A Context-Encoder-Based Thermal Anomaly Detection Network for Dual-Sensor Surveillance

Thermal and visible dual-sensor imaging are commonly used in video surveillance, with anomaly detection being the focus of monitoring. Existing deep-learning-based methods for thermal image anomaly detection suffer from the issue of generated images not being clear enough. Meanwhile, scarcity of anomalous data is a common headache in anomaly detection. To alleviate the problem of blurry generated images, this paper proposes a Context-Encoder-based thermal anomaly detection network. Leveraging the capabilities of the Context-Encoder deep learning network in context feature learning and image generation, the proposed network establishes a mapping from visible images to thermal images, generates thermal images, compares them with actual thermal infrared imaging, and subsequently detects anomalies. Further, it also incorporates adversarial loss to improve the realism of the generated samples, thus reducing blurriness. To overcome the challenge of sparse anomalous data, the paper also proposes a thermal pseudo anomaly generation method to test the proposed anomaly detection approach. Experimental results on public LLVIP dataset demonstrate the effectiveness of the proposed method.