CN115661530A - Image data enhancement method and system and electronic equipment - Google Patents
Image data enhancement method and system and electronic equipment Download PDFInfo
- Publication number
- CN115661530A CN115661530A CN202211338374.5A CN202211338374A CN115661530A CN 115661530 A CN115661530 A CN 115661530A CN 202211338374 A CN202211338374 A CN 202211338374A CN 115661530 A CN115661530 A CN 115661530A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- disturbance
- resolution
- image data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of image enhancement, solves the technical problem of unstable confrontation of a high-resolution target image generation model, and particularly relates to an image data enhancement method, which comprises the following steps: preprocessing an acquired image data set, training a generation countermeasure network constructed based on a progressive incremental deep convolution neural network by adopting a procedural incremental training mode to obtain a trained generation model, inputting data to be enhanced and random Gaussian noise into the trained generation model for characteristic processing to obtain generated disturbance noise, and restoring the disturbance noise to obtain real noise to finish image data enhancement. According to the method, the generation countermeasure network constructed based on the progressive incremental deep convolutional neural network is trained by adopting the procedural incremental mode, so that a generation model capable of generating high-resolution disturbance is obtained, and the diversity and the generalization of the high-resolution image classification data set are improved.
Description
Technical Field
The present invention relates to the field of image enhancement technologies, and in particular, to a method, a system, and an electronic device for enhancing image data.
Background
Because the image is influenced by factors such as imaging equipment, imaging environment, light refraction and shooting personnel technical level, a large number of images have degradation phenomena such as low resolution, noise, distortion and the like, so that the useful information in the image needs to be enhanced by adopting an image enhancement technology, the image quality is improved, and the image is convenient to process in the later stage.
At present, a commonly used data enhancement method generally expands data volume of original data by performing random angle rotation on an actually acquired target image, and then enhances the data volume by using an enhancement model, although a good effect is obtained on a low-resolution image, the data enhancement method cannot solve the influence of factors such as light and lens vibration, and the model has poor stability when a high-resolution image is enhanced, and is prone to mode collapse, confusion and other problems.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides an image data enhancement method, an image data enhancement system and electronic equipment, solves the technical problems that the existing high-resolution target image generation model is unstable in countermeasure and is easy to cause mode collapse and confusion, and achieves the purposes that the high-resolution target image generation model is more stable and the image restoration is more real.
In order to solve the technical problems, the invention provides the following technical scheme: an image data enhancement method, comprising the processes of:
s1, acquiring a plurality of groups of image data sets acquired by image acquisition equipment;
s2, preprocessing each group of image data sets to obtain a disturbance data set, a standardized disturbance data set and label data;
s3, constructing a generation countermeasure network taking a progressive growth type deep convolutional neural network as a basic framework, wherein the generation countermeasure network comprises a generator and a discriminator;
s4, training the generated countermeasure network in a procedural incremental training mode according to the standardized disturbance data set to obtain a trained generated model;
and S5, performing characteristic processing on the data to be enhanced and random Gaussian noise according to the pre-trained VGG network and the generation model to obtain disturbance noise, performing standardized inverse transformation on the disturbance noise according to the mean value and the standard deviation in the step S2 to obtain real noise, and completing image data enhancement.
Further, the step S2 specifically includes:
s21, randomly selecting an image from each group of image data sets and zooming the image to a preset resolution as label data;
s22, scaling and rotating the original images in each group of image data sets to the same position, and then respectively carrying out difference on the original images under the same resolution to obtain a disturbance data set;
and S23, calculating the mean value and the standard deviation of the disturbance data set under the same resolution, and standardizing the mean value and the standard deviation to obtain the standardized disturbance data set.
Furthermore, the input of the generator is gaussian noise, the output is the normalized disturbance data, the generator and the discriminator both comprise a pre-trained VGG network, and the output layer of the generator adopts a pixel normalization layer.
Further, the input of the discriminator is the disturbance data after standardization, the output is an evaluation score, a small batch standardization layer is merged before the last layer, and a feature splicing layer for splicing the feature map obtained by the VGG network is merged before the last layer of the discriminator and after the first layer of the generator.
Further, the step S4 specifically includes:
s41, inputting a disturbance data set corresponding to an original image with the resolution of R multiplied by R and corresponding label data to start training a countermeasure network;
s42, after the R multiplied by R resolution training is finished, 2R multiplied by 2R resolution training is carried out, the result obtained by the R multiplied by R resolution network and the result obtained by the 2R multiplied by 2R resolution network are subjected to linear interpolation, the weight of the linear interpolation is set to be a, the initial value of a is equal to 0, and the number of times of iteration a in each training is increased by 0.001;
s43, judging whether the weight a of the linear interpolation is equal to 1, if a is less than 1, returning to execute the step S42, otherwise, executing the next step S44;
and S44, judging whether the resolution of the corresponding training image is equal to 896 when the weight a =1 of the linear interpolation, if so, finishing the training to obtain a trained generated model, otherwise, returning to the step S41.
Further, the step S5 specifically includes:
s51, inputting the data to be enhanced into a trained VGG network for feature processing to obtain a feature map of the data to be enhanced;
s52, splicing an intermediate result obtained by inputting random Gaussian noise into the generator through a first-layer network with the characteristic diagram, and taking the spliced diagram as the input of a network layer of a subsequent generator to obtain disturbance noise;
s53, reducing the generated disturbance noise according to the mean value and the standard deviation in the step S2 to obtain real noise, and finishing image data enhancement.
The invention also provides a technical scheme that: a system for implementing the image data enhancement method described above, comprising:
an image dataset acquisition module for acquiring sets of image datasets acquired by an image acquisition device;
the data preprocessing module is used for preprocessing each group of image data sets to obtain a disturbance data set, a standardized disturbance data set and label data;
the network construction module is used for constructing a generation countermeasure network which takes a progressive growing type deep convolutional neural network as a basic framework, and the generation countermeasure network comprises a generator and a discriminator;
the network training module is used for training the generated countermeasure network in a procedural incremental training mode according to the standardized disturbance data set to obtain a trained generated model;
and the real noise generation module is used for carrying out feature processing on the data to be enhanced and random Gaussian noise according to the pre-trained VGG network and the generation model to obtain disturbance noise, carrying out standard inverse transformation on the disturbance noise according to the mean value and the standard deviation in the step S2 to obtain real noise, and completing image data enhancement.
Furthermore, the input of the generator is gaussian noise, the output is a normalized disturbance data set, the output layer of the generator adopts a pixel normalization layer, and the generator and the discriminator both comprise a pre-trained VGG network, the input of the VGG network is label data, and the output of the VGG network is a feature map of the label data.
Further, the input of the discriminator is the normalized perturbation data set, the output is an evaluation score, a small batch normalization layer is merged before the third last layer, and a feature splicing layer for splicing the feature map obtained by the VGG network is merged before the second last layer of the discriminator and after the first layer of the generator.
The invention also provides a technical scheme that: an electronic device for implementing the image data enhancement method is characterized by comprising: a processor, a memory, and a program or instructions stored on the memory and executable on the processor, which when executed by the processor, implements the steps of the method of any one of claims 1 to 6.
By means of the technical scheme, the invention provides an image data enhancement method, an image data enhancement system and electronic equipment, which at least have the following beneficial effects:
1. the method comprises the steps of preprocessing an acquired image data set, using a gradually-growing deep convolutional neural network as a basic frame for generating a countermeasure network, inserting a feature splicing layer before the penultimate layer of a discriminator and after the first layer of the generator for splicing a feature map obtained by a VGG network, adding a small batch of standard difference layers before the penultimate layer of the discriminator, using a pixel normalization layer as an output layer of the generator, carrying out forward reasoning on all convolutional layers in a mode of balanced learning rate, training in a procedural incremental training mode to generate a countermeasure network model, and using a Wtherstein factor distance as a loss function to obtain a generation model capable of generating high-resolution disturbance, so that the problems of difficult diamond classification, unstable countermeasure of the classification model, mode collapse and mode confusion of the existing high-resolution diamond are solved.
2. According to the method, the characteristic diagram of the data to be enhanced is obtained by performing characteristic processing on the VGG network to be enhanced and input to be pre-trained, then the intermediate result obtained by inputting random Gaussian noise into a generator through a first layer of network is spliced with the characteristic diagram to obtain a spliced diagram, the spliced diagram is used as the input of a subsequent generator network layer to obtain generated disturbance noise, then the generated disturbance noise is subjected to standard inverse transformation according to the mean value and the standard deviation to obtain real diamond noise, the enhancement of high-resolution image data is completed, the image restoration is more real, and the diversity and the generalization of a data set of high-resolution image classification are improved.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application. In the drawings:
FIG. 1 is a flow chart of a method for enhancing image data according to the present invention;
FIG. 2 is a flow chart of data preprocessing in the image data enhancement method provided by the present invention;
FIG. 3 is a schematic diagram of data preprocessing in the image data enhancement method provided by the present invention;
FIG. 4 is a frame diagram of a countermeasure network generated in the image data enhancement method provided by the present invention;
FIG. 5 is a schematic diagram of generation of confrontation network training in the image data enhancement method provided by the present invention;
FIG. 6 is a flowchart of generation of confrontational network training in the image data enhancement method provided by the present invention;
FIG. 7 is a schematic representation of cross-resolution training in the image data enhancement method provided by the present invention;
FIG. 8 is a flow chart illustrating the generation of true noise in the image data enhancement method according to the present invention;
FIG. 9 is a schematic diagram illustrating the generation of real noise in the image data enhancement method according to the present invention;
FIG. 10 is a schematic diagram of an image data enhancement system provided by the present invention;
fig. 11 is a block diagram of an image data enhancement electronic device according to the present invention.
In the figure: 10. an image dataset acquisition module; 20. a data preprocessing module; 30. a network construction module; 40. a network training module; 50. and a real noise generation module.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below. Therefore, the realization process of how to apply technical means to solve the technical problems and achieve the technical effects can be fully understood and implemented.
It will be understood by those skilled in the art that all or part of the steps in the method for implementing the above embodiments may be implemented by relevant hardware instructed by a program, and therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
Overview of scenes
Because the diamond image is influenced by factors such as shooting environment, light refraction, diamond cutting and the like, the classification of the high-resolution diamonds is extremely difficult, the diamond data needs to be expanded in order to ensure the stable confrontation of a classification model, but the common rotation expansion cannot solve the diamond data under the influence of factors such as light, lens vibration and the like, and the artificial shooting cost is extremely high, therefore, the method for generating the confrontation network is adopted to process the diamond image, new diamond disturbance is generated, the newly generated disturbance is added to the diamond image, the new diamond data is obtained, the task of data enhancement is completed, and the purpose of quickly and accurately classifying the high-resolution diamond image is achieved.
Examples
Referring to fig. 1 to 8, a specific implementation of the present embodiment is shown, in the present embodiment, a generative countermeasure network constructed based on a convolutional neural network is trained by using a procedural incremental training method, and a wotherstein distance is used as a loss function to obtain a generative model capable of generating high-resolution disturbance, so that the problems of difficulty in diamond classification, unstable countermeasure of the classification model, and easiness in mode collapse and confusion of the existing high-resolution diamond are solved, and the purpose of rapidly and accurately classifying high-resolution images is achieved.
As shown in fig. 1, an image data enhancement method includes the steps of:
s1, a plurality of groups of image data sets acquired by image acquisition equipment are acquired.
Sequentially acquiring original images of a plurality of diamonds at the same resolution through an image acquisition device, and rotating the original images of the same diamond at different angles to obtain a group of original image sets of the same diamond at different rotation visual angles, wherein the set of original image sets is marked as X, namely X = { X = { (X) } 1 ，x 2 ，…，x n And n denotes the number of images.
And S2, preprocessing each group of image data sets to obtain a disturbance data set, a standardized disturbance data set and label data. As shown in fig. 2 and 3, the specific steps include:
and S21, randomly selecting an image from each group of image data set and zooming to a preset resolution as label data.
Randomly selecting an original image in each data set, scaling to a preset resolution of 112 × 112 to obtain label data, and marking as x * 。
S22, scaling and rotating the original images in each group of image data sets to the same position, and then respectively performing difference on the original images under the same resolution ratio to obtain a disturbance data set.
First, the image dataset X is scaled to eight resolutions, 7 × 7, 14 × 14, 28 × 28, 56 × 56, 112 × 112, 224 × 224, 448 × 448, 896 × 896, resulting in eight datasets, denoted as X h I.e. by
h∈{7，14，28，56，112，224，448，896}。
Then, sequentially adopting a mode of respectively making differences on the images under the same resolution ratio to obtain a disturbance data set marked as E h I.e. by
And is provided with
And S23, calculating the mean value and the standard deviation of the disturbance data set under the same resolution, and standardizing the mean value and the standard deviation to obtain the standardized disturbance data set.
The formula for calculating the mean and standard deviation of the perturbed data set at the same resolution is as follows:
in the above equation, N represents the number of disturbance data. The mean and standard deviation are then normalized, and the normalized calculation formula is as follows:
standardizing the mean value and standard deviation of the disturbance data set under each resolution according to the formula to obtain a standardized disturbance data set, and recording the standardized disturbance data set as
Namely that
And S3, constructing a generation countermeasure network taking the progressive growth type deep convolutional neural network as a basic framework. As shown in fig. 4, the generation of the countermeasure network includes two parts, a generator and a discriminator.
The input of the generator is Gaussian noise, the output is a standardized disturbance data set, the output layer of the generator adopts a pixel normalization layer, the generator and the discriminator both comprise a pre-trained VGG network, and the input of the VGG network is label data x * And the characteristic diagram with output being label data is marked as vgg (x) * ) (ii) a The input of the discriminator is the disturbance data after standardization, the output is an evaluation score, a small batch standardization layer is merged before the last but one layer, and a feature splicing layer for splicing the feature map obtained by the VGG network is merged before the last but one layer of the discriminator and after the first layer of the generator.
The output layer of the generator adopts a pixel normalization layer, all the convolution layers adopt a mode of balanced learning rate to carry out forward reasoning, and a method of balanced learning rate is adopted to ensure the stability of network output, the training of different resolution stages has different corresponding learning rates and different training iteration times, and the batch size of each stage is different according to the bearing capacity of a computer. The forward reasoning calculation formula of the small batch standard deviation calculation layer is as follows:
y＝concat(conv_layer(x),vgg(x))
MiniBatchStd(x)＝concat(x,mean(std(x,dim＝0)))×ones_like(x)
in the above equation, miniBatchStd represents the forward inference calculation formula of the small batch standard deviation calculation layer, concat represents tensor concatenation, mean function represents mean function, std (x, dim = 0) represents the calculation standard deviation along the input batch dimension, ones _ like (x) represents the full 1 tensor as large as x.
The forward reasoning calculation formula of the pixel normalization layer is as follows:
in the above formula, pixnrormlayer represents the forward inference calculation formula of the pixel normalization layer, mean (x) 2 Dim = 1) represents averaging along the input channel dimension.
In the above equation, equizedlrlayer represents the forward inference calculation formula of each layer of the convolutional network, conv _ layer represents the convolutional layer, and n _ dim (x) represents the dimension of the input.
And S4, training the generation countermeasure network by adopting a procedural incremental training mode according to the standardized disturbance data set to obtain a trained generation model.
In order to generate high-resolution noise, the generation countermeasure network is trained in a procedural incremental training manner, and as shown in fig. 5, the training of the low-resolution generation model is gradually transited to the training of the high-resolution generation model.
As shown in fig. 6, the specific steps of training the generation countermeasure network include:
and S41, inputting a disturbance data set corresponding to the original image with the resolution of R multiplied by R and corresponding label data to start generating the countermeasure network.
Specifically, training is started from R × R resolution, and wotherstein distance (Wasserstein distance) is used as a loss function. In this embodiment, a perturbation data set corresponding to an original image with a resolution of 7 × 7 and corresponding label data are input to start training to generate a countermeasure network. Wherein, the formula adopted by the loss function is as follows:
in the above formula, D represents a discriminator, G represents a generator, X represents a true disturbance, Z represents Gaussian noise, and P represents r Score representing true disturbanceThe cloth is made of a material having a high thermal conductivity,
indicating a gaussian distribution.
And S42, after the R multiplied by R resolution training is finished, 2R multiplied by 2R resolution training is carried out, linear interpolation is carried out on the result obtained by the R multiplied by R resolution network and the result obtained by the 2R multiplied by 2R resolution network, the weight of the linear interpolation is set to be a, the initial value of the a is equal to 0, and the a is increased by 0.001 every training iteration time.
As shown in fig. 7, after the R × R resolution training is completed, 2R × 2R resolution training is performed, in the training process, in order to ensure the stability of the network cross-resolution training, the generator performs upsampling on the Output result of the R × R resolution training to obtain an Output result Output _ R of the 2R × 2R resolution training, and performs interpolation with the Output result Output _2R directly passing through the 2R × 2R resolution, that is:
(1-a)*Output_R+a*Output_2R
in the above equation, a represents the weight of linear interpolation, the initial value is 0, and a is increased by 0.001 every time one iteration is trained.
And S43, judging whether the weight a of the linear interpolation is equal to 1, if a is less than 1, returning to the step S42, and if not, executing the next step S44.
And S44, judging whether the resolution of the corresponding training image is equal to 896 when the weight a =1 of the linear interpolation, if so, finishing the training to obtain a trained generated model, otherwise, returning to the step S41.
In this embodiment, a model capable of generating high-resolution disturbance is obtained by training the model in a procedural incremental training manner and using the wotherstein distance as a loss function, so that the problems of difficulty in classifying the existing high-resolution diamond, instability in resisting the classification model, mode collapse and mode confusion are solved.
And S5, performing characteristic processing on the data to be enhanced and random Gaussian noise according to the pre-trained VGG network and the generation model to obtain disturbance noise, performing standard inverse transformation on the disturbance noise according to the mean value and the standard deviation in the step S2 to obtain real noise, and completing image data enhancement. As shown in fig. 8 and 9, the method specifically includes the following steps:
and S51, inputting the data to be enhanced into a pre-trained VGG network for feature processing to obtain a feature map of the data to be enhanced.
Specifically, the designated diamond image is scaled to 112 × 112 resolution as the tag data x * And inputting the pre-trained VGG network for feature processing to obtain a feature map of the data to be enhanced. It should be noted that the VGG network is a convolutional network framework which is currently relatively standard and common, and is pre-trained.
And S52, splicing the intermediate result obtained by inputting the random Gaussian noise into the generator through the first-layer network with the characteristic diagram, and taking the spliced diagram as the input of the network layer of the subsequent generator to obtain the disturbance noise.
Specifically, the generator inputs random gaussian noise, splices an intermediate result obtained by passing the gaussian noise through a first-layer network with the feature map of the tag data, and uses the spliced result as the input of a subsequent generator network layer to obtain the generated disturbance noise.
And S53, carrying out standard inverse transformation on the generated disturbance noise according to the mean value and the standard deviation in the step S2 to obtain real noise, and finishing image data enhancement.
Specifically, the generated disturbance noise is subjected to normalization inverse transformation according to the mean value μ and the standard deviation σ calculated in step S2, so as to obtain real diamond noise, and thus the goal of image data enhancement is completed.
According to the embodiment, firstly, an acquired image data set is preprocessed, then a progressive-growth type deep convolutional neural network is used as a basic frame for generating an antagonistic network, a course incremental training mode is adopted for training to generate an antagonistic network model, wosestein factor distance is used as a loss function to obtain a generation model capable of generating high-resolution disturbance, finally, the VGG network to be enhanced is input into the pre-training for feature processing to obtain a feature map of data to be enhanced, then a random Gaussian noise input generator is spliced with an intermediate result obtained through a first layer network and the feature map and used as input of a subsequent generator network layer to obtain generated disturbance noise, the generated disturbance noise is reduced according to a mean value and a standard difference to obtain real diamond noise, the enhancement of high-resolution image data is completed, the image reduction is more real, and the diversity and the generalization of the data set of high-resolution image classification are improved.
Referring to fig. 10, the present embodiment further provides a system for implementing the image data enhancement method, including:
an image data set acquiring module 10, configured to acquire image data sets composed of sets of diamond images with different rotation angles acquired by an image acquiring device.
And the data preprocessing module 20 is configured to preprocess each group of image data sets to obtain a disturbance data set, a standardized disturbance data set, and label data.
The network construction module 30 is used for constructing a generation countermeasure network based on a progressive-growth type deep convolutional neural network, and the generation countermeasure network comprises a generator and a discriminator;
the input of the generator is Gaussian noise, the output of the generator is a standardized disturbance data set, an output layer of the generator adopts a pixel normalization layer, the generator and the discriminator both comprise a pre-trained VGG network, the input of the VGG network is label data, and the output of the VGG network is a feature map of the label data;
the input of the discriminator is a normalized disturbance data set, the output is an evaluation score, a small batch normalization layer is merged before the third last layer, and a feature splicing layer for splicing the feature map of the convolutional neural network is merged before the second last layer of the discriminator and after the first layer of the generator.
All convolutional layers of the generator and the discriminator adopt a mode of equalizing learning rate to carry out forward reasoning.
And the network training module 40 is used for training the generation countermeasure network by adopting a course incremental training mode according to the standardized disturbance data set to obtain a trained generation model.
And a real noise generation module 50, configured to perform feature processing on the data to be enhanced and the random gaussian noise according to the pre-trained VGG network and the generation model to obtain disturbance noise, and perform standardized inverse transformation on the disturbance noise according to the mean value and the standard deviation in step S2 to obtain real noise, so as to complete image data enhancement.
Referring to fig. 11, the present embodiment further provides a technical solution: an electronic device for implementing the image data enhancement method includes:
a processor, a memory, and a program or instructions stored on the memory and executable on the processor, the program or instructions when executed by the processor implementing the steps of the method as claimed in any one of claims 1 to 6 including.
According to the embodiment, a progressive-growth type deep convolutional neural network is used as a basic framework for generating the countermeasure network, a feature splicing layer is inserted before the penultimate layer of the discriminator and after the first layer of the generator and is used for splicing the feature map, a small-batch standard deviation layer is added before the penultimate layer of the discriminator, a pixel normalization layer is adopted as an output layer of the generator, forward reasoning is carried out on all convolutional layers in a mode of balanced learning rate, a countermeasure network model is generated by training in a procedural incremental training mode, and a generation model capable of generating high-resolution disturbance is obtained by taking the Wosestein factor distance as a loss function.
The present invention has been described in detail with reference to the foregoing embodiments, and the principles and embodiments of the present invention have been described herein with reference to specific examples, which are provided only to assist understanding of the methods and core concepts of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present invention.
Claims (10)
1. An image data enhancement method, characterized by comprising the processes of:
s1, acquiring a plurality of groups of image data sets acquired by image acquisition equipment;
s2, preprocessing each group of image data sets to obtain a disturbance data set, a standardized disturbance data set and label data;
s3, constructing a generation countermeasure network taking a progressive growth type deep convolutional neural network as a basic framework, wherein the generation countermeasure network comprises a generator and a discriminator;
s4, training the generated countermeasure network in a procedural incremental training mode according to the standardized disturbance data set to obtain a trained generated model;
and S5, performing characteristic processing on the data to be enhanced and random Gaussian noise according to the pre-trained VGG network and the generation model to obtain disturbance noise, performing standard inverse transformation on the disturbance noise according to the mean value and the standard deviation in the step S2 to obtain real noise, and completing image data enhancement.
2. The method according to claim 1, wherein the step S2 specifically comprises:
s21, randomly selecting an image from each group of image data set and zooming the image to a preset resolution as label data;
s22, scaling and rotating the original images in each group of image data sets to the same position, and then respectively carrying out difference on the original images under the same resolution to obtain a disturbance data set;
and S23, calculating the mean value and the standard deviation of the disturbance data set under the same resolution, and standardizing the mean value and the standard deviation to obtain the standardized disturbance data set.
3. The method of claim 1, wherein the input of the generator is gaussian noise, the output is normalized perturbation data, and both the generator and the discriminator comprise a pre-trained VGG network, and the output layer of the generator employs a pixel normalization layer.
4. The image data enhancement method of claim 1, wherein the input of the discriminator is the normalized perturbation data, the output is an evaluation score, and a small lot normalization layer is merged before the third last layer, and a feature merging layer for merging the feature maps obtained by the VGG network is merged before the second last layer of the discriminator and after the first layer of the generator.
5. The method according to claim 1, wherein the step S4 specifically comprises:
s41, inputting a disturbance data set corresponding to an original image with the resolution of R multiplied by R and corresponding label data to start training a countermeasure network;
s42, after the R multiplied by R resolution training is finished, 2R multiplied by 2R resolution training is carried out, the result obtained by the R multiplied by R resolution network and the result obtained by the 2R multiplied by 2R resolution network are subjected to linear interpolation, the weight of the linear interpolation is set to be a, the initial value of a is equal to 0, and the number of times of iteration a in each training is increased by 0.001;
s43, judging whether the weight a of the linear interpolation is equal to 1, if a is less than 1, returning to execute the step S42, otherwise, executing the next step S44;
and S44, judging whether the resolution of the corresponding training image is equal to 896 when the weight a =1 of the linear interpolation, if so, ending the training to obtain a trained generated model, otherwise, returning to the step S41.
6. The image data enhancement method according to claim 1, wherein the step S5 specifically comprises:
s51, inputting the data to be enhanced into a pre-trained VGG network for feature processing to obtain a feature map of the data to be enhanced;
s52, splicing an intermediate result obtained by inputting random Gaussian noise into the generator through a first-layer network with the characteristic diagram, and taking the spliced diagram as the input of a network layer of a subsequent generator to obtain disturbance noise;
and S53, carrying out standard inverse transformation on the generated disturbance noise according to the mean value and the standard deviation in the step S2 to obtain real noise, and finishing image data enhancement.
7. A system for implementing the image data enhancement method of any one of claims 1 to 6, comprising:
an image dataset acquisition module (10), the image dataset acquisition module (10) being for acquiring sets of image datasets acquired by an image acquisition device;
the data preprocessing module (20) is used for preprocessing each group of image data sets to obtain a disturbance data set, a standardized disturbance data set and label data;
the network construction module (30) is used for constructing a generation countermeasure network based on a progressive growing type deep convolutional neural network, and the generation countermeasure network comprises a generator and a discriminator;
the network training module (40) is used for training the generated countermeasure network in a process incremental training mode according to the standardized disturbance data set to obtain a trained generation model;
and the real noise generation module (50) is used for performing characteristic processing on the data to be enhanced and the random Gaussian noise according to the pre-trained VGG network and the generation model to obtain disturbance noise, performing standard inverse transformation on the disturbance noise according to the mean value and the standard deviation in the step S2 to obtain real noise, and completing image data enhancement.
8. The image generation system of claim 7, wherein the input of the generator is gaussian noise, the output is a normalized perturbation data set, the output layer of the generator employs a pixel normalization layer, and the generator and the discriminator both include a pre-trained VGG network, the input of the VGG network is label data, and the output is a feature map of the label data.
9. The image generation system of claim 7, wherein the input of the discriminator is the normalized perturbation data set, the output is an evaluation score, and a small batch normalization layer is merged before the third last layer, and a feature merging layer for merging the feature map obtained by the VGG network is merged before the second last layer of the discriminator and after the first layer of the generator.
10. An electronic device for implementing the image data enhancement method according to any one of claims 1 to 6, comprising: a processor, a memory, and a program or instructions stored on the memory and executable on the processor, which when executed by the processor, implements the steps of the method of any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211338374.5A CN115661530A (en) | 2022-10-28 | 2022-10-28 | Image data enhancement method and system and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211338374.5A CN115661530A (en) | 2022-10-28 | 2022-10-28 | Image data enhancement method and system and electronic equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115661530A true CN115661530A (en) | 2023-01-31 |
Family
ID=84992776
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211338374.5A Pending CN115661530A (en) | 2022-10-28 | 2022-10-28 | Image data enhancement method and system and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115661530A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117808685A (en) * | 2024-02-29 | 2024-04-02 | 广东琴智科技研究院有限公司 | Method and device for enhancing infrared image data |
| CN117975201A (en) * | 2024-03-29 | 2024-05-03 | 苏州元脑智能科技有限公司 | Training data generation method, device, computer equipment and storage medium |
-
2022
- 2022-10-28 CN CN202211338374.5A patent/CN115661530A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117808685A (en) * | 2024-02-29 | 2024-04-02 | 广东琴智科技研究院有限公司 | Method and device for enhancing infrared image data |
| CN117808685B (en) * | 2024-02-29 | 2024-05-07 | 广东琴智科技研究院有限公司 | Method and device for enhancing infrared image data |
| CN117975201A (en) * | 2024-03-29 | 2024-05-03 | 苏州元脑智能科技有限公司 | Training data generation method, device, computer equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113240580B (en) | Lightweight image super-resolution reconstruction method based on multi-dimensional knowledge distillation | |
| Quan et al. | Image inpainting with local and global refinement | |
| CN115661530A (en) | Image data enhancement method and system and electronic equipment | |
| Yu et al. | A unified learning framework for single image super-resolution | |
| CN109727195B (en) | Image super-resolution reconstruction method | |
| CN109636721B (en) | Video super-resolution method based on countermeasure learning and attention mechanism | |
| CN111626932A (en) | Super-resolution reconstruction method and device for image | |
| US20220414838A1 (en) | Image dehazing method and system based on cyclegan | |
| CN113592715B (en) | Super-resolution image reconstruction method for small sample image set | |
| CN115345866B (en) | Building extraction method in remote sensing image, electronic equipment and storage medium | |
| CN112184582B (en) | Attention mechanism-based image completion method and device | |
| KR20220055970A (en) | Image Upscaling Apparatus And Method Based On Learning With Privileged Information | |
| CN116563916A (en) | Attention fusion-based cyclic face super-resolution method and system | |
| CN111753670A (en) | Human face overdividing method based on iterative cooperation of attention restoration and key point detection | |
| Hesabi et al. | Structure and texture image inpainting | |
| Shi et al. | Space targets adaptive optics images blind restoration by convolutional neural network | |
| CN113538254A (en) | Image restoration method and device, electronic equipment and computer readable storage medium | |
| CN116091893A (en) | Method and system for deconvolution of seismic image based on U-net network | |
| CN115239967A (en) | Image generation method and device for generating countermeasure network based on Trans-CSN | |
| CN113888424A (en) | Historical relic photo color restoration method and device, electronic equipment and storage medium | |
| CN114549314A (en) | Method for improving image resolution | |
| CN113628139A (en) | Fuzzy image restoration method and system based on generation countermeasure network | |
| Viriyavisuthisakul et al. | Parametric loss-based super-resolution for scene text recognition | |
| CN112016456A (en) | Video super-resolution method and system based on adaptive back projection depth learning | |
| CN112287938A (en) | Text segmentation method, system, device and medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |