Research On Forged Image Detection Method Based On Frequency Domain Conversion

The image generation technology based on deep learning method shines brightly in the fields of style transfer and face attribute editing,but it also lowers the threshold of image forgery and brings negative effects to people’s reputation and other aspects.In such an environment,the forged image detection technology has become one of the research hotspots in the field of image processing.Existing detection techniques are mainly oriented to lowresolution images,and the detection effect is good in the generated results of a specific forged image generation method,but the detection effect is poor on images generated by other methods.However,in practical application scenarios,most of the image sources are unknown,so targeted detection cannot be performed.Moreover,the resolution of the current forged images is getting higher and higher,and the images need to be down-sampled to adapt to the input format of the training model,which will cause the loss of image information and affect the detection effect.Be directed against the problem mentioned above,this paper proposes to improve the generalization of the model from the perspective of frequency domain and improve the adaptability of the model to high-resolution images.The main research content includes:1)In view of the difference in the effect of spatial domain images and frequency domain images in the detection method of forged images,the particularity of the application of frequency domain in the field of forged image detection is analyzed and summarized,and the preference of forged image detection feature extraction on the underlying information is found,and it is demonstrated that compared with the spatial domain image,the highfrequency frequency domain image brings the improvement in detection generalization.Based on the underlying principle of the frequency domain conversion method,the input preprocessing method of the neural network architecture based on the frequency domain image is redesigned,which provides a good theoretical basis for the subsequent detection work.2)Aiming at the problem of weak generalization of deep fake image detection in traditional spatial domain detection methods,a forged image detection method is proposed to enhance the generalization of detection models through frequency domain transformation and highfrequency filtering.This method uses the high-frequency information of images to learn the common features of forged images combined with the attention mechanism of channel and space,and designs a frequency domain detection method to mine the common latent features of forged images.Experiments show that this method can improve the detection generalization of the model on 11 popular forged image generation methods while reducing the time required for training and the order of magnitude of training images.3)In view of the problem that the current high-quality generated images need downsampling to adapt to the existing detection framework,resulting in information loss and reducing the detection effect,this paper designs and implements a forged image detection based on unconventional discrete cosine frequency domain transformation.The method uses unconventional discrete cosine image segmentation to reduce the spatial input scale of the image,so that the network can accept images that are at least 16 times larger than the existing images,and the information of the input image is guaranteed to be complete.At the same time,two types of channel screening methods and attention mechanism are combined to improve the accuracy of forged image detection on the basis of preserving input image information.The experimental results show that the method improves the adaptability to unknown image input and obtains a stronger generalization effect.