Local Feature Modeling Based On Transform Domains For Visual Smoke Detection

Visual smoke detection has become a research hotspot and difficulty in the early fire detection field,and its detection rate directly affects the application of visual fire detection technology.In recent years,with the rapid development of technologies such as digital image processing,video analysis,pattern recognition,and machine learning,visual fire detection has higher requirement for the detection rate of smoke detection.The visual features of smoke generally show up as varied forms,various colors,extremely differences among transparencies,irregular movements,etc.,are easily affected by the external environments,and have less stability,which makes it difficult to extract the smoke features with the good expression ability and robustness from videos and images to increase the smoke detection rate.Therefore,studying the smoke feature modeling methods in visual smoke detection is of great theoretical value and practical significance for early fire alarm.The feature of transform domain and the static texture feature are two importance smoke features.The transform domain contains the higher-level abstract information of the original image,and extracting the feature from the transform domain is helpful to enhance the ability to express smoke feature.The local feature of smoke can well represent the static texture information of smoke,and is a stable smoke feature.Therefore,this dissertation focuses on the smoke feature modeling based on transform domains and local feature patterns,so as to improve the detection rate and reduce the false alarm and error rates for smoke detection,and further promote the technological development of visual smoke detection.The main researches of the dissertation are as follows.(1)A local feature modeling method based on edge transform domain is proposed.In the image,smoke usually has the blurry curved edges and few straight edges,while other artificial objects have the clear straight edges.According to this phenomenon,this method extracts the local feature from the edge feature map to enhance the expression ability of the smoke feature.In order to ensure the adaptability of the edge detection algorithm,this algorithm detects the edges with an adaptive Canny operator to generate the edge feature map.For the binary edge feature map,this method presents two local feature patterns,namely local boundary summation pattern(LBSP)and local region summation pattern(LRSP).After extracting LBSP and LRSP features from the edge feature map,and the local binary pattern(LBP)features from the original image and the edge feature map,this method obtains the final feature by selecting the extracted features.The method can use edge transform domain to solve the problem that the edge information of the target objects in the image is not used.(2)A local feature modeling method based on Gabor transform domain is proposed.Image transformation is usually used in the traditional visual smoke detection.The process of smoke feature extraction generally uses only a single-scale transformation domain,and rarely applies Gabor transform domain with multi-scale.Therefore,this method introduces Gabor wavelets with multi-scale and multi-direction to enhance the expression ability of smoke feature by extracting the local features of Gabor feature maps.First,this method constructs the aggregated Gabor kernels to reduce the number and the redundant information of the traditional Gabor kernels.Second,this method improves the encoding method of the traditional local binary pattern,and uses a custom comparison function instead of the original binarization function,so that the abundant pixel value information in the Gabor feature map is well used.Finally,the improved local binary pattern is used to extract the features from the original image and the Gabor feature maps,and the extracted features are concatenated as the final feature.The method can apply Gabor transform domain to solve the problem that the traditional local features do not have the multi-scale and multi-directional characteristics.(3)A local feature modeling method based on Gabor transform domain and edge transform domain is proposed.The Gabor feature maps can extract the abstract features contained by the original image,and have the rich pixel value information.In order to make better use of the pixel value information in the Gabor feature maps,this method improves local ternary pattern.Aiming at the problem that the fixed threshold of the traditional local ternary pattern is not suitable for all images,this method presents the confidence level-based local ternary pattern(CLLTP).CLLTP calculates the high and low thresholds of the ternaryzation function by utilizing the confidence level threshold,and designs the new ternaryzation function to generate the encoding feature maps,so as to improve the adaptability of CLLTP.To solve the problem that the feature fusion method of the features extracted from the Gabor feature maps is too simple to reflect the importance of each extracted feature,this method applies the information entropy of each Gabor feature map to calculate the weight of the corresponding feature to generate the weighted feature.Moreover,the edge feature map can capture the specific high-frequency information in the original image and reflect the edge features of the objects,so this method also produces the edge feature map of the original image.After utilizing CLLTP to calculate the features of the original image and the edge feature map,and the weighted features of the Gabor feature maps,this method achieves the final feature by feature fusion and feature dimension reduction on these features.The method can solve the problem that the fusion method of the local features extracted from the multiple Gabor transform domains is too simple.At the same time,two kinds of abstract information in the original image can be captured by using Gabor transform domain and edge transform domain.(4)A local feature modeling method based on multilayer Gabor transform domain is proposed.In the traditional feature extraction methods with Gabor kernels,each Gabor kernel usually only performs one convolution operation,so that the abstract features at different levels in the original image cannot be extracted at the same time.Hence,this method constructs a multilayer Gabor convolutional network to extract the abstract features at multiple levels in the original image,and generates the multilayer Gabor feature maps.The network contains an image input layer and multiple Gabor feature calculation layers.The Gabor feature calculation layer includes a Gabor convolution sublayer and a Gabor feature fusion sublayer.The convolution sublayer can apply the Gabor kernels to generate the Gabor feature maps.The feature fusion sublayer can produce new Gabor feature maps by fusing the Gabor feature maps from the convolution sublayer,in order to avoid an exponential increase in the number of the output maps for the Gabor feature calculation layer.After the Gabor convolutional network output the maps at multiple levels,this method calculates LBP features of these images outputted by the Gabor convolutional network.Finally,this method selects LBP features of each sublayer by using a weight vector of feature selection,and achieves the final feature by feature fusion and feature dimension reduction on the selected features.To solve the problem that Gabor kernels only perform one convolution operation,the method constructs the Gabor convolution network to capture the abstract information at multiple levels in the original image.(5)In order to verify the performances of four above-mentioned local feature modeling methods,the four proposed methods are applied to extract the smoke features in the smoke recognition and video smoke detection experiments.The video smoke detection experiments use the classification models trained by the smoke recognition experiments.The experimental results show that the four methods can achieve higher detection rates in smoke recognition experiments,and the features extracted by the four methods have good discriminative ability for smoke.In the video smoke detection experiments,the classification models can well detect smoke from the smoke videos,and achieve good performances for early smoke alarm.