Smoke detection in dynamic scenes using box plot and fully convolutional network

Smoke has diverse forms in early fires, and owing to its strong transparency, time information is easily affected by dynamic backgrounds, resulting in low recognition accuracy and environmental interference in video-based single-stage smoke detection networks, thereby limiting their applicability. To address these problems, we propose a two-stage smoke detection algorithm that employs box plot background modeling (BPB) and a fully convolutional classification network (FCDN). The experimental results demonstrated that the proposed algorithm greatly reduces the miss rate and false alarm rate in dynamic scenes, enabling the basic model to accept non-square images directly; increases the detection efficiency, optimizes the model size; and enhances the applicability and reliability of the overall solution.