Image Inpainting Passive Forensics Research Based On Deep Learning

Image inpainting is a technique that uses information from known areas of an image to repair missing or damaged areas in the image.With the rapid development of image inpainting technology,image editing software based on this is more and more common in life,they are powerful and easy to operate,and people can easily edit images using this software,Once the image inpainting technology is used to maliciously remove the content of the image,it will bring a crisis of confidence to the real image.Therefore,the image inpainting forensics technology to verify whether the image is real and complete has become a research hotspot in the field of security,and it also has important practical significance.The current research on image inpainting forensics can only detect a certain type of image inpainting.Although they can detect the inpainted area to a certain extent,the detection effect is not ideal,and they cannot simultaneously detect the exemplar-based technology and the deep learning-based technology image inpainting area that look like real images.at the same.In response to this problem,this paper conducts in-depth research on passive forensics of image inpainting based on deep learning.The main work is as follows:In this paper,the corresponding datasets are first created using two algorithms with excellent inpainting effects at present.After an in-depth understanding of traditional inpainting forensics,a passive forensic method for image inpainting based on dual branches network is proposed.The High-Pass filtered Convolutional Network(HPCN)in the dual branch network first uses a set of high-pass filters to attenuate the low-frequency components in the image and then uses four residual blocks to extract features,and capture the differences in content and texture between the inpainted region and the pristine region.The Dual-Attention Feature Fusion(DAFF)branch in the dual branch network first uses a preprocessing module to add local binary pattern feature maps to the image and then uses a dual-attention convolution block to adaptively integrate local features and global features dependent,to capture the differences in content and texture between the inpainted region and the pristine region.The fusion module fuses the outputs of the two branches at multiple scales to obtain accurate detection results.Extensive experiments show that the method not only outperforms existing image forensics algorithms but also can detect sample block image inpainting and deep learning image inpainting simultaneously.In addition,sharper and more accurate edges can be obtained on the boundary of the locating image to remove the object,and it is resistant to post-processing operations of the image.Next,to further improve the efficiency and accuracy of passive forensics in image inpainting.A passive forensics method based on the dual-stream improved U-network is proposed.The encoder structure based on the RGB stream directly extracts image features and automatically learns the inconsistent features between the inpainted region and the pristine region in the RGB image through training.The encoder structure based on noise flow performs feature extraction on the local noise residual map of the image and learns the inconsistent features between the inpainted region and the pristine region in the noise residual image.The two complement each other and are passed to the multi-feature map decoder for decoding.The improved U-shaped network continuously strengthens and consolidates the learning through the ring residual.The U-shaped structure transfers the feature map jumps of different layers of the encoder to the decoder structure to improve the precision of the inpainted region in which the decoder is located.After experiments,it is found that this method can effectively improve the accuracy of image inpainting passive forensics,and plays a positive role in promoting image inpainting forensics to practical applications.