Image Splicing Forgery Detection Based On Deep Learning

With the development of deep learning technology,various convolutional neural network models emerge in an endless stream,and the efficiency of detection algorithms based on computer vision has also been greatly improved.As a small branch of computer vision technology application field,image mosaic tampering detection is rapidly rising in the field of weak feature extraction.The original method based on manual feature extraction is not only difficult to extract features,but also has extremely low detection efficiency.Therefore,the method based on manual feature extraction is no longer applicable.Various methods based on deep learning technology to detect and locate tampered areas of image splicing have begun to receive attention.It has been extensively studied and is gradually playing a leading role in the field of image security.Based on the UNet and U~2Net models,this study optimizes the model structure,introduces a mixed attention mechanism,and designs three different network structures in combination with noise characteristics and frequency domain characteristics,and names them AttRUNet,SU~2Net and SFU~2Net respectively,and finally realizes image stitching Determination and segmentation of tampered areas.The main work of this research is as follows:1.On the basis of the existing serial mixed attention mechanism,a parallel structure of channel attention mechanism and spatial attention mechanism is proposed,and it is used as the design of the model attention module,and then combined with UNet with the residual module The model is combined to propose a network structure AttRUNet based on the mixed attention mechanism.2.Introduce the image residual noise spatial domain rich model SRM extraction module in the image splicing forgery detection,and combine the feature pyramid module PPM to obtain the noise image feature extraction module,combine it with U~2Net in the salient target detection,and propose based on noise characteristics SU~2Net structure for image tampering detection and localization.3.Combine the convolutional neural network with the frequency domain characteristics,set a learnable filter,divide the frequency domain coefficients of the image into high frequency,intermediate frequency and low frequency,and then divide the spatial domain images corresponding to three different frequency coefficients It is used for training,and combined with the spatial convolution pooling feature pyramid module ASPP to design a frequency domain feature extraction module,and finally merged with SU~2Net to propose an image splicing tamper detection and localization neural network SFU~2Net based on frequency domain feature fusion.4.Create an image splicing and tampering dataset through the COCO2017 dataset,and conduct ablation experiments,comparison experiments,and noise interference tests on each network,which fully proves the effectiveness of each module,the performance of the entire model,and the robustness to noise interference.And use the front-end and back-end separation mode,and use the Flask framework to deploy the model on the Web side.