Multi-mode Feature Extraction Of Security Images And Image Classification Of MKL

With the advent of the era of big data,the types of data used to describe images are increasingly diversified.The image feature extraction and selection as the key link of image classification can determine the performance of the whole image classification system.Image classification system consists of three parts:target region selection,feature extraction and classifier modeling.In the field of public security inspection,most of the domestic security check adopts the method of manual image evaluation to identify the digital X-ray security image whether contains contraband at present,this method has low efficiency,strong subjectivity,and poor stability of identification results affected by environmental factors.In the face of massive X-ray security image recognition,the low-level features of a single-mode digital image cannot accurately describe the key content of a security image from the overall perspective.Therefore,it is of great significance to study how to choose useful features in X-ray security images,how to extract multi-mode of features from X-ray images with large noise,and the effective fusion of multi-mode features to improve the classification accuracy has great significance.In this dissertation,the extraction of multi-mode features and image classification of X-ray images are studied.Firstly,aimed to maintain the invariance of features when the local information scale changed,Hu improved moment invariant algorithm is proposed,and the K-means classifier is used to train and test the image data set.Based on the principle of copy-paste image tamper detection,this dissertation proposed an algorithm for security image specific target detection which can resist large scale geometric change.On the basis,the feature information of the target image region is obtained by expanding the peak value of Dirac function through the improved band limiting coefficient,and provides data support for the following image classification work.Then the relationship between the knowledge of security experts on the components of security images and the features of Elecroencephalogram(EEG)signals,EEG signals induced by image visual stimulation were all studied,the method of extracting EEG signal is proposed,this dissertation establish a classification model of EEG features.In order to make full use of the multi-source heterogeneous of features,a post data fusion scheme of multi-mode feature fusion is adopted,and a Simple Multiple Kernel Learning(SimpleMKL)model is used to automatically learn the fusion features of images,an automatic X-ray security image classifier is designed and implemented.The main contents and contributions of this dissertation are as follows:1.An improved Hu invariant moment algorithm is proposed to calculate the shape feature of the image.The traditional Hu moment invariants are invariable when describing the linear continuous image feature space,while the improved Hu moment invariants are invariable when 2D digital images are shifted,rotated and scaled,which is more stable against large-scale image geometric attacks.In this dissertation,the improved K-means classifier was used for cluster analysis based on the improved Hu moment feature,and the security check image classification accuracy was improved.The experimental results showed that compared with the classification results of the improved Hu moment feature and the traditional Hu moment feature,the former improved the average accuracy of the classification of the special target in the security inspection image by 17%.2.Based on the "copy-paste" image tamper detection algorithm,the research is firstly carried out based on the image coordinate theory in the log-polar coordinatedomain,so that any feature recognition algorithm based on the image translation invariant feature has the possibility to detect the "copy-paste" image region.Then an improved adaptive band limiting algorithm is proposed in the frequency domain by using the time-frequency signal of the image in the log-polar coordinate system.By automatically selecting the useful frequency band of the image,the peak value of the Dirac function is calculated to extract the image features of the contraband.Finally,the method of feature marking in the image area of contraband is used to realize the detection of security inspection images,which provides data basis for the design of image classifier with multi-mode feature fusion.The experimental results show that the improved frequency band limit algorithm can detect the image region of contraband adaptively,and the algorithm is robust.The average TPR(True Positive Rate)was 71%,while the average FPR(False Positive Rate)was 29%.3.In order to make the features of security inspection images diversified and break away from the shackles of traditional image detection methods,the feature extraction of X-ray security inspection images was studied from different perspectives.In view of the relationship between security experts’ cognition of the content in the images and EEG features,an EEG signal extraction scheme based on visual stimulus was designed,and an EEG image classification model based on EEG features was constructed.First,EEG signals induced by continuous visual stimuli in X-ray security images were collected,and the δ band signals were extracted by spatial decomposition and reconstruction of 7 layers wavelet packets.After that,Hilbert transform(HT)method was used to calculate the phase-related brain network features and analyze the correlation of each lead corresponding to the complex brain network features.Finally,the binary classification of X-ray security images is realized by using the features of brain network graph theory.The experimental results showed that the correlation between the leads of the brain network was more complicated in the experiment scene with blocked objects than in the experiment scene without blocked objects.Using the optimized Support Vector Machine(SVM)to train EEG signal features to achieve the accuracy of image two-category classification is 91.75%(no blocking scene),76.89%(with blocking scene),83.39%(mixed scene).4.The background of big data X-ray security images in the application,aiming at the feature problem of making full use of heterogeneous multi-mode,an automatic classifier of multi-mode features fusion security images based on Multiple Kernel Learning(MKL)was designed.Firstly,a post-fusion scheme of feature fusion is designed to meet the requirement of multi-mode feature data fusion in security inspection images.Then,the SimpleMKL model is used to automatically learn fusion features,and the optimal synthetic kernel is used to solve the problem of kernel function selection.Finally,an automatic security image classifier based on MKL based on Radial Basis Function(RBF)is designed.The experimental results show that the average recognition accuracy of image dichotomies obtained by using the data of 8 independent experiments is 93.19%(unblocked scene),77.26%(blocked scene)and 85.15%(mixed scene),respectively.The classification accuracy of security check images based on multi-mode feature fusion has been improved,and the technology of image classification based on single modal feature of traditional security check images has been improved in this dissertation.