Copy-move Forgery Detection Based On Block Matching And Clustering Algorithm

The digital era facilitates people's life.However,it also expands the scope of counterfeiting,reduces the cost of counterfeiting and allows some offenders to fake data,information and news for certain interests.Image forgery is also one of the simple.Quick and easy methods for multimedia technology and digital image processing software,image forgery has become indiscriminate.Image tamper detection technology came into being.Copy-move forgery is a common tamper technique for image forgery.In view of this tampering technology,the current mainstream research directions are divided into feature point detection and block matching detection,each of which has its own advantages.However,in terms of detection efficiency,the accuracy of feature point detection is higher and the applicable range is wider.However,feature extraction and matching results rely too much on the selected scale space and pixel gradient direction.There is more room for improvement in block matching detection.The detection method based on block matching in this paper focuses on improving the accuracy of block matching detection and reducing the time complexity.The main research methods are as follows:1.Feature extraction and clustering algorithm applicability analysis: After the color image is converted into a grayscale image,it is more common to tamper with a certain size of the block to extract certain features of each block for block matching.In this paper,we make innovations in feature selection.In addition to selecting common spatial domain features,we apply the concept of self-coding in machine learning and deep learning to extract features.The two types of features are clustered to analyze the suitability of the clustering algorithm for copying tampering in the same graph.2.Spatial domain features clustering algorithm: The spatial domain feature clustering model is relatively simple.The k-means clustering algorithm is used to cluster the feature matrix directly.Then the Euclidean distance determination and the random sample consensus(RANSAC)algorithm are used to remove the abnormal block,which calls abnormal block removal algorithm.Finally,the correct detection area is obtained.3.Sparse autoencoder features clustering algorithm: This one is relatively complex and combines the clustering algorithm to detect images.Select a sparse self-encoder training a given sample set,the sample set from the same atlas with a number of tampered copies of the same composition,through the self-learning to find the internal structure of the image and the underlying law to get hidden layer weight parameter matrix,so fast The sparse self-coding feature of the test set is extracted,and the feature clustering is performed by combining the kmeans clustering algorithm.The clustering result will remove the smooth area of the image,which greatly reduces the computational complexity.K-means clustering results can be detected as a simple structure of the image.The images with complex structure still need further testing.K-means secondarily clusters the texture features and gray features of the remaining blocks.If the detection result contains abnormal blocks,an abnormal block removal algorithm is used to obtain the correct detection area.The experimental results show that the SAE feature clustering algorithm improves the overall accuracy by 14.3% and the time efficiency by 72% compared with other algorithms.For JPEG-compressed images,it has strong robustness.Setting the value of compression factor [75,95],the precision and recall value of the detection results are above 0.8.Aiming at the rotation operation of image tampering region,the algorithm in this paper has some adaptability.For the images with low structural complexity,the five angles were chosen for testing successfully.In terms of time efficiency,the proposed algorithm greatly reduces the time complexity of the detection process.For block-based matching detection,the clustering algorithm is suitable for different features,and the efficiency of the feature matching process is improved through block clustering.Feature extraction which can be quickly extracted through self-learning is combined with the field of deep learning to greatly reduce the detection time.However,the feature clustering algorithm model needs to be improved for operations such as rotation and blurring of complex structure images.