Real-time video-based smoke detection with high accuracy and efficiency. (October 2020)

• Record Type: Journal Article
• Authors: Li, Chenghua; Yang, Bin; Ding, Hao; Shi, Hongling; Jiang, Xiaoping; Sun, Jing
• Is Part Of: Fire safety journal. Volume 117(2020)
• Publication Date: 2020-10
• Subjects: smoke Detection -- Convolutional neural networks -- Regularization loss function; Fire prevention -- Periodicals; Incendies -- Prévention -- Recherche -- Périodiques; Fire prevention -- Research; Periodicals; 628.92205
• DOI: 10.1016/j.firesaf.2020.103184
• Languages: English
• ISSNs: 0379-7112

Abstract

Abstract: Video-based smoke detection is a challenging task due to the large variance of smoke color, brightness and shape. Handcrafting discriminative features for smoke detection is a complicated and expensive task, which cannot represent smoke characteristics accurately. There are some smoke detection methods based on Convolutional Neural Networks (CNNs), but most of them are computationally expensive and difficult to achieve real-time detection. In this paper, we propose a novel data processing pipeline based on deep learning algorithm to improve the accuracy and efficiency in the smoke detection task. In the pipeline, Smoke Region Proposal is proposed to extract the suspected smoke regions and in order to achieve the purpose of real-time detection, the convolutional neural network model is pruned and reconstructed. The reconstructed CNN model is named SCCNN. To enhance the effect of the SCCNN model, we propose a regularization loss function named Score Clustering to reduce the overfitting problem caused by fitting the one-hot label and improve the accuracy of the model. Experimental results demonstrated that the SCCNN model has higher Accuracy Rate, Recall Rate and F1 Score on smoke datasets with fewer parameters and the running speed reached 61.2 frames per second.