Deep counterfeit image detection method fusing depth learning and width learning

Abstract

The invention relates to a depth counterfeit image detection method fusing depth and width learning, which comprises the following steps of: firstly, converting an image to be detected into a gray level image, segmenting the gray level image into two groups of image blocks, extracting a centralized frequency domain magnitude spectrum of each image block, applying an attention mechanism to the centralized frequency domain magnitude spectrum, and carrying out channel connection on the two groups of image blocks to obtain a primary feature; secondly, constructing a channel convolution self-encoding module, performing feature fusion on the primary features by using an encoder of the pre-trained channel convolution self-encoding module to obtain two intermediate features, and respectively taking the two intermediate features as inputs of a feature mapping stream and a feature enhancement stream to obtain two mapping features and an enhancement feature; and finally, constructing three classifiers according to a width learning system principle, and carrying out weighted average on output results of the three classifiers to obtain a final detection result. According to the method, the attention mechanism is applied to the image block, the region with obvious tampering traces can be concerned from global information, data and time required by model training are less, and both accuracy and efficiency are realized.

Description

technical field [0001] The invention belongs to the technical field of fake image detection, in particular to a deep fake image detection method integrating depth and width learning. Background technique [0002] With the development of computer technology, it is more and more easy to tamper with or synthesize images through artificial intelligence (Artificial Intelligence, AI), ProGAN, AttGAN and other technologies. The images obtained by these methods are called deep fake images. At present, deep fake images have reached At the level of falsehood, there is a greater threat to the security field. [0003] At present, classic deep learning models such as VGG, DenseNet, and Xception are widely used to detect deep fake images. Although the deep learning model has achieved good detection results on a single data set, with the improvement of detection performance, it also brings the amount of parameters and data requirements. The sharp increase in the number of images requires ...

AI Technical Summary

Problems solved by technology

Most feature extraction methods based on GAN fingerprints mainly rely on the structure of the GAN generator, which makes the detection model overfit on the specific forged images participating in the training, and cannot handle the forged images generated by unknown generators, feature representation capabilities and methods poor generalization

Images