CN105631473A - Camera source identification method in finite labeled sample condition - Google Patents

Abstract

The present invention relates to a camera source identification method in a finite labeled sample condition, belonging to the technical field of signal and information processing. The method comprises a step of carrying out feature extraction of a labeled training sample set, a step of using a labeled sample feature training classifier to classify all samples and forming a prototype set according to the sorting of posterior probability, and a step of integrating the projected vectors of each sample in each prototype set to form an integrated feature, training the classifier by using the integrated feature formed by the labeled samples, and classifying a unlabeled sample.

Description

A kind of camera source discrimination method having existence exemplar condition

Technical field

The invention belongs to Signal and Information Processing technical field, relate to a kind of camera source discrimination method having under existence exemplar number limited conditions.

Background technology

For image capture devices such as digital camera, mobile phone and scanneies, there is different equipment manufacturer of all kinds and model various in style. After determining the acquisition device type of digital picture, differentiate and the model of its collecting device of collecting evidence, be the important content that differentiates of digital picture source. One of acquisition equipment that digital picture is the most commonly used is exactly digital camera, therefore the source camera model for digital photograph image differentiates, also it is the most concerned, is called the source authentication technique based on camera model (Model-Based) by people. The research emphasis of the present invention is camera source and differentiates.

Digital photograph image for camera shooting carries out source discriminating, and the most directly perceived and simplest method is EXIF (ExchangeableImageFileFormat) information of examination digital picture. EXIF is the standard formulated for the image file format of digital camera use, formulated in 1996 by Japanese Electronic Industries Development Association at first, and issued up-to-date 2.2 edition with in March, 2002, its objective is to increase the version information of perhaps image processing software in relevant photographing information for digital camera images. In EXIF definition about in the important information of photo property, the shooting camera model of in store photograph image, acquisition parameters and time etc. Unfortunately, these attributes in the picture with clear-text way preserve, and the such many picture browsings of such as ACDSee and process software, all support the amendment to these attributes, deletion, it might even be possible to replicate completely with the such software of JHead. Therefore, examination EXIF information differentiates the digital camera model of image, is only capable of as the reference that camera model source is differentiated, its result due to EXIF modifiability and and unreliable.

M.Kharrazi et al. thinks, the hardware device adopted due to different digital cameras is different with software algorithm, the digital photograph image of shooting can there are differences in color, picture quality etc., therefore it is from the angle of digital picture statistical nature, extract the features such as color correlation, neighborhood distribution centroid, color energy ratio, and use for reference wavelet character and Image quality measures, the camera source of photograph image is classified. M.Tsai et al. it is also proposed similar method. F.J.Meng et al., then on the basis of M.Kharrazi method, introduces the bispectrum feature of image, and all of feature is optimized selection. This kind of method in the small numbers of situation of digital camera, can reach the source of average about 90% and differentiate accuracy rate. But the digital camera sample being as obtaining image increases, or when existing with brand camera image, the discriminating accuracy rate of this kind of method will decline. Meanwhile, the statistical nature of these extractions, also it is easily subject to the impact of picture material and image capturing environment.

Other researcher is thought, extracts hardware device or the software algorithm feature of different model camera from digital photograph image, it is possible to carry out taxonomic history for the camera model of image is originated. K.S.Choi et al. finds, different camera lenses exist different light distortions, detect and quantify this distortion, it is possible to as the camera model source diagnostic characteristics of digital photograph image. They propose a kind of by detecting the method that the straight line distortion in image describes lens distortions, and apply it in the camera model source discriminating of image, camera image for three kinds of different models, reached 91.5% on average differentiate accuracy rate, but this method needs to extract sufficiently long straight line information in image. On the other hand, K.S.Choi often adopts the phenomenon of different JPEG quantization table always according to the digital camera of different model, and by calculating in jpeg image 8 �� 8 piecemeal, the percentage ratio of the non-zero quantised DCT coefficient of each position differentiates the camera model of image. The image that four kinds of digital cameras are shot by this method, has reached the discriminating accuracy rate of average 92%.

Table is quantified as outside image sources diagnostic characteristics other than with lens distortions and JPEG, another one important operation CFA (ColorFilterArray) interpolation in digital camera imaging process, also by many researcheres one of key character as digital camera, originate for the camera model of digital picture and differentiate. H.Farid et al. points out, cfa interpolation operation inevitably introduces the dependency of local pixel in the image of digital camera shooting, and on frequency domain, it is reflected as the peak point of energy, it is possible to detect this dependency by EM (ExpectationMaximization) algorithm. S.Bayram et al. is based on this thought, it is proposed to utilize EM algorithm detection neighborhood of pixels periodically, and using the peak point position of the interpolation weights coefficient obtained and Two-dimensional Probabilistic figure as feature, differentiates the camera model source of digital picture. For the image that the digital camera of three kinds of different brands shoots, it on average differentiates that rate of accuracy reached is to 96%. Y.J.Long and A.Swaminathan et al. then uses secondary correlation model and linear interpolation model respectively, by solving the interpolation coefficient of minimization problem estimation neighbourhood CFA, and it is respectively adopted BP neutral net and SVM as grader, the image of different model digital camera shooting is carried out source discriminating. In the less situation of number of samples of digital camera, the method for Y.J.Long and A.Swaminathan be attained by more than 95% on average differentiate accuracy rate. The image that then digital camera of 19 kinds of different models of 9 kinds of different brands shot of method carried out source and differentiated evidence obtaining experiment, reached current best classifying quality, it on average differentiates that accuracy rate is up to 85.9%

Existing correlational study, is all based on an important condition hypothesis, namely different image acquisition equipments, owing to the hardware/software in imaging system is different, capital introduces some statistical nature in image, by these statistical natures of classification/recognition, it is possible to its source is differentiated. It is however noted that, owing to these features are all based in statistical significance, the method that therefore has is required for substantial amounts of known label sample and carries out extraction and the training of feature, thus obtaining reliable and stable statistical nature model. In fact, due to the existing method of great majority, its characteristic dimension is all in tens magnitudes tieing up even hundreds of dimension, therefore to ensure the problem that the grader of training did not produce training, often every kind needs hundreds of known label sample to be trained, and is only possible to the relatively stable and desirable taxonomic history accuracy rate of acquisition. But in the actual environment, each classification is carried out to the acquisition of great amount of samples, be a very difficult job. Once enough known label samples cannot be obtained, its statistical nature is not it is possible to possess statistical significance, or training problem occurs in training grader process, causes that the taxonomic history accuracy rate of grader reduces rapidly.

Therefore, how in the actual environment of finite sample, it is achieved the accurate and effective of numeral image sources is differentiated classification, is a problem with significant application value equally, this is also the key problem solved in the present invention.

Summary of the invention

Present invention is generally directed to existing source discrimination method limited have an exemplar when the not high problem of Detection accuracy, invented and built based on prototype collection and the samples sources under existence tag part that has of integrated mapping differentiates. The method carries out feature extraction to there being label training sample set. Then utilize and have exemplar features training grader and to all sample classifications, constitute prototype collection according to the sequence of posterior probability. Then by each sample integrally formed integration characteristic of projection vector on each prototype collection, utilize the integration characteristic having exemplar to be formed to train grader, then classify to without exemplar.

Technical scheme is as follows:

1. camera source discrimination method general introduction when finite sample

There is camera source discrimination method when existence exemplar, it is characterised in that first to there being label training sample set to carry out 354 dimension LBP feature extractions; In order to obtain there being fully describing of existence exemplar, utilize and have exemplar training grader and to all sample classifications, constitute prototype collection according to the sequence of posterior probability; Then by each sample integrally formed integrated vector of projection vector on each prototype collection, this feature is designated as integration characteristic, utilizes the integration characteristic having exemplar to be formed to train grader, then classify to without exemplar.

2. prototype collection builds

Described a kind of camera source discrimination method when there being existence exemplar, it is characterized in that, build multiple prototype collection, thus obtaining there being fully describing of existence exemplar, building the method for prototype collection is from being had label training characteristics to concentrate to randomly select m dimensional feature and constitute new training characteristics collection by LBP structural feature, new training characteristics training is utilized to practice grader and to all sample classifications (include without label and have exemplar), so each sample will obtain the posterior probability belonging to all kinds of, if the probability that sample belongs to each class is all identical, at this moment it is considered that this sample is noise sample, the impact on classification results of this sample should be removed, when according to maximum entropy theory equiprobability, entropy is maximum, wherein the computing formula of entropy is shown below:

$entropy=-\underset{i=1}{\overset{N}{\Σ}}p\left(c_i\right){\log }_{2}p\left(c_i\right)---\left(1\right)$

Wherein p (c_i) represent that sample belongs to c_iThe probability of class, threshold value e is set according to this formula, noise sample will be taken as less than the sample of this entropy threshold to cast out, again sample is ranked up by remaining sample according to sample posterior probability, each class is chosen front n sample and is constituted a prototype collection, repeatedly repeats said process and can obtain the prototype collection of multiple description sample.

3. integrated mapping

Described one camera source when there being existence exemplar differentiates, it is characterized in that, integrated mapping theory is introduced source differentiate, utilize each self-training grader of prototype collection built, each have exemplar can obtain, after grader is classified, the posterior probability values belonging to each class on each prototype collection, it is designated as projection vector, undertaken integrated to obtain integrated vector (also known as integration characteristic) by the projection vector obtained on each prototype collection, utilize the feature set training grader in the enough cities of the integration characteristic having exemplar, and classify to without exemplar, namely final classification results is obtained.

The invention has the beneficial effects as follows:

The present invention is directed in the situation having label training sample limited, propose a kind of theoretical based on prototype collection and integrated approach camera source discrimination method, overcome the impact on source discriminatory analysis under the insufficient condition of exemplar, it is possible to discriminating of effectively originating.

Accompanying drawing explanation

Fig. 1 is that the prototype collection mentioned in the present invention builds framework.

Fig. 2 is the framework of the integrated mapping mentioned in the present invention.

Detailed description of the invention

The specific embodiment of the present invention is described in detail below in conjunction with accompanying drawing and technical scheme.

In order to make the object, technical solutions and advantages of the present invention clearly, describe the present invention below in conjunction with accompanying drawing.

Fig. 1 and Fig. 2 is the block diagram that the present invention carries out integration characteristic extraction when finite sample, and the method comprises the following steps:

Step one: to without exemplar and have exemplar to carry out feature extraction

In this step, it is possible to use existing source diagnostic characteristics, such as CFA feature, LBP feature etc., utilize feature extraction algorithm to without exemplar with there is exemplar to carry out feature extraction, obtaining without label test sample set D_uAnd have label training sample set D_l��

Step 2: build prototype collection, it is achieved to there being fully describing of existence exemplar

In order to make full use of the label information of existence exemplar, the theory of prototype collection is incorporated in the discriminating of source. Sample class number scale is N, constitutes k character subset from there being label training sample to concentrate by k random m dimension of choosing, is designated asNext this k character subset training SVM classifier it is utilized respectively, and to there being sample D more_uAnd D_lClassify, it is thus achieved that k classification results. For the result of each grader, according to the posterior probability of each apoplexy due to endogenous wind sample, sample is ranked up, n sample that selected and sorted is forward plus label, constitutes the prototype collection comprising original training sample collection partial information of a N �� n. Then k grader can obtain k prototype collection altogether. It is designated as { P₁,P₂,��,P_k}��

In above process, if to belong to the probability of each class all identical for sample before being ranked up, at this moment it is considered that this sample is noise sample, the impact on classification results of this sample should be removed, when according to maximum entropy theory equiprobability, entropy is maximum, and wherein the computing formula of entropy is shown below:

$entropy=-\underset{i=1}{\overset{N}{\Σ}}p\left(c_i\right){\log }_{2}p\left(c_i\right)---\left(1\right)$

Wherein p (c_i) represent that sample belongs to c_iThe probability of class, sets threshold value e according to this formula, will be taken as noise sample less than the sample of this entropy threshold and cast out, then according to sample posterior probability, sample is ranked up by remaining sample.

Step 3: integrated mapping, is undertaken integrated by the classification information obtained from each prototype collection.

Building in the process of prototype collection and add label to without the sample of label, therefore each prototype collection can be considered as one new has label training sample set. This k prototype collection is utilized to be respectively trained grader, and to D_lIn each sample classify, then each sample can obtain the posterior probability belonging to all kinds of on a grader, and this posterior probability is designated as vector v_i, by the vector { v of k grader acquisition₁,v₂,��,v_kCarry out integrated namely may make up integration characteristic V^N��k,1. To all D_lMiddle sample extraction integration characteristic, constitutes new training sample set, utilizes this training sample set training grader SVM, and to D_uIn classify without exemplar.

Namely achieved target integrated for the classification information of multiple prototype collection by said process, and achieve the effective classification without exemplar.

Claims (1)

1. the camera source discrimination method having existence exemplar condition, it is characterised in that following steps,

Step one: to without exemplar and have exemplar to carry out feature extraction

Use existing source diagnostic characteristics, utilize feature extraction algorithm to without exemplar and have exemplar to carry out feature extraction, obtaining without label test sample set D_uAnd have label training sample set D_l;

Step 2: build prototype collection, it is achieved to there being fully describing of existence exemplar

The theory of prototype collection is incorporated in the discriminating of source; Sample class number scale is N, constitutes k character subset from there being label training sample to concentrate by k random m dimension of choosing, is designated asNext this k character subset training SVM classifier it is utilized respectively, and to there being sample D more_uAnd D_lClassify, it is thus achieved that k classification results; For the result of each grader, according to the posterior probability of each apoplexy due to endogenous wind sample, sample is ranked up, n sample that selected and sorted is forward plus label, constitutes the prototype collection comprising original training sample collection partial information of a N �� n; Then k grader obtains k prototype collection altogether; It is designated as { P₁,P₂,��,P_k;

In above process, if to belong to the probability of each class all identical for sample before being ranked up, this sample is noise sample, it should remove the impact on classification results of this sample, when according to maximum entropy theory equiprobability, entropy is maximum, and wherein the computing formula of entropy is shown below:

$entropy=-\underset{i=1}{\overset{N}{\Σ}}p\left(c_i\right){\log }_{2}p\left(c_i\right)---\left(1\right)$

Wherein p (c_i) represent that sample belongs to c_iThe probability of class, sets threshold value e according to this formula, will be taken as noise sample less than the sample of this entropy threshold and cast out, then according to sample posterior probability, sample is ranked up by remaining sample;

Step 3: integrated mapping, is undertaken integrated by the classification information obtained from each prototype collection;

Building in the process of prototype collection and add label to without the sample of label, therefore each prototype collection is considered as one new has label training sample set; This k prototype collection is utilized to be respectively trained grader, and to D_lIn each sample classify, then each sample can obtain the posterior probability belonging to all kinds of on a grader, and this posterior probability is designated as vector v_i, by the vector { v of k grader acquisition₁,v₂,��,v_kCarry out integrated namely may make up integration characteristic V^N��k,1; To all D_lMiddle sample extraction integration characteristic, constitutes new training sample set, utilizes this training sample set training grader SVM, and to D_uIn classify without exemplar;

Namely achieved target integrated for the classification information of multiple prototype collection by said process, and achieve the effective classification without exemplar.