Research On Attention-based Video Image Recognition Methods For Disasters

In recent years, with the development of science and technology, all of the large, medium and small enterprises in our country have accessed to more fast development. Nowadays many enterprises have also realized factory automation, and the productivity of enterprises more and more rely on the many kinds of the advanced equipment. Equipment safety has come under more and more attention with the development of the equipment automation. To prevent and avoid the natural or artificial disasters is an important subject of the enterprise production safety, and the fire and spill disasters are the most frequently happened and extremely destructive disasters. Therefore, the research of fire disasters detection methods has very important practical significance.The earliest fire disaster detection methods used various sensors and signal processing technology to detect the fire disasters, for example, smoke detector, heat detector, flame detector and compound detector, etc. But all kind of the detectors are very vulnerable to air flow velocity and dust, space position, the influence of environment temperature and humidity environments, thus it usually caused a certain degree of omissions or misstatements. With the development of computer technology and digital image processing technology, more and more researchers began to study the fire detection methods and techniques based on video image, and important progresses have been made.This thesis mainly studies disaster video image recognition methods based on attention mechanism. Different from the traditional video image recognition methods, the proposed methods uses underlying human visual characteristic to make spotlight focus rapidly on significant areas, and then extract the features of the significant areas under fire and smoke features. It is the first time of the thesis that apply the visual attention mechanism to the disaster video image recognition, therefore, it is a creative origination, which remains to be more scientific research in this direction.