A high-quality image classification method and system based on frequency band information

Abstract

A high-quality image classification method and system based on frequency band information relates to the field of intelligent image evaluation technology, addressing the existing need for an accurate high-quality image classification method and system. The method comprises the following steps: obtaining a training set, assigning labels to the images in the training set, wherein the labels are divided into two categories: whether they are damaged and whether they are oversharpened and / or have artifacts; obtaining frequency band information of the images using Gaussian high-pass filtering; obtaining image features using the frequency band information; training an image structure discrimination model and an oversharpening artifact discrimination model using the image features of the images in the training set; and obtaining high-quality images from the image set based on the image structure discrimination model and the oversharpening artifact discrimination model. The present invention accurately discriminates image quality, greatly reduces the burden on staff, improves the efficiency of image evaluation, solves the problem of poor-quality images interfering with image evaluation work, and fills a gap in image quality classification.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent image evaluation, and in particular to a high-quality image classification method and system based on frequency band information. Background Art

[0002] With the rapid development of image processing, effective evaluation of processed images and quantitative assessment of algorithm performance have become hot research topics. Given the lack of available references in practical engineering applications, existing research has largely focused on evaluation without reference images. However, when image processing is the premise, not all restoration algorithms can accurately process all image types. The acquired images may exhibit structural distortion, oversharpening, and significant artifacts. These low-quality images, due to the excessive artifact details, severely interfere with image evaluation, resulting in objective metrics that clearly differ from subjective visual perception. Therefore, an accurate, high-quality image classification method and system are needed. Summary of the Invention

[0003] In order to solve the above problems, the present invention provides a method and system for high-quality image classification based on frequency band information, which adopts a method for classifying images to conveniently eliminate low-quality images and obtain high-quality images.

[0004] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0005] A high-quality image classification method based on frequency band information, comprising:

[0006] A training set is obtained, and labels are set for images in the training set. The labels are divided into two categories: whether the images are damaged, and whether there are oversharpening and / or artifacts.

[0007] Use Gaussian high-pass filtering to obtain the frequency band information of the image;

[0008] Use frequency band information to obtain image features;

[0009] The image features of the images in the training set are used to train an image structure discrimination model and an over-sharpening artifact discrimination model; high-quality images in the image set can be obtained based on the trained image structure discrimination model and over-sharpening artifact discrimination model.

[0010] A high-quality image classification system based on frequency band information, comprising:

[0011] A first acquisition module is configured to obtain a training set, wherein images in the training set are provided with labels, and the labels are divided into two categories: whether the images are damaged, and whether there are oversharpening and / or artifacts;

[0012] A first calculation module is used to obtain frequency band information of the image using Gaussian high-pass filtering;

[0013] A second calculation module is used to obtain image features using frequency band information;

[0014] The training module is used to train an image structure discrimination model and an over-sharpening artifact discrimination model using the image features of the images in the training set; high-quality images in the image set can be obtained based on the trained image structure discrimination model and over-sharpening artifact discrimination model.

[0015] A method for obtaining high-quality images comprises the following steps:

[0016] S1. Obtain the image to be judged;

[0017] S2. Use Gaussian high-pass filtering to obtain frequency band information of the image to be judged;

[0018] S3. Obtaining image features of the image to be determined based on frequency band information of the image to be determined;

[0019] S4, using the image features obtained in S3 as input to the image structure discrimination model;

[0020] S5. Inputting the image features corresponding to the image structure not being destroyed output by the image structure discrimination model in S4 into the over-sharpening artifact discrimination model, and inputting the image features corresponding to the image without over-sharpening and / or artifacts output by the over-sharpening artifact discrimination model;

[0021] S6. Obtain a high-quality image based on the image features corresponding to the absence of over-sharpening and / or artifacts obtained in S5;

[0022] The image structure discrimination model and the over-sharpening artifact discrimination model are both obtained by using the high-quality image classification method based on frequency band information.

[0023] The beneficial effects of the present invention are:

[0024] The present invention proposes a high-quality image classification method, system, and method for obtaining high-quality images based on frequency band information. These methods construct and train models for structural damage and over-sharpening artifacts. The accuracy of the present invention for distinguishing high-quality images exceeds 90%, and the results are generally consistent with human subjective visual perception. This invention significantly reduces the burden on image quality assessment personnel, improves the efficiency of image evaluation, addresses the problem of poor-quality images interfering with image evaluation, and fills a gap in image quality classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The figure is a flow chart of a high-quality image classification method based on frequency band information of the present invention.

[0026] Figure 2This is a schematic diagram showing that each image feature has relatively obvious differences under different circumstances in a high-quality image classification method based on frequency band information of the present invention.

[0027] Figure 3 The figure is a flow chart of a method for obtaining high-quality images according to the present invention. DETAILED DESCRIPTION

[0028] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0030] A high-quality image classification method based on frequency band information includes the following steps:

[0031] Obtaining a training set consisting of images, and setting labels for the images in the training set, wherein the labels include whether the images are oversharpened and / or have artifacts, and whether the structures are destroyed;

[0032] Use Gaussian high-pass filtering to obtain the frequency band information of the images in the training set;

[0033] Use frequency band information to obtain image features;

[0034] The training set is used to train the image structure discrimination model and the over-sharpening artifact discrimination model.

[0035] The trained image structure discrimination model and over-sharpening artifact discrimination model can be used to obtain high-quality images in the image set.

[0036] As a preference, the label further comprises a high-quality image.

[0037] Specifically, a test set is also required. Both the training set and the test set include high-quality images and low-quality images. Labels are set for the images in the training set and the test set. The labels include high-quality images, whether the structure is destroyed, and whether there is over-sharpening and / or artifacts. In this embodiment, "whether there is over-sharpening and / or artifacts" is a class label. When there is over-sharpening and no artifacts, when there is over-sharpening and artifacts, and when there is no over-sharpening but artifacts, the label is marked as "yes". After training the image structure discrimination model and the over-sharpening artifact discrimination model, the test set is used to test the accuracy of the image structure discrimination model and the over-sharpening artifact discrimination model.

[0038] The training set can also be divided into a first training set and a second training set. The first training set has labels indicating whether the images have structural damage, used to train the image structure discrimination model; the second training set has labels indicating whether the images have oversharpening and / or artifacts, used to train the oversharpening artifact discrimination model. Similarly, the test set can be divided into a first test set and a second test set, which will not be detailed here.

[0039] The image structure discrimination model is used to determine whether the image structure is destroyed, and the oversharpening artifact discrimination model is used to determine whether the image has oversharpening and / or artifacts. In this embodiment, an image with intact structure, no oversharpening, and no artifacts is considered a high-quality image. Of course, a high-quality image is not limited to the two requirements of whether there is oversharpening and / or artifacts, and whether the image structure is destroyed.

[0040] The following is a detailed description of each step. Figure 1 shown.

[0041] Step 1: Based on the existing images, divide them into two parts: training set and test set, and set the labels corresponding to the images. The training set and test set are collectively referred to as the sample image set.

[0042] Step 2: For all images in the training set and test set, use Gaussian high-pass filtering to obtain the frequency band information of the image.

[0043] The transfer function of the Gaussian high-pass filter is:

[0044]

[0045] Among them, D0 is the cutoff frequency; D(u,v) is the distance from the point (u,v) in the frequency domain to the frequency origin, that is, u represents the abscissa of a point in the frequency domain, and v represents the ordinate of a point in the frequency domain.

[0046] Step 3: Use frequency band information to obtain image features of all images in the training set and test set;

[0047] Among them, the frequency band information corresponds to high-frequency components and low-frequency components, and the image features include E1, E2, E3, AP and AG. Figure 2 As shown in the figure, by observing the five images, it can be clearly seen that the features of each image have obvious differences in different situations, which is suitable for solving classification problems. Figure 2 (a) corresponds to E1, Figure 2 (b) corresponds to E2, Figure 2 (c) corresponds to E3, Figure 2 (d) corresponds to AP, Figure 2(e) Corresponding to AE, break indicates an image with a destroyed structure, OverSharp indicates an image with over-sharpened artifacts, and Highquality indicates a high-quality image.

[0048] The image features obtained using frequency band information are expressed as:

[0049]

[0050] Among them, E1, E2 and E3 are the image information entropy, high-frequency component information entropy and medium- and low-frequency component information entropy respectively; AP is the mean of the high-frequency component; AG is the gradient mean of the high-frequency component; I is the input image; I_high is the high-frequency component corresponding to I (when obtaining the high-frequency component of the image, the filter radius of the Gaussian high-pass filter is set to 100), I-I_high is the medium- and low-frequency component corresponding to I; I_high_g is the gradient of the high-frequency component; m and n are the image sizes, m is the width of the image, and n is the height of the image; i is the pixel number; entropy is the one-dimensional entropy of the image.

[0051] The one-dimensional entropy of an image is expressed as:

[0052]

[0053] Among them, i is a pixel value of the image, p i is the probability that i appears in the image.

[0054] Step 4: Use a decision tree, specifically a binary tree, to decompose the optimization objective into two binary classification problems. Use the image features of the images in the training set and the labels corresponding to the image features to train the image structure discrimination model and the over-sharpening artifact discrimination model of the support vector machine (SVM).

[0055] The support vector machine kernel function described in step 4 selects the Gaussian radial basis kernel function (RBF):

[0056]

[0057] Among them, exp is the e exponent; σ is the width parameter of the function, which controls the radial range of the function, x i and x j are the feature vectors of the two samples in high-dimensional space.

[0058] Use the test set to test the accuracy of the two models and debug the relevant parameters.

[0059] Step 5: Integrate the obtained discriminant model, which can achieve the function of high-quality image classification. Specifically, the image structure discriminant model and the over-sharpening artifact discriminant model can be integrated into a model, called an image classification model.

[0060] A high-quality image classification system based on frequency band information, comprising:

[0061] A first acquisition module is configured to obtain a training set, wherein the images in the training set are provided with labels, wherein the labels include whether the structure is destroyed and whether there is oversharpening and / or artifacts;

[0062] A first calculation module is used to obtain frequency band information of the image using Gaussian high-pass filtering;

[0063] A second calculation module is used to obtain image features using frequency band information;

[0064] The training module is used to train an image structure discrimination model and an over-sharpening artifact discrimination model using the image features of the images in the training set; high-quality images in the image set can be obtained based on the trained image structure discrimination model and over-sharpening artifact discrimination model.

[0065] A method for obtaining a high-quality image of the model obtained according to the above method, such as Figure 3 , including the following steps:

[0066] S1. Obtain the image to be judged, usually an image set;

[0067] S2. Use Gaussian high-pass filtering to obtain the frequency band information of the image to be judged;

[0068] S3, using the above formula (2) to obtain the image features corresponding to the image to be judged;

[0069] S4. Inputting the image features into an image structure discrimination model, the image structure discrimination model can determine which images have intact image structures, and the image structure discrimination model outputs image features indicating that the image structures have not been destroyed;

[0070] S5. The image features corresponding to the image structure not destroyed obtained in S4 are used as the input of the over-sharpening artifact discrimination model. According to the over-sharpening artifact discrimination model, it can be known which images are images without over-sharpening and / or artifacts. The over-sharpening artifact discrimination model outputs image features without over-sharpening and / or artifacts, that is, it outputs image features with intact image structure and without over-sharpening and / or artifacts. Then, images with intact structure and without over-sharpening and / or artifacts can be obtained, that is, high-quality images can be obtained. According to a high-quality image acquisition method, it can be completed to determine whether all images are high-quality images and to complete the classification of high-quality images and non-high-quality images in the image set.

[0071] The output of the above-mentioned image structure discrimination model is not limited to image features and may also be an image with intact image structure or an image with a label indicating that the image structure is intact. The output of the above-mentioned over-sharpening artifact discrimination model is not limited to image features and may also be an image without over-sharpening and / or artifacts or an image with a label indicating that the image is free of over-sharpening and / or artifacts. The definition of the degree of over-sharpening is not specifically defined and can be determined based on actual needs. For example, over-sharpening may correspond to severe over-sharpening, while no over-sharpening may correspond to no significant over-sharpening.

[0072] The present invention proposes a high-quality image classification method and system based on frequency band information, which uses a binary tree to decompose the optimization target into two binary classification problems that can be solved using a support vector machine. It is used to solve the problem that images acquired with undesirable structures, such as destroyed structures, over-sharpening, and severe artifacts, which seriously affect the evaluation of images by the non-reference image evaluation method based on image processing. Through actual testing, the accuracy of the present invention in distinguishing high-quality images has reached more than 90%, and the distinction results are basically consistent with the subjective visual effects of people. The present invention greatly reduces the burden on staff in the work of distinguishing image quality and improves the efficiency of image evaluation. The present invention solves the problem of interference of poor-quality images on image evaluation work and fills the gap in image quality classification. A high-quality image acquisition method based on this has accurate results and is applicable to various fields. It replaces the human eye's discrimination, improves the recognition efficiency of high-quality images, and has certain engineering significance.

[0073] The present invention creatively proposes a method of obtaining image features through frequency band information, and creatively proposes image features - image information entropy, high-frequency component information entropy, medium and low-frequency component information entropy, the mean of the high-frequency component and the gradient mean of the high-frequency component. Based on image features, training is performed to obtain a model that accurately judges whether an image is of high quality, thereby enabling accurate discrimination and classification of high-quality images.

Claims

1. A high-quality image classification method based on frequency band information, characterized in that: include: A training set is obtained, and labels are set for images in the training set. The labels are divided into two categories: whether the images are damaged, and whether there are oversharpening and / or artifacts. Use Gaussian high-pass filtering to obtain the frequency band information of the image; Use frequency band information to obtain image features; Using the image features of the images in the training set, an image structure discrimination model and an over-sharpening artifact discrimination model are trained; high-quality images in the image set can be obtained based on the trained image structure discrimination model and over-sharpening artifact discrimination model; The image structure discrimination model and the training over-sharpening artifact discrimination model are both machine learning models based on a binary tree support vector machine; The method of obtaining image features by using frequency band information is specifically as follows: （2） in, 、 and They are image information entropy, high-frequency component information entropy, and medium- and low-frequency component information entropy respectively; is the mean of the high-frequency components; is the gradient mean of the high-frequency component; is the input image; for The corresponding high-frequency components, for The corresponding mid- and low-frequency components; is the gradient of the high-frequency component; and is the size of the image, m is the width of the image, and n is the height of the image; Number the pixels; is the one-dimensional entropy of the image; The one-dimensional entropy of the image Expressed as: （3） in, is a pixel value of the image, for The probability of appearing in the image.

2. The high-quality image classification method based on frequency band information according to claim 1, characterized in that: The method also includes the steps of obtaining a test set consisting of images and setting the labels for the images in the test set, wherein the test set is used to test the accuracy of the obtained test image structure discrimination model and over-sharpening artifact discrimination model.

3. The high-quality image classification method based on frequency band information according to claim 1, characterized in that: The transfer function of the Gaussian high-pass filter is: （1） in, is the cutoff frequency; is a point in the frequency domain The distance to the frequency origin, that is .

4. The high-quality image classification method based on frequency band information according to claim 1, characterized in that: The image structure discrimination model is used to determine whether the image structure is destroyed, and the over-sharpening artifact discrimination model is used to determine whether the image has over-sharpening and / or artifacts. The structure of the high-quality image is not destroyed and there is no over-sharpening and / or artifacts.

5. A high-quality image classification system based on frequency band information, characterized in that: include: A first acquisition module is configured to obtain a training set, wherein images in the training set are provided with labels, and the labels are divided into two categories: whether the images are damaged, and whether there are oversharpening and / or artifacts; A first calculation module is used to obtain frequency band information of the image using Gaussian high-pass filtering; A second calculation module is used to obtain image features using frequency band information; A training module is used to train an image structure discrimination model and an over-sharpening artifact discrimination model using image features in a training set; high-quality images in the image set can be obtained based on the trained image structure discrimination model and over-sharpening artifact discrimination model; The image structure discrimination model and the training over-sharpening artifact discrimination model are both machine learning models based on a binary tree support vector machine; The method of obtaining image features by using frequency band information is specifically as follows: （2） in, 、 and They are image information entropy, high-frequency component information entropy, and medium- and low-frequency component information entropy respectively; is the mean of the high-frequency components; is the gradient mean of the high-frequency component; is the input image; for The corresponding high-frequency components, for The corresponding mid- and low-frequency components; is the gradient of the high-frequency component; and is the size of the image, m is the width of the image, and n is the height of the image; Number the pixels; is the one-dimensional entropy of the image; The one-dimensional entropy of the image Expressed as: （3） in, is a pixel value of the image, for The probability of appearing in the image.

6. A method for obtaining high-quality images, characterized in that: The steps include: S1. Obtain the image to be judged; S2. Use Gaussian high-pass filtering to obtain frequency band information of the image to be judged; S3. Obtaining image features of the image to be determined based on frequency band information of the image to be determined; S4, using the image features obtained in S3 as input to the image structure discrimination model; S5. Inputting the image features corresponding to the image structure not being destroyed output by the image structure discrimination model in S4 into the over-sharpening artifact discrimination model, and inputting the image features corresponding to the image without over-sharpening and / or artifacts output by the over-sharpening artifact discrimination model; S6. Obtain a high-quality image based on the image features corresponding to the absence of over-sharpening and / or artifacts obtained in S5; The image structure discrimination model and the over-sharpening artifact discrimination model are both obtained by using a high-quality image classification method based on frequency band information as described in any one of claims 1 to 4.

Images