CN-121998847-A - Image water-jet printing method, device, equipment, medium, product and model training method

CN121998847ACN 121998847 ACN121998847 ACN 121998847ACN-121998847-A

Abstract

The embodiment of the application relates to an image water-jet printing method an apparatus, device, medium, product, and model training method. The method comprises the steps of erasing a first target image in a first format through a rough erasing sub-network of a target watermarking model to obtain a first image feature of the first target image, wherein the first target image is an image containing a watermark, respectively carrying out regional convolution on the first image feature through an image region corresponding to the first image feature according to a preset regional division mode through a thinning sub-network of the target watermarking model to obtain a second image feature, and carrying out up-sampling operation on the second image feature to obtain a second target image, wherein the second target image is an image not containing the watermark.

Inventors

ZHANG YUNRU
DONG HUIYAN
Zeng chan
XIE YING

Assignees

人保信息科技有限公司
中国人民保险集团股份有限公司

Dates

Publication Date: 20260508
Application Date: 20251230

Claims (10)

1. A method of image watermarking, the method comprising: Erasing a first target image in a first format through a rough erasing sub-network of a target watermarking model to obtain first image characteristics of the first target image, wherein the first target image is an image containing a watermark; Respectively carrying out region convolution on the first image features according to the image regions corresponding to the first image features in a preset region division mode through a refinement sub-network of the target de-watermarking model so as to obtain second image features; And carrying out up-sampling operation on the second image characteristic to obtain a second target image, wherein the second target image is an image which does not contain a watermark.
2. The method according to claim 1, wherein the performing, by the refinement sub-network of the target de-watermarking model, region convolution on the first image feature by using the image region corresponding to the first image feature in a preset region division manner, to obtain a second image feature includes: Downsampling the first image feature to obtain a sparse image feature map, wherein the sparse image feature map is a set of sparse feature vectors corresponding to feature points of the sparse image feature map; For a first target feature point in feature points of the sparse image feature map, dividing the sparse image feature map into a plurality of feature areas by taking the first target feature point as a center; Acquiring a region image feature vector of each of the plurality of feature regions; Aggregating regional image feature vectors of the plurality of feature regions, and updating sparse feature vectors corresponding to the first target feature points by the aggregated regional image feature vectors to obtain updated sparse feature vectors of the first target feature points; And updating sparse feature vectors corresponding to all feature points in the sparse image feature map to obtain the second image feature.
3. The method of claim 2, wherein the acquiring the region image feature vector for each of the plurality of feature regions comprises: Acquiring feature vectors of second target feature points of target feature areas in the feature areas; acquiring a characteristic point weight matrix; And performing activation and maximum pooling operation on the feature vector of the second target feature point and the feature point weight matrix to obtain regional image feature vectors of the target feature regions in the plurality of feature regions.
4. The method according to claim 2, wherein aggregating the regional image feature vectors of the plurality of feature regions, and updating the sparse feature vector corresponding to the first target feature point with the aggregated regional image feature vector, to obtain the updated sparse feature vector of the first target feature point, includes: acquiring a regional weight matrix; obtaining the product of the regional weight matrix and the regional image feature vector; And summing products of the regional weight matrixes of all the regions and the regional image feature vectors to obtain the updated sparse feature vectors of the first target feature points.
5. The method according to claim 2, wherein updating the sparse feature vectors corresponding to all feature points in the sparse image feature map to obtain the second image feature comprises: updating the sparse feature vectors corresponding to all feature points in the sparse image feature map, and taking the set of the sparse feature vectors updated by all feature points as the second image feature.
6. A method of training an image watermarking model, the method comprising: Training the image watermarking model by adopting a LOGO data set, wherein the LOGO data set comprises a watermarked image and a non-watermarking truth value image corresponding to the watermarked image; and in the training process, adopting AdamW optimizers, cosine annealing learning rate adjustment strategies and L1 loss functions to optimize network parameters of the image watermarking model.
7. An image watermarking apparatus, the apparatus comprising: The first acquisition module is used for erasing the first format characters of a first target image through a rough erasing sub-network of the target watermarking model so as to obtain first image characteristics of the first target image, wherein the first target image is an image containing a watermark; the second acquisition module is used for carrying out regional convolution on the first image features through the refinement sub-network of the target de-watermarking model on the image regions corresponding to the first image features according to a preset regional division mode so as to obtain second image features; And the third acquisition module is used for carrying out up-sampling operation on the second image characteristics so as to obtain a second target image, wherein the second target image is an image which does not contain a watermark.
8. An electronic device, comprising: A memory having a computer program stored thereon; A processor for executing the computer program in the memory to implement the steps of the method of any one of claims 1-6.
9. A non-transitory computer readable storage medium having stored thereon a computer program, characterized in that the program when executed by a processor realizes the steps of the method according to any of claims 1-6.
10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1-6.

Description

Image water-jet printing method, device, equipment, medium, product and model training method Technical Field The embodiment of the application relates to the technical field of image processing, in particular to an image dewatering printing method, an image dewatering printing device, image dewatering printing equipment, an image dewatering printing medium, an image dewatering printing product and a model training method. Background In the field of insurance, image watermarking techniques may be used to remove watermarks from security documents, contracts, or other documents of interest to facilitate copying, scanning, or electronic version archiving. This helps to protect the copyright and privacy of the file and to ensure the authenticity and integrity of the information. The current mainstream watermarking technology splits the whole watermarking process into two tasks, watermark positioning and watermark removal. The two tasks realize end-to-end training through a deep learning technology, and further finish the water mark removing process cooperatively. In particular, the technique employs Unet symmetrical structure networks to encode and decode images. In the encoding stage, unet network extracts the features of the image through a series of convolution operations while gradually reducing the resolution of the image. This process aims to encode watermark information in the image into the underlying features of the network. Then, in the decoding stage, unet network executes two different tasks through two up-sampling branches, the first branch focuses on the prediction of watermark mask, namely, generates a mask indicating the watermark position, and the second branch restores the original picture after watermark removal. However, the traditional convolutional neural network adopted in the downsampling process of the encoder in the prior art has limited receptive field, cannot capture global information of an image, has various image watermarks, can be distributed in a small area of the image, can occupy the whole image, and has limited perceived area and cannot pay attention to the global information of the whole image, so that the effect of removing the large-area watermark is poor, and the integrity and the authenticity of the image after the watermarking are affected. Disclosure of Invention In order to solve the problems, the application provides an image water-jet printing method, an image water-jet printing device, an image water-jet printing medium, an image water-jet printing product and a model training method. According to a first aspect of an embodiment of the present application, there is provided an image watermarking method, the method comprising: Erasing a first target image in a first format through a rough erasing sub-network of a target watermarking model to obtain first image characteristics of the first target image, wherein the first target image is an image containing a watermark; Respectively carrying out region convolution on the first image features according to the image regions corresponding to the first image features in a preset region division mode through a refinement sub-network of the target de-watermarking model so as to obtain second image features; And carrying out up-sampling operation on the second image characteristic to obtain a second target image, wherein the second target image is an image which does not contain a watermark. In one embodiment, the performing, by the refinement sub-network of the target watermark removal model, region convolution on the first image feature by using the image region corresponding to the first image feature in a preset region division manner, to obtain a second image feature includes: Downsampling the first image feature to obtain a sparse image feature map, wherein the sparse image feature map is a set of sparse feature vectors corresponding to feature points of the sparse image feature map; For a first target feature point in feature points of the sparse image feature map, dividing the sparse image feature map into a plurality of feature areas by taking the first target feature point as a center; Acquiring a region image feature vector of each of the plurality of feature regions; Aggregating regional image feature vectors of the plurality of feature regions, and updating sparse feature vectors corresponding to the first target feature points by the aggregated regional image feature vectors to obtain updated sparse feature vectors of the first target feature points; And updating sparse feature vectors corresponding to all feature points in the sparse image feature map to obtain the second image feature. In one embodiment, the acquiring the region image feature vector of each of the plurality of feature regions includes: Acquiring feature vectors of second target feature points of target feature areas in the feature areas; acquiring a characteristic point weight matrix; And performing activation and maximum pooling operation on