Video Fire Shape Deformation Identification Algorithm based on Self-Supervised Contrastive Learning

Real-time video detection of fire is widely used in fire prevention and security fields, allowing for timely fire warnings and reducing personal and property damage. Existing algorithms often adopt object detection models cascaded with other criteria to determine whether fire exists in video frames. These criteria are mostly limited to single-frame image features, which fail to effectively suppress false alarms caused by high-brightness interference objects (such as car headlights or streetlights at night). Therefore, this paper proposes a video fire shape deformation recognition algorithm, FireShapeCLR(Contrastive Learning Representation), based on self-supervised contrastive learning. Firstly, a sample collection strategy is proposed to construct anchor samples, positive samples, and negative samples triplets. Secondly, a feature extraction network is utilized to compute the feature of samples and calculate the cosine distance between the triplets. Finally, Info-NCE loss function is employed to drive model training. This algorithm effectively leverages the visual characteristic of high-frequency shape variations of fire in the time domain, combined with self-supervised contrastive learning, to address the limitation of traditional algorithms in determining whether an object has deformed, significantly reducing false alarm rates. Experimental results show that the cascaded FireShapeCLR algorithm can reduce the false alarm rate by 23% while maintaining the detection rate of the original model.