Image recognition system, method and device based on image channel correlation

Abstract

The invention belongs to the technical field of digital images, and particularly relates to an image recognition system, method and device based on image channel correlation. The system. The system comprises a plurality of mixed feature extraction modules which are used for obtaining fusion features of features, mixed with channel and neighborhood correlation, of a to-be-identified image; the feature fusion module is used for superposing the fusion features output by the plurality of mixed feature extraction modules into a total feature map, and fusing the total feature map into a high-dimensional feature representation through a plurality of convolution; and the image classification module is used for respectively acquiring classification probabilities of the natural image and the rendered image based on the high-dimensional feature representation, and outputting the image with the larger classification probability as an recognition result. According to the invention, the recognitionaccuracy and efficiency of the convolutional network for the rendered image are improved.

Description

technical field [0001] The invention belongs to the technical field of digital images, and in particular relates to an image identification system, method and device based on image channel correlation. Background technique [0002] Visual information is an efficient and reliable medium of interaction due to its immediacy and comprehension. As an important carrier of visual information, natural images are used to record and disseminate important information. In recent years, with the rapid development of entertainment industries such as movies and games, the images created by computer graphics rendering technology have become more and more realistic, so that it has become difficult for the naked eye to distinguish rendered images from photographic images. Therefore, the identification of rendered images has become an important research topic in the field of digital image forensics. [0003] Before the emergence of deep learning methods, the vast majority of methods were des...

AI Technical Summary

Problems solved by technology

Post-processing operations such as image interpolation and image compression can eliminate image generation traces to a certain extent, and weaken the difference between rendered images and natural images in feature space, resulting in a decrease in identification accuracy.

Images