Encoder-decoder based convolutional neural network (EDCNN) for video classification of smoke and fire image.

An increase in fire accident rate has been reported to be growing recently due to a lack of advanced technology for detecting the presence of smoke and fire at the initial stage. To date, electronic sensor-based monitoring methods are the most common technique that has been widely practiced. However, the existing methods are not securely proven to detect the presence of smoke and fire at an initial stage with higher accuracy. This study presents a preliminary method for initial stage detection of smoke and fire using Deep Learning method. One of DL methods namely Encoder-decoder based convolutional neural network (EDCNN) is used to classify both smoke and fire density levels. In this study, the images data are extracted from a sample video sequence, and it is further divided into three different classes such as background or negative, smoke, and fire. An EDCNN with VGG-16 architecture is used to train and test the 3 classes of images data. The global accuracy of the proposed classification method reaches up to 95%. The classification model of EDCNN is then applied to the simulated fault electronic circuit video due to loose wires. According to the video application result, the proposed method is potential for further real time application where the existence of smoke and fire can be distinguished by the EDCNN classification model. [ABSTRACT FROM AUTHOR]