Research On Color Image Splicing Detection Based On Feature Combination Deep Learning

Today,our lives flooded with a variety of tampered images.One of the most common ways to tamper with images is splicing,which brings various social problems.For this tampering method,many detection schemes have been proposed.But most methods analyze the inherent features of the image,which extract the corresponding features and use a classifier to classify them.However,these methods usually rely on a specific feature,and need to manually extract.This way easily leads to problems such as too strong dependency and insufficient feature extraction.At the same time,in addition to detecting splicing,it is often necessary to know which region is the splicing region.This is more difficult and more meaningful than detection.The current existing splicing localization methods have the problem of low detection accuracy and serious edge missing.For the above-mentioned splicing detection and localization problems,this thesis proposes color image tampering detection and localization schemes based on deep learning and attention.Main tasks are as follows:1)For the problems of relying on single feature and inadequacy of feature extraction in the splicing detection of current methods,this thesis proposes a neural network detection scheme based on weight combination strategy.For color image splicing detection,the splicing marks are often reflected in many aspects.The single feature as the detection basis cannot fully excavate these splicing marks.Therefore,this solution uses YCb Cr features,edge features,and pattern noise(PRNU)features for splicing detection.At the same time,in order to solve the problem of the insufficient extraction of feature,this solution also proposes a weight combination strategy.By adding weights to the above three features and combining them according to the weight ratio,the purpose of further removing redundant information is achieved.Experiments show that this scheme can achieve better results.2)For the problems of low accuracy and edge missing of color splicing localization,this thesis proposes a splicing localization scheme based on attention deep learning.FCN is selected as the basic framework for this scheme,and edge features are selected as the detection features.The pixel-level detection capability of FCN and the edge detail information possessed by edge features are valid for these problems.At the same time,in order to further enhance the splicing localization capability,this solution also uses an attention mechanism.The enhanced extraction capability of this mechanism can further enhance the extraction of the content and location of input features.Experiments show that this scheme is effective.