Smoke detection in a digital image with the use of convolutional network

摘要

The article presents the concept of how to use convolutional networks as a method for processing digital images acquiredin visible region of light for the needs of smoke detection in large open area. The meaning and consequences of massiveblaze were underlined on the basis of statistical data concerning the forest fires. The proposal to overcome the difficultiesin using traditional methods for detection of fire threat by image processing techniques was discussed. The idea, innerstructure and properties of a convolutional neural network as a tool for automatic feature generation and imagerecognition were presented. The algorithms of data processing used in vision systems for fire detection were analyzedincluding the solutions implementing the networks. On the basis of the analysis the proposal to develop a neural networkfor smoke detection with the use of the strategy called transfer learning was presented. Using the image base of firesavailable on the web, the quantified assessment of the proposed approach was conducted. In the research the AlexNetframework was adopted to recognize smoke in images. The processing of the net was illustrated with examples ofactivations of selected layers when fed with images containing smoke. The 99% sensitivity reached by the proposedprocessing together with the 1% of false alarm rate seems to be very promising for the system of fire surveillance basedon watchtowers or air vessels monitoring large open areas