General image invisible watermark detection method based on few-sample learning

Abstract

The invention discloses a general image invisible watermark detection method based on few-sample learning, which comprises four steps of watermark feature extraction, graph construction, label propagation and loss calculation. A feature embedding step of the method is improved on the basis of a few-sample transduction propagation network framework; the transformed feature embedding part is composed of three sub-steps of preprocessing, multi-scale feature fusion and feature embedding. The method can be used as a general invisible watermark detector in an actual image invisible watermark detection task, and a specific invisible watermark embedding algorithm does not need to be trained independently. In addition, an image invisible watermark general detection model can be trained on the basis of a few watermark images of an invisible watermark embedding algorithm, so that the actual image invisible watermark detection process is more convenient, the limitation of actual conditions is better met, and the detection of the image invisible watermark under the actual condition can be really met.

Description

technical field [0001] The invention relates to a detection method in the computer field, in particular to a general detection method for image invisible watermarks based on few-sample learning. Background technique [0002] The image invisible watermark detection method based on deep learning generally only has a good detection effect on a specific invisible watermark embedding algorithm, and the image invisible watermark detection model based on deep learning needs a large amount of watermark image data sets corresponding to the watermark embedding algorithm during training, while In the actual image invisible watermark detection task, it is often necessary to detect an invisible watermark embedding algorithm whose detection model is unknown, and a large amount of watermark image data sets for this invisible watermark embedding algorithm are also difficult to obtain. How to train a general image invisible watermark detection method under the condition of a small number of ...

AI Technical Summary

Problems solved by technology

[0002] The image invisible watermark detection method based on deep learning generally only has a good detection effect on specific invisible watermark embedding algorithms, and the image invisible watermark detection model based on deep learning needs a large amount of watermark image data sets corresponding to the watermark embedding algorithm during training, while In the actual image invisible watermark detection task, it is often necessary to detect invisible watermark embedding algorithms with unknown models, and a large number of watermark image data sets for this invisible watermark embedding algorithm are also difficult to obtain

How to train a general image invisible watermark detection method under the condition of a small number of watermark image samples of the target invisible watermark embedding algorithm is a great challenge

Images