Exactly recognizable distortion prediction model and method suitable for machine vision task

By using the combined model of JRD image generation network, discriminant network and residual-guided regression network in machine vision tasks, the problems of low efficiency and lack of universality in the prior art are solved, and efficient and accurate predictive of appropriately identified distortions are achieved. , suitable for a variety of machine vision tasks.

CN119011835AActive Publication Date: 2024-11-22HANGZHOU DIANZI UNIV
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202410948217.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-22
Estimated Expiration
2044-07-16

AI Technical Summary

Technical Problem

The appropriately recognizable distortion prediction model in existing machine vision tasks is inefficient and lacks universality, which cannot meet the practical application needs. In particular, the full-reference or semi-reference prediction model requires encoding and compression of the original image, and the accuracy performance is low. .

Method used

Using a model that includes JRD image generation network G, discriminant network D and residual guides JRD regression network R, false JRD images are generated through uncompressed original images, and regression prediction is performed using residual information, optimal quantization parameters are generated, and discriminant networks are identified. Training against the generative network to improve the accuracy of the generated results, using a U-Net structure and a multi-scale discriminator, combined with the VGG16 feature extractor and an adversarial learning framework, the loss function is optimized to improve prediction accuracy.

Benefits of technology

It achieves improved reference-free prediction efficiency and accuracy, reduces the redundancy of image and video data in transmission and storage, and is suitable for a variety of machine vision tasks and meets practical application needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119011835A_ABST
    Figure CN119011835A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of video image compression, and discloses a just recognizable distortion prediction model and method suitable for a machine vision task, and the model comprises a JRD image generation network G, a discrimination network D and a residual error guide JRD regression network R; the JRD image generation network G takes an uncompressed original image as input and takes a false JRD image as output; afterwards, the residual error of the generated image and the original image serves as compensation information of the original image and is input into a residual error guiding JRD regression network R together, a predicted JRD score is finally output, and a judgment network D and a generation network G are trained together in an adversarial mode so as to help the G to generate a generation result closer to a true JRD image. In the non-reference model provided by the invention, the original image does not need to be compressed, coded and decoded for a plurality of times, so that the prediction efficiency is greatly improved. Meanwhile, in the regression network, the residual error of the generated image is input as supplementary information, so that the final prediction result of the model is more accurate and reliable.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of video image compression, and in particular relates to a precisely identifiable distortion prediction model and method suitable for machine vision tasks. Background Art

[0002] Under the influence of various psychological and physiological factors, the human visual system (HVS) cannot perceive some changes in images and videos, that is, the HVS has just noticeable distortion (JND). Exploring the threshold of JND has guiding significance for improving the encoding method and bit rate control of images and videos. Image and video encoding based on a reasonable JND model can significantly reduce the bit rate required in the storage and transmission process without affecting the HVS perception level.

[0003] However, the encoding method and bit rate control of image videos based on JND threshold improvement are only optimized for the characteristics of HVS. With the improvement of hardware performance and the development of deep learning algorithms, the performance of machine vision tasks continues to improve, and video image data for machine vision algorithms rather than HVS is growing exponentially. In addition, it is well known that the accuracy of machine vision tasks depends not only on the advancement of hardware and algorithms, but also on the quality of compressed images or videos. Therefore, studying the mapping relationship between the accuracy of machine vision tasks and the quality of compressed data is of great significance for guiding image and video compression. Therefore, in order to reduce the redundancy of video and image data during transmission and storage, distortion threshold prediction methods suitable for machine vision tasks need to be studied urgently.

[0004] In recent years, research on the relationship between machine vision performance and image or video frame quality has begun to advance. Some researchers have optimized the encoding of specific machine vision tasks, including methods for generating realistic adversarial images that can deceive both machine vision algorithms and humans, and generating semantically guided JND for image classification tasks through unsupervised learning. However, these works all rely on innovations in specific machine vision tasks and are not universal for a variety of machine vision tasks.

[0005] Some researchers have been inspired by the concept of JND and have explored whether there is a distortion threshold similar to JND in machine vision. Some researchers have conducted comprehensive experiments on various machine vision tasks and proved that machine vision also has a threshold similar to JND in HVS. They have also proposed a new concept called Just Recognizable Distortion (JRD) to describe the maximum distortion of image and video encoding at an acceptable level of machine vision performance. Assume that I0 is the original image and its compressed variants are I1, I2, I3, ..., I n , where a larger n indicates a worse quality image. For the machine vision model M, the JRD of I0 is defined as:

[0006]

[0007] Where j is the quantization parameter (QP) selected for encoding, ε is an arbitrary positive integer, and M refers specifically to the target detection task. Some people have also proposed the concept of machine satisfaction rate (SMR) to evaluate image quality from the perspective of machine satisfaction, and proposed a deep learning-based model to predict the SMR of compressed images or video frames. These studies were conducted under the conditions of using standard video coding such as HEVC. The concepts of JRD and SMR are defined to describe the acceptable degree of distortion of image and video coding for machine vision tasks. However, in these studies, an accurate and efficient reference-free prediction JRD method framework has not been proposed.

[0008] In summary, existing work has preliminarily explored the relationship between the accuracy of machine vision tasks and the quality of compressed images and videos described by concepts similar to JND (such as JRD and SMR). However, related research is still insufficient, mainly reflected in the following aspects: 1) Most of the existing JRD or SMR models are full-reference or semi-reference prediction models, which means that these methods need to encode and compress the original image at least once for reference, which is inefficient. 2) The distortion threshold prediction of images and videos usually depends on specific machine vision task algorithms. The versatility of the JRD prediction framework needs to be further improved to adapt to different machine vision tasks. 3) The existing JRD prediction model has low accuracy performance and cannot meet the needs of practical applications. Summary of the invention

[0009] The purpose of the present invention is to provide a method for accurately recognizing distortion prediction suitable for machine vision tasks to solve the above-mentioned technical problems.

[0010] In order to solve the above technical problems, the specific technical solutions of the present invention for a precisely identifiable distortion prediction model and method applicable to machine vision tasks are as follows:

[0011] A precisely recognizable distortion prediction model suitable for machine vision tasks comprises a JRD image generation network G, a discriminant network D and a residual-guided JRD regression network R; the JRD image generation network G takes an uncompressed original image as input and a fake JRD image as output; then, the residual of the generated image and the original image are input into the residual-guided JRD regression network R as compensation information for the original image, and finally a predicted JRD score, i.e., an optimal quantization parameter for image compression, is output, wherein the discriminant network D is trained together with the generation network G in an adversarial manner to help G produce a generation result that is closer to a true JRD image.

[0012] Furthermore, the model uses the residual image between the original image and the corresponding fake JRD image as supplementary information. Figure 1The input residual guides the JRD regression network and finally generates the predicted JRD score, which is the optimal quantization parameter for image compression.

[0013] Furthermore, the generator part in the JRD image generation network G is implemented using the U-Net infrastructure. In the U-Net network, the input passes through a series of layers and is gradually down-sampled until the bottleneck layer, after which a transposed convolution is performed to achieve up-sampling, and a large amount of low-level information between the input and output is transmitted through a jump connection structure.

[0014] Furthermore, the discriminant network D uses a multi-scale discriminator, using two discriminators with the same structure, but making them work at different image scales. The two discriminators working at different scales are called D1 and D2. D1 and D2 work at the original scale of the real image and the generated image and the scale of 2 times downsampling, respectively.

[0015] Furthermore, the main function of the residual-guided JRD regression network R is to evaluate the degree of distortion of the fake JRD image relative to the original image;

[0016] Choose to input the original image and residual into the feature extractor respectively, and then perform linear regression calculation on their fused features. The feature extractor uses VGG16. The network parameters of VGG16 use the network parameters pre-trained on the ImageNet dataset. During training, only the network parameters of the linear regression part need to be learned and optimized.

[0017] Furthermore, the model includes a loss function L R During the training process, if the predicted JRD score φ(i) is greater than the true value Φ(i), MSE loss is used, otherwise L1 loss is used.

[0018]

[0019] N is the total number of images, φ(i) is the JRD prediction score of the i-th image, and Φ(i) is the true JRD of the image, where the predicted value φ(i) is rounded down.

[0020] The present invention also discloses a training method for a distortion prediction model suitable for machine vision tasks, comprising the following steps:

[0021] Step 1: Create a dataset containing uncompressed original images or video frames and their corresponding JRD images;

[0022] Step 2: Generate adversarial network.

[0023] Furthermore, the step 1 comprises the following steps:

[0024] For the machine vision task of YOLOv7 target detection, a dataset of 130,000 uncompressed original images or video frames and their corresponding JRD images is created, and their corresponding JRD values ​​are annotated. The original images of the dataset are from COCO2017. The YOLOv7 algorithm is used to infer the original images or video frames, and the detected object boxes are intercepted. The intercepted image I0 and its variants I1, I2, I3, ..., I after compression with different quantization parameters are n The YOLOv7 algorithm with the same parameters is used for inference, and finally the JRD image I corresponding to I0 is selected j .

[0025] Furthermore, the step 2 comprises the following steps:

[0026] The adversarial learning framework uses the basic network structure of cGAN.

[0027]

[0028] Where x is the original image, y is the true JRD image, z is the noise, G(x,z) is the output of G with x and z as input, that is, the generated fake JRD image. In the above formula, G tries to minimize L cGAN To counteract D, which tries to maximize the goal, it can be expressed as:

[0029]

[0030] Introducing multi-scale feature matching loss function L fm And the perceptual loss function L vgg Optimize the loss function of cGAN,

[0031] Among them, the multi-scale feature matching loss function L fm It is expressed as:

[0032]

[0033] Denotes discriminators D of different scales k The i-th layer feature extractor of represents the features extracted by the kth discriminator on the i-th layer with the original image and the true JRD image as input, The original image and fake JRD image are taken as input;

[0034] The VGG loss function is introduced as a perceptual loss, L vgg The calculation formula is as follows:

[0035]

[0036] i represents the i-th image among a total of n images, F represents the feature map of the generated fake JRD image, R represents the feature map of the true JRD image, C, H, and W represent the number of channels, height, and width of the feature map, respectively;

[0037] The ultimate goal of adversarial learning is as follows:

[0038]

[0039] The invention provides a precisely identifiable distortion prediction model and method suitable for machine vision tasks, which has the following advantages:

[0040] The invention of the present invention is to decompose the difficult task of predicting the JRD score into two relatively simple tasks, namely the generation task and the regression task. In the reference-free model proposed by the present invention, the original image does not need to be compressed and decoded several times, which greatly improves the prediction efficiency. At the same time, in the regression network, the residual of the generated image is input as supplementary information, making the final prediction result of the model more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is the main structural diagram of the prediction framework of the present invention.

[0042] Figure 2 This is a flow chart of the production of the JRD dataset taking the target detection task as an example of the present invention.

[0043] Figure 3 This is a schematic diagram of the adversarial training of the JRD image generation network G and the discriminant network D of the present invention.

[0044] Figure 4 It is a structural diagram of the multi-scale discriminator of the present invention. DETAILED DESCRIPTION

[0045] In order to better understand the purpose, structure and function of the present invention, the following further describes in detail a recognizable distortion prediction model and method suitable for machine vision tasks of the present invention in conjunction with the accompanying drawings.

[0046] like Figure 1 , Figure 3As shown, a recognizable distortion prediction model suitable for machine vision tasks of the present invention includes a JRD image generation network G, a discriminant network D and a residual-guided JRD regression network R. The JRD image generation network G takes the uncompressed original image as input and a fake JRD image as output. Afterwards, the residual of the generated image and the original image are input into the residual-guided JRD regression network R as compensation information for the original image, and finally the predicted JRD score is output, that is, the optimal quantization parameter for the image compression. The discriminant network D is trained together with the generation network G in an adversarial manner to help G produce a generation result that is closer to the true JRD image.

[0047] The model of the present invention uses the residual image between the original image and the corresponding pseudo JRD image as supplementary information. Figure 1 The input residual guides the JRD regression network, and finally generates a predicted JRD score, which is the optimal quantization parameter for image compression. The method proposed in the present invention guides the compression of images and videos, and can reduce data redundancy in transmission and storage as much as possible while ensuring high accuracy of machine vision tasks.

[0048] JRD image generation network G:

[0049] The generator part of the present invention is implemented using the basic structure of U-Net, and its network structure is as follows: Figure 1 As shown in the figure, U-Net is a fully convolutional neural network. Compared with the common encoder-decoder structure network that first downsamples to a low dimension and then upsamples to the original resolution, the difference of U-Net lies in its unique skip connection structure. The skip connection allows the feature layer in the encoding to be directly spliced ​​with the feature layer of the same dimension after decoding in the form of channel superposition to preserve pixel-level detail information at different scales.

[0050] A characteristic of image-to-image translation problems is that they map a high-resolution input grid to a high-resolution output grid. The input and output are not exactly the same in content, but are both rendered at different qualities based on the same underlying content. Therefore, the structure in the input is roughly aligned with the structure in the output. In the U-Net network, the input passes through a series of layers and is gradually downsampled until the bottleneck layer, after which transposed convolutions are performed to achieve upsampling. Such a network design allows all information to flow through all layers. For the image-to-image translation problem that the present invention needs to solve, there is a large amount of low-level information sharing between the input and the output, so it is effective to transmit this information directly through the jump connection structure.

[0051] Therefore, the structural design of U-Net enables it to combine the information of deep features and shallow features. Among them, shallow features can provide more refined feature information for image generation, which has an effect on improving the detail fineness of image generation. Deep features can provide contextual semantic information of image content in the entire image. Since the input and output images in the present invention correspond, these contextual information helps to improve the quality of generated images under the same underlying structure.

[0052] Discriminant network D:

[0053] The JRD values ​​corresponding to images with different semantics are different. In order to generate content that is closer to the quality of real JRD images, the generative network cannot do without the help of the discriminator. This also poses a major challenge to the design of the GAN discriminator. In order to better distinguish the generated images from the real images, the present invention uses a multi-scale discriminator with the following structure: Figure 4 As shown. The present invention uses two discriminators with the same structure, but makes them work at different image scales. The two discriminators working at different scales are called D1 and D2. D1 and D2 work at the original scale of the real image and the generated image and the scale of 2 times downsampling respectively. The D2 discriminator working at a coarse-grained scale has a larger receptive field. It can discriminate images from a more global perspective and guide the generator to generate more globally consistent images. The D1 discriminator working at a finer-grained scale can make judgments at a finer scale and help the generator generate finer details. Because this multi-scale discriminator structure does not use an overly deep network structure or an overly large convolution kernel, it takes up less computing resources and reduces the risk of overfitting.

[0054] Residual guided JRD regression network R:

[0055] The main function of the residual-guided JRD regression network is to evaluate the degree of distortion of the fake JRD image relative to the original image. It is a very difficult task to directly predict the JRD score of an image when only the original image is input. Therefore, the design strategy of the present invention is to introduce the network module and use the residual image between the generated JRD image and the original image as supplementary information. Its structure is as follows: Figure 1 shown.

[0056] Since the JRD score of an image is highly correlated with the semantic content of the image, the present invention selects to input the original image and the residual into the feature extractor respectively, and then performs linear regression calculation on their fusion features. The feature extractor uses VGG16, which has a parameter volume of up to 138M and a strong fitting ability. Its structure of increasing the network depth through convolution can effectively extract the semantic information of the image. The network parameters of VGG16 use the network parameters pre-trained on the ImageNet dataset. Therefore, in training, only the network parameters of the linear regression part need to be learned and optimized, which greatly reduces the difficulty of training.

[0057] In order to give priority to ensuring the accuracy of the image compressed by the prediction score φ(i) for the machine vision task, the present invention designs a loss function L for the network R , as shown in the following formula. During the training process, if the predicted JRD score φ(i) is greater than the true value Φ(i), MSE loss is used, otherwise L1 loss is used. The proposed loss function L R The predicted JRD scores Φ(i) that are greater than the actual JRD scores are more penalized, which ensures that under the guidance of the method of the present invention, the compressed video frames and images still have high machine vision task accuracy.

[0058]

[0059] N is the total number of images, φ(i) is the JRD prediction score of the i-th image, and Φ(i) is the true JRD of the image. The predicted value φ(i) is rounded down.

[0060] A method for accurately recognizing distortion prediction training applicable to machine vision tasks of the present invention comprises the following steps:

[0061] Step 1: Create a dataset containing uncompressed original images or video frames and their corresponding JRD images;

[0062] Because adversarial learning is often used to learn complex correspondences, in theory, as long as there is enough data, the GAN network can learn the mapping relationship between machine vision task accuracy and image distortion. Therefore, the present invention believes that the GAN network can learn the JRD scores corresponding to different images under specific machine vision tasks under the drive of a large amount of data. Therefore, the image generation effect of the generative network is very dependent on the data set. In the experiment of the present invention, the machine vision model M is defined as the YOLOv7 target detection algorithm. For the machine vision task of YOLOv7 target detection, the present invention creates a data set containing 130,000 uncompressed original images or video frames and their corresponding JRD images, and annotates their corresponding JRD values. The original images of the data set are derived from COCO2017. The present invention uses the YOLOv7 algorithm to infer the original images or video frames and intercept the detected object frames. The intercepted image I0 and the variants I1, I2, I3, ..., I after compression with different quantization parameters are as follows: n The YOLOv7 algorithm with the same parameters is used for inference, and finally the JRD image I corresponding to I0 is selected j The production process of the JRD dataset is as follows: Figure 2 shown.

[0063] Step 2: Generate adversarial network;

[0064] The adversarial learning framework of the present invention mainly uses the basic network structure of cGAN (conditional-GAN), such as Figure 3 As shown. cGAN is an improved network based on the Generative Adversarial Network (GAN). Unlike the GAN network that generates images directly from noise, cGAN achieves conditional and stable generation effects by giving the original GAN ​​generator (abbreviated as G) and discriminator (abbreviated as D) additional conditional information. The core operation of cGAN is to add conditional information to G and D, where the conditional information can be images, labels or other auxiliary information. In other words, cGAN can use the original image and the corresponding labeled image for training, so that the generation network can transform the original image into an image with specified content and specific changes.

[0065]

[0066] like Figure 3 As shown, in the present invention, x is the original image, y is the true JRD image, z is the noise, and G(x,z) is the output of G with x and z as input, that is, the generated false JRD image. In the above formula, G attempts to minimize L cGAN To counteract D, which tries to maximize the goal, it can be expressed as:

[0067]

[0068] In order to enhance the stability of network training, the present invention proposes to introduce a multi-scale feature matching loss function L fm And the perceptual loss function L vgg Optimize the loss function of cGAN. Specifically, the network will extract features from the multi-level discriminator, learn the correspondence between feature parameters in different levels, and achieve multi-level feature matching between the original image and the fake JRD image.

[0069] Among them, the multi-scale feature matching loss function L fm It is expressed as:

[0070]

[0071] Denotes discriminators D of different scales k The i-th layer feature extractor of represents the features extracted by the kth discriminator on the i-th layer with the original image and the true JRD image as input, The original image and the fake JRD image are taken as input.

[0072] In addition, the present invention introduces the VGG loss function as a perceptual loss to ensure that the quality of image generation in multiple dimensions is closer to the true JRD image. vgg The calculation formula is as follows:

[0073]

[0074] i represents the i-th image among a total of n images, F represents the feature map of the generated fake JRD image, and R represents the feature map of the true JRD image. C, H, and W represent the number of channels, height, and width of the feature map, respectively.

[0075] Therefore, the ultimate goal of adversarial learning is as shown in formula (6):

[0076]

[0077] The reference-free exactly recognizable distortion prediction method for machine vision tasks proposed in the present invention is suitable for guiding most image and video data compression tasks oriented to machine vision algorithms (such as image classification, target detection, etc.).

[0078] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A precisely identifiable distortion prediction model for machine vision tasks, It is characterized in that The invention comprises a JRD image generation network G, a discriminant network D and a residual guided JRD regression network R; the JRD image generation network G takes an uncompressed original image as input and a fake JRD image as output; then, the residual of the generated image and the original image is input into the residual guided JRD regression network R together as compensation information for the original image, and finally a predicted JRD score, i.e., the optimal quantization parameter for image compression, is output, wherein the discriminant network D is trained together with the generation network G in an adversarial manner to help G produce a generation result that is closer to a true JRD image.

2. The just recognizable distortion prediction model suitable for machine vision tasks according to claim 1, It is characterized in that The model uses the residual image between the original image and the corresponding pseudo JRD image as supplementary information, and inputs it into the residual guided JRD regression network together with the original image, and finally generates a predicted JRD score, that is, the optimal quantization parameter for image compression.

3. The just recognizable distortion prediction model suitable for machine vision tasks according to claim 1, It is characterized in that The generator part in the JRD image generation network G is implemented using the U-Net infrastructure. In the U-Net network, the input passes through a series of layers and is gradually down-sampled until the bottleneck layer, after which transposed convolution is performed to achieve up-sampling, and a large amount of low-level information between the input and output is transmitted through a skip connection structure.

4. The just recognizable distortion prediction model suitable for machine vision tasks according to claim 1, It is characterized in that The discriminant network D uses a multi-scale discriminator, using two discriminators with the same structure, but making them work at different image scales. The two discriminators working at different scales are called D 1 , D 2 , D 1 , D 2 We work on the original scale of real and generated images as well as on a 2x downsampled scale.

5. The just recognizable distortion prediction model suitable for machine vision tasks according to claim 1, It is characterized in that The main function of the residual-guided JRD regression network R is to evaluate the degree of distortion of the fake JRD image relative to the original image; Choose to input the original image and residual into the feature extractor respectively, and then perform linear regression calculation on their fused features. The feature extractor uses VGG16. The network parameters of VGG16 use the network parameters pre-trained on the ImageNet dataset. During training, only the network parameters of the linear regression part need to be learned and optimized.

6. The just recognizable distortion prediction model suitable for machine vision tasks according to claim 1, It is characterized in that The model includes the loss function L R During the training process, if the predicted JRD score φ(i) is greater than the true value Φ(i), MSE loss is used, otherwise L1 loss is used. N is the total number of images, φ(i) is the JRD prediction score of the i-th image, and Φ(i) is the true JRD of the image, where the predicted value φ(i) is rounded down.

7. A method for training a distortion prediction model suitable for machine vision tasks as claimed in any one of claims 1 to 6, It is characterized in that The steps include: Step 1: Create a dataset containing uncompressed original images or video frames and their corresponding JRD images; Step 2: Generate adversarial network.

8. The training method according to claim 7, It is characterized in that The step 1 comprises the following steps: For the machine vision task of YOLOv7 target detection, a dataset of 130,000 uncompressed original images or video frames and their corresponding JRD images is created, and their corresponding JRD values ​​are annotated. The original images of the dataset are from COCO2017. The original images or video frames are inferred by the YOLOv7 algorithm, and the detected object boxes are captured. The captured images I 0 And the variants after compression with different quantization parameters I 1 , I 2 , I 3 ,…,I n The YOLOv7 algorithm with the same parameters was used for reasoning, and finally I was selected 0 The corresponding JRD image I j .

9. The training method according to claim 7, It is characterized in that The step 2 comprises the following steps: The adversarial learning framework uses the basic network structure of cGAN. Where x is the original image, y is the true JRD image, z is the noise, G(x,z) is the output of G with x and z as input, that is, the generated fake JRD image. In the above formula, G tries to minimize L cGAN To counteract D, which tries to maximize the goal, it can be expressed as: Introducing multi-scale feature matching loss function L fm And the perceptual loss function L vgg Optimize the loss function of cGAN, Among them, the multi-scale feature matching loss function L fm It is expressed as: Denotes discriminators D of different scales k The i-th layer feature extractor of represents the features extracted by the kth discriminator on the i-th layer with the original image and the true JRD image as input, The original image and fake JRD image are taken as input; The VGG loss function is introduced as a perceptual loss, L vgg The calculation formula is as follows: i represents the i-th image among a total of n images, F represents the feature map of the generated fake JRD image, R represents the feature map of the true JRD image, C, H, and W represent the number of channels, height, and width of the feature map, respectively; The ultimate goal of adversarial learning is as follows:

Citation Information

Patent Citations

  • Image quality evaluation method and system, and computer readable medium

    CN114926459A

  • Target level just recognizable distortion generation method for machine vision coding

    CN116233429A

  • Generative residual error repairing method and device for low-bit-rate image compression

    CN116939226A

  • HVS-JND spectrum residual network robust watermarking method based on CNN

    CN116958690A

  • Encoder and decoder for encoding and decoding images

    US11546614B1