Research On Source Camera Model Recognition Algorithm Based On Attention Mechanism

In today's society,artificial intelligence products are more and more widely used in people's daily lives.Cameras,smart phones and other camera devices are becoming indispensable tools for people to record their lives.However,technological progress is a double-edged sword.In recent years,incidents of using computer software to maliciously tamper with digital pictures to distort facts have occurred.Some criminals can take advantage of this opportunity to obtain improper benefits,creating unnecessary fear in society,security issues and real problems of digital images have increasingly become the focus of attention.Therefore,digital image forensics technology has also attracted much attention,and this paper is a research on the important branch of digital image forensics for source camera model recognition.At present,there are two mainstream methods for source camera model recognition.One is the traditional method,which mainly relies on processing the pattern noise such as Photo Response Non-Uniform noise,filtering it,and obtaining the residual pattern noise part.Compare with the existing pattern noise to identify and classify the query pictures;the other is based on deep learning,which mainly extracts the feature map in the pictures through the convolutional neural network,and then the network autonomously extracts the characteristics of the image are processed,and finally the classification results are given to achieve the purpose of source camera model recognition.This paper chooses deep learning methods to carry out source camera model recognition.An attention mechanism is introduced to classify the source camera model.Through two branches in the Residual Attention Network:the backbone branch and the mask branch,the features are extracted from the picture at the same time,and then the featrures extracted by the two branches are fused according to different ratios to generate a feature map.The Residual Attention Network obtains the final classification result through the Softmax layer by processing the feature map.The Residual Attention Network in this paper adopts the method of extracting patch as the input.Therefore,this paper has improved patch extraction,data set,and data preprocessing.Finally,the method proposed in this paper achieves a recognition accuracy of 98.63%on the source camera model recognition.In view of the low recognition rate of some special brand cameras(mainly Sony cameras)in the current network,this paper proposes a Cascade Network to improve the algorithm's recognition rate for this type of camera.For Sony source camera recognition,the network structure based on the Residual Network is used for training and testing,and the patch extraction rules are re-improved,the network data pre-processing operation is improved,and the problem of low recognition rate of Sony cameras is greatly improved,which has achieved an overall accuracy rate of 93.29%for the three Sony classes,and has also significantly improved the serious confusion in the Sony camera category.During the test,the Residual Network and the Residual Attention Network are fused using the idea of cascading,so that the network is integrated as a whole.The image to be tested only needs to be sent to the input layer,and the final network will automatically give a final recognition.The classification result is convenient,fast and easy to apply.The proposed framework can also be used as a reference for other data sets.After experimental validation,the network proposed in this paper can perform the task of source camera model recognition excellently,and has certain robustness and scalability.