Maksymiv O. Information technology of flame identification by convolutional neural networks in video surveillance systems

In the dissertation work, on the basis of the research carried out, an important scientific and applied problem was solved. This problem concerns the development of information technology for the identification of a flame in images and their sequences at the initial stages. Was analyzed the features of the following color models: RGB, YUV, HSI, YCbCr and L * a * b. The obtained results show that despite the widespread use of the color segmentation method in fire detection problems, such method has a large number of false-positive challenges. In contrast to existing approaches to flame detection in images, it was decided to develop a hypothesis generator that would allow to separate areas of images that by their characteristics can not resemble a flame. In comparison with other works, where the authors focused their efforts on achieving the max level of accuracy, the hypothesis generator aims to ensure the max level of precision. Experimental results on benchmark fire datasets reveal the effectiveness of the proposed information technology and validate its suitable for life safety systems. The accuracy of the model on images (without motion information) is 86% and for image sequences (video) 92%. Keywords: information technologies, machine learning, video cameras, fires, convolutional neural networks, artificial intelligence, data processing, video surveillance systems, generation, verification, transfer training.