Research On Flame Recognition Algorithm Based On Improved ViBe And Deep Learning

Flame is important in people’s production and life.But improper use of flame can cause fires and pose a serious threat to people’s life and property safety.If the fire can be detected early,appropriate measures can be taken in time to reduce or even eliminate the harm caused by the fire.Image-based flame recognition algorithms have a wider detection range than sensor-based flame recognition algorithms,which has important implications for fire prevention.This thesis combines the moving target detection technology in image processing and the image classification model in deep learning to propose an accurate and efficient flame recognition algorithm.The main research contents are as follows:The visual background extractor(ViBe)algorithm in classical moving target detection algorithms is improved to filter out some non-flame targets and extract suspected flame area.This thesis analyzes the color and shape structure features of flame and integrates them into the ViBe algorithm.Firstly,by setting rules for flame color discrimination,the falsely detected background targets are filtered out,and the effect of the filtering method is verified.Then,by setting rules for circularity discrimination,the flame-like targets are filtered out,and the effect of the filtering is also verified.Different image enhancement algorithms were compared for their effectiveness on enhancing fire images,and the best-performing image enhancement algorithm was selected to enhance the images in the dataset.In order to overcome the low accuracy and poor generalization ability of traditional fire image classifiers,the GoogLeNet in convolutional neural networks was used for fire image classification.The channel attention mechanism was added to the Inception module of GoogLeNet to achieve more accurate classification results.The image classification model was then combined with the moving target detection algorithm in Chapter 3 to detect moving targets in the image sequence,filter out non-fire interferences,and extract suspected flame areas.The suspected flame area images were then input to the trained flame recognition model for fire recognition.The experiments show that the flame recognition method proposed in this thesis can accurately identify fires in various scenarios,providing favorable criteria for fire alarm.