Convolutional neural network-based flame detection method and device

Abstract

The invention provides a convolutional neural network-based flame detection method. The method includes the following steps that: the sample image of a label is selected, the sample image is trained by using a convolutional neural network, so that a trained model can be obtained; a suspected flame point in a color scene image is extracted, and a binary image is obtained with the suspected flame point adopted as a foreground point; connected region processing is performed on the binary image through using a connected region method, so that a series of connected regions can be obtained; the connected regions are screened, so that candidate regions can be obtained; and the trained model is utilized to recognize the candidate regions, and a recognition result is outputted. Compared with an existing flame detection technology, the convolutional neural network-based flame detection method of the invention has the advantages of high flame detection accuracy and high robustness.

Description

technical field [0001] The invention relates to image processing, video monitoring and fire protection, in particular to a flame detection method and device. Background technique [0002] Fire is a common and frequent disaster, which seriously threatens the safety of national property and people. In order to effectively and timely warn the occurrence of fire and avoid the spread of fire, it is of great practical and economic significance to introduce an efficient flame detection method. Since the flame itself contains complex and changeable static and dynamic features, how to accurately describe the flame characteristics so that it can effectively distinguish between flames and non-flames in complex scenes has always been a difficult and core issue in the field of flame detection. [0003] Existing flame detection mainly revolves around flame features, including flame static and dynamic features, and introduces relevant probability and classification models on the basis of ...

AI Technical Summary

Problems solved by technology

The color feature is mainly based on RGB, YCbCr and other color spaces, and the detection speed is fast, but it cannot correctly distinguish the target that is close to the color of the flame

The outline feature mainly describes the outline information such as the area, perimeter, and shape of the flame. However, due to the uncertainty and diversity of the flame outline, it is impossible to accurately describe the distinction between the flame domain and the non-flame from the outline.

At present, the dynamic characteristics of flames are only a kind of frequency change that simply describes the local or global flames, but they cannot be distinguished for flashing lights, shaking red flags, etc.

Images