Research On The Data Acquisition Based On Depth Learning

With the development and popularization of powerful image processing software, digital images tend to be easily edited and forged without leaving any traces. The modified photos can bring us some visual enjoyment and entertainment, however, once the fake pictures are being abused, it will lead to serious consequence. Human cognition may be misled and personal privacy may be invaded. Moreover, image forgery in some fields, like economy, politics, media, military affairs, healthcare, judicature, will have great damage on the country and society. As a result, the research on digital image forgery forensics has been a very popular and significant research topic.There are various digital image forgery methods and image splicing is one of the most common forgery approaches. The existing image forgery detection methods are mainly based on the feature extraction of the changes of image intrinsic information caused by image splicing. Generally speaking, these methods use sallow learning to realize the classification. As a pop new network model, deep learning has been used in fields like pattern recognition and target tracking, which overcomes the shortcomings of sallow learning such as inadequate learning and lack of depth. This work mainly focuses on the research on passive detection of image splicing and explores how to apply the theory of deep learning to the issue.In this paper, the image splicing detection algorithm based on deep learning is discussed in detail. We use deep belief network (DBN) to realize the image splicing detection by mimicking a person’s learning procedure. Firstly, the Markov feature transition probability matrix is extracted from an image’s Discrete Cosine Transform (DCT) coefficients as the input of DBN’s bottom layer. Secondly, contrastive divergence (CD) algorithm is used to train the Restricted Boltzmann Machine (RBM). After that, BP algorithm is applied to fine-tune the parameters of each layer with labeled data. Finally, the trained deep network is ready to predict the image category. The experimental results show that the average of accuracy can achieve91.26%which is higher than the methods based on sallow learning. In addition, the effects of main characteristic parameters of DBN on image splicing detection are achieved by a multitude of experiments, including the number of hidden layers, the number of hidden layer units, learning rate, iteration times and the number of training samples. Finally, the optimal network parameters are attained, which makes meaningful contribution to the application of DBN in the field of passive image forgery detection. DBN can fully develop the intrinsic information of data by unsupervised learning and supervised network fine-tuning, which opens up new horizon for the problem of image forgery detection.