Research On Color Image Median Filtering Forensics Based On Quaternion Convolution Neural Network

Image,as a carrier of information transmission,has been widely spread and applied with the development of science and technology.The image carries the function of transmitting information,and at the same time,the malicious tamper will tamper with the content carried by it,so that the tamper can achieve its purpose.Image forensics is a technique developed in this context to detect whether an image has been tampered with.Median Filtering(MF)is a smooth denoising technique because its good edge-holding characteristics are widely used,but it is also often used by tampering to mask operations that modify image content,such as stitching and scraping.And caused traces of edge changes.Therefore,research on image median filtering detection has become increasingly important.In this paper,convolution neural network technology is used to study the color image median filtering forensics.The main research contents are as follows:(1)Based on the quaternion discrete cosine transform(QDCT),an image forensics scheme based on QDCT is proposed.By analyzing and comparing the advantages and disadvantages of QDCT amplitude and amplitude matrix,combined with the structural characteristics of convolution neural network,a quaternion amplitude-angle convolution neural network is proposed for the median filtering forensic problem of color image,compared with conventional convolution.The effect of neural network on color image median filtering forensics has improved the detection accuracy.(2)Based on the quaternion convolution,the components of the conventional convolution neural network are extended to the quaternion domain.An image forensics scheme based on quaternion convolution neural network is proposed.By studying the process of quaternion convolution and pooling,the convolution neural network operation for quaternion is realized.The color image can be directly represented as a quaternion matrix and input into the proposed network for training and classification.Since the color channel information of the color image can be fully utilized,the detection effect is improved compared with the quaternion amplitude-angle convolution neural network mentioned above.