclusions:

• SFA-Net effectively addresses the challenge of object detection in rainy weather by selectively absorbing features.
• The srRain dataset provides a valuable resource for advancing research in adverse weather object detection.
• The proposed method offers a promising solution for real-world applications requiring reliable object detection under various weather conditions.