Research On Source Camera Identification Of Digital Image Based On Deep Learning

Digital images are an important information carrier in the digital age.With the in-depth development and popularization of image processing technology and applications,it becomes easier to tamper with and forge images.Digital images face a serious problem of not being trusted,so the research of digital image forensics is of great significance.Image source forensics is an important branch of digital image forensics.Source forensics is to identify the information related to the equipment used to capture images from unknown sources.Although existing research methods have been developed in camera source identification,there are still many problems.Although the existing research methods have been relatively mature in camera sources identification,there are still many problems:1.A large amount of training data is still required when training the model;2.It is difficult to distinguish challenging camera models;3.Existing image camera source models have poor performance for JPEG recompressed images。Aiming at these problems,the main achievements of this paper include:(1)A camera source identification algorithm based on multi-scale feature fusion is proposed.The algorithm uses the feature map output by the shallow network to subtract the feature map mapped by several convolutional layers pixel by pixel,whose purpose is to eliminate the higher semantic features which is related to image content information learned by the deep convolutional layers.A series of experiments are completed based on the Dresden image set.The experimental results show that when only 20 images are selected for each model,our model(MCCmodel)can identify the source camera of 23 models in the Dresden forensic image library and the accuracy rate is as high as 99.51%.The source identification of all camera models in the Dresden forensic image library is also carried out,including challenging camera models,and the accuracy rate is 95.3%,of which the accuracy rate of Sony DSCH50,Sony DSCT77 and Sony DSCW170 reached 82%,99.5%,and 90.7%,respectively.At the same time,the performance of our model is compared with several state-of-the-art methods.(2)A camera source identification algorithm based on multi-column constrained convolution layer and attention mechanism is proposed.This algorithm uses the feature fusion of multi-column constrained convolution layers with different scales to adaptively learn sufficient features that are conducive to camera source identification and combines SENet to weight different feature channels,enhances the features that are useful for the current classification task and suppresses the features that are not useful to enhance the importance of the camera source identification related features.The experimental results show that our model(MCA-model)has an accuracy of 99.61%for camera identification of 9 brands and 98.6%for 17 models of cameras.All the test images are recompressed by JPEG with different quality factors(100,95,90),and the average accuracy rate of identifying image blocks in various scenes reaches 82.34%,which is about 3%higher than the existing methods for recompressed images under the same conditions.Using the recompressed images to train the model,in four scenarios:original images,JPEG compression with different quality factors(100,95,90),the average accuracy rate is as high as 86.53%.Experimental results show that our model not only outperforms existing convolutional neural network based methods on raw images,but also has strong competition for JPEG compressed images compared to the current camera source methods for recompressed images.