Research On Image Splicing Blind Identification Algorithm Based On LGPQ-advanced Features

As the most common and convenient information carrier,images play an increasingly important role in the modern society,which involves all aspects of people's life,from judiciary and media to personal entertainment and social interaction.Meanwhile,various edit tools developed rapidly,and the flourish of image tampering methods brings a great challenge to the task of tampering image identification.In this paper,we focus on an important area of the tampering image identification field,called splicing image identification.One of the most effective solution is to extract the statistical features of the texture of the images,which can effectively capture some features left by the tampering process.LPQ as a most promising feature can achieve a good result,but simply utilize the pure LPQ feature cannot have a satisfied result.Therefore,LGPQ feature is proposed by researcher and applied in tampering image identification field to capture the LPQ features from multiscale and multi-orientation.Although the performance of LGPQ feature is better than it of LPQ feature,there are also some inherent defects of LGPQ feature.Based on the above,we proposed two LGPQ-advanced features and two splicing image identification models correspondingly,which called LGPQ-Dispersion feature and Global-LGPQ feature respectively.Next,we will explain the two features from the following three aspects: motivation(that is why we proposed the feature),what inherent defects of LGPQ feature can solved by the proposed feature(that is the contribution of the proposed feature),validity analysis(that is why the feature can effectively distinguish authentic images and splicing images):1)LGPQ-Dispersion(called LGPQ-D in the paper).key-pixel plays different roles under different window size(key-pixels are the pixels in the authentic area side of tampering edge).In the other words,key-pixel exists in authentic area in the small-scale window,meanwhile,it also exists in the splicing area in the large-scale window.Since the calculation of pure LGPQ feature is in fixed-window size,that is,the pixels can play only one role,which cannot reflect the multiple-role of key pixels.Based on the defects of LGPQ feature described above,we proposed LGPQ-D feature.The intention of LGPQ-D feature is to distinguish real image and mosaic image by highlighting key pixels.The theory of LGPQ-D feature is that we utilized the Standard Deviation(called Std in paper),which can reflect the dispersion of a data set.The effective combination of Std and LGPQ can capture the dispersion degree of different LGPQ feature values under different window sizes for each pixel,we calculate this dispersion degree and regard as the LGPQ-D value for each pixel.After analysis,the dispersion degree is proportional to the possibility of multiple role of pixels.Experiment results shows that the more discrete a pixel is(same as the LGPQ-D value of the pixel more larger),the more likely the pixel is a key pixel.2)Global-LGPQ feature.As a local texture feature,the LGPQ feature cannot reflect the globality of an image,since we proposed Global-LGPQ.As the name implies,GlobalLGPQ can perfectly cover the defect that LGPQ cannot capture the global situation.Global-LGPQ feature is different from the existing features.On the one hand,GlobalLGPQ features can reflect the relationship between pixels and local neighborhood pixels,which is called inter-block correlation,this relationship reflects the local nature of GlobalLGPQ feature.On the other hand,the Global-LGPQ feature can reflect the relationship between pixels and the non-neighborhood pixels in the image,which is called intra-block correlation,this relationship reflects the global nature of Global-LGPQ feature.Based on the above two texture features,two image splicing blind identification models are proposed,which are divided into four stages: 1)image preprocessing stage;2)feature extraction stage;3)feature dimensionality reduction stage;4)classifier training and testing stage.The two models are different in feature extraction,and the other three stages are basically the same.The procedure of the proposed two models can be summarized as follows: Firstly,in the image preprocessing stage,RGB image is transformed into YCbCr image,and Gabor transform is applied to Cr channel to obtain multi-scale and multi-directional Gabor images,then the texture features of the above Gabor images are extracted,and non-negative matrix is utilized to reduce the dimension of features.Finally,the feature vector after dimension reduction is fed to SVM.