| Conflagration is a serious threat to people’s lives,and an effective flame detection algorithm can minimize the harm of conflagration.Current models are based on flame color,contour and area,yet they overlook the gradual change of color of flame area.Thus,the thesis employs the gradient feature to represent the change of color of flame area;and extends depth learning model of rigid body to the field of flame detection.(1)In this thesis,a fire detection algorithm based on gradient feature is proposed.A dataset of non-flame pixels has been established based on the collected flame pixels by utilizing a modified k-Nearest Neighbor algorithm.Then,a classifier for obtaining candidate areas has been trained based on the aforementioned dataset.The feature of gradient has been defined on this basis,and two calculation methods have been designed.Accordingly,the candidate region has been classified according to the gradient feature,and the region can be accurately judged as the flame candidate region or the non-flame candidate region.This method is mainly based on the static characteristics of the flame image and combines with the mobile object detection technology,Visual Background extractor,to get a high precision flame detection based on a video.Because gradient feature contains the information of color changes in the flame area,the experiment based on the gradient feature works well.(2)The deep learning model in this thesis employs the design mode of rigid body detection model-Single Shot MultiBox Detector(SSD).Comparing with a rigid body,the contour of flame is variable,and the flame has fewer detailed features.Therefore,the structure of the network has been modified for flame detection.Because of the imbalance of the positive and negative samples,the weights of samples have been normalized based on SSD loss function.Compared with the original model,the modified one has fewer channels of feature map layer in extract candidate boxes,and thus becomes faster.The advantages and limitation of several algorithms,the flame detection algorithm proposed by Khatami et al,flame detection algorithm based on gradient feature,and modified SSD algorithm for flame detection,have been compared in this thesis.The modified models speed is faster in flame detection.Furthermore,a comparison with state-of-the-art preceding method and gradient feature mcxiel shows that the modified model has higher accuracy,while the gradient feature model has higher recall rate.The improved SSD model works better when the flame zones are independent,while the gradient feature model works better when the flame area is extremely irregular. |
