Blur Detection for Surveillance Camera Systems: Dataset and Model

Surveillance videos taken in unconstrained environments can be tampered with due to different environmental factors and malicious human activities. They often blur the video content and introduce difficulty in identifying the events in the scene. The problem is particularly acute for smart surveillance systems that need to make real-time decisions based on the video. Automatic detection of the blur anomalies in the video is crucial to these systems. In this research, a learning-based approach for camera blur detection is proposed. The proposed model was trained with a newly constructed image dataset, which contains 17,000 surveillance images with common blur anomalies. All testing images of the dataset were taken from real scenes; they serve as a reference for evaluating the performance of the comparing methods. To fully utilize the abundant positive samples, the proposed model adopted a self-supervised learning method. A color attention module is included based on our observation that some blurry images have a special distribution in their 2-dimensional color histograms. Our experiment results show that the proposed model significantly outperforms state-of-the-art approaches for blur detection while keeping the model size small for edge applications.