CN110782445A - No-reference image quality evaluation method and system - Google Patents
No-reference image quality evaluation method and system Download PDFInfo
- Publication number
- CN110782445A CN110782445A CN201911023113.2A CN201911023113A CN110782445A CN 110782445 A CN110782445 A CN 110782445A CN 201911023113 A CN201911023113 A CN 201911023113A CN 110782445 A CN110782445 A CN 110782445A
- Authority
- CN
- China
- Prior art keywords
- image
- network model
- evaluated
- distorted
- training
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30168—Image quality inspection
Landscapes
- Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a no-reference image quality evaluation method and a no-reference image quality evaluation system, wherein the method comprises the following steps: acquiring a distorted image to be evaluated from a database; restoring the distorted image to be evaluated by using a preset restoration model to obtain a restored image; and inputting the image difference data of the distorted image to be evaluated and the restored image into a preset evaluation network model, and obtaining an image quality evaluation result of the distorted image to be evaluated, which is output by the preset evaluation network model. The image difference data is large if the distortion degree of the distorted image is large, and the image difference data is small if the distortion degree of the distorted image is small. And inputting the image difference data into a preset evaluation network model so as to obtain an image quality evaluation result of the distorted image. According to the method, the distorted image is identified without manually selecting statistical characteristics or knowing the distortion type of the distorted image, and an accurate and effective image quality evaluation result can be obtained by simulating the processing operation of a human visual system.
Description
Technical Field
The invention relates to the technical field of communication, in particular to a method and a system for evaluating the quality of a non-reference image.
Background
Images are an important source of information for human perception and computer vision. Due to imperfections in the imaging system, transmission media, compression schemes, storage devices, etc., it is inevitable that images will be distorted and degraded to varying degrees. In some image processing systems or video processing systems, the image quality of an image to be processed needs to be known so as to be used as a judgment basis for whether to perform subsequent operations, so that the evaluation of the image quality has important practical application value.
Currently, image quality evaluation (IQA) can be divided into subjective evaluation and objective evaluation. Subjective evaluation, in which a plurality of observers take subjective feeling as image quality scores and obtain an average subjective score (MOS) of an image, is superior in accuracy and practicality, but is inefficient in manual processing.
Objective evaluation, namely simulating the perception process of the human visual system to the image quality, and constructing an automatic image quality evaluation algorithm which is consistent with subjective evaluation as much as possible. The image quality evaluation can be divided into full-reference image quality evaluation (FR-IQA), half-reference image quality evaluation (RR-IQA), and no-reference image quality evaluation (NR-IQA), depending on whether or not an original image is used as a reference in the image quality evaluation process.
At present, an original reference image cannot be obtained usually when image quality evaluation is carried out in an actual application scene, and image quality evaluation can be carried out under the condition that the original reference image is absent without reference image quality evaluation, and the image quality evaluation is consistent with the actual application scene, so that the method is a key point of current research.
The conventional non-reference image quality evaluation method has the idea that some statistical characteristics of a distorted image and an original reference image (i.e., an undistorted image) are different, so that the statistical characteristics can be manually found, and a model is trained according to the statistical characteristics to calculate an evaluation result based on the trained model. However, in actual scenes, the distortion types of images vary widely, multiple distortion types usually exist simultaneously, and manual searching for statistical features is difficult.
Currently, some schemes propose to achieve reference-free image quality evaluation using a deep neural network DNN. The main principle is to train the DNN based on annotation data so that the DNN learns the mapping from the distorted image to the annotated mean subjective score (MOS), and subsequently use the trained DNN to calculate an image quality assessment.
However, DNN works poorly in the face of distorted images of unknown type. The number of IQA annotation datasets currently disclosed is limited, containing only a few specific distortion types (e.g., white gaussian noise, gaussian blur, JPEG compression distortion, etc.). Therefore, DNNs trained using these labeling data are difficult to achieve satisfactory results in the face of images of unknown distortion types in real scenes.
Disclosure of Invention
In view of this, the present invention provides a method and a system for evaluating quality of a non-reference image. The method does not need to manually select statistical characteristics, can effectively deal with the distorted images of unknown distortion types, and obtains more accurate and effective image quality evaluation results.
In order to achieve the above object, the present application provides the following technical features:
a no-reference image quality evaluation method, the method comprising:
acquiring a distorted image to be evaluated from a database;
restoring the distorted image to be evaluated by using a preset restoration model to obtain a restored image;
inputting image difference data of the distorted image to be evaluated and the restored image into a preset evaluation network model to obtain an image quality evaluation result of the distorted image to be evaluated, which is output by the preset evaluation network model;
and the image quality evaluation result of the distorted image to be evaluated is used as a judgment basis for subsequent processing operation.
Optionally, the repairing the distorted image to be evaluated to obtain a repaired image includes:
under the condition that the preset restoration model is a post-training restoration network model, inputting the distorted image to be evaluated to the post-training restoration network model to obtain a restoration image output by the post-training restoration network model; or the like, or, alternatively,
and under the condition that the preset restoration model is the trained variational self-encoder, inputting the distorted image to be evaluated to the trained variational self-encoder to obtain a restoration image output by the trained variational self-encoder.
Optionally, when the preset restoration model is a trained restoration network model, before obtaining the distorted image to be evaluated from the database, the method further includes: storing the trained restoration network model;
under the condition that the preset restoration model is a trained variational self-encoder, before obtaining a distorted image to be evaluated from a database, the method further comprises the following steps: storing the trained variational self-encoder;
the training process of the trained restoration network model comprises the following steps:
initializing a repair network model and a discrimination network model; wherein the discrimination network model is used for assisting in training the restoration network model;
alternately training a repairing network model and a judging network model in a confrontation mode so that a repairing image output by the repairing network model is continuously close to an original reference image;
and obtaining a post-training repair network model after the training end condition is reached.
Optionally, the alternately training the repairing network model and the discriminating network model by adopting a countermeasure mode includes:
under the condition that a repairing network model is not changed temporarily, obtaining a repairing image output after the repairing network model repairs a distorted image and an original reference image corresponding to the distorted image, adding the group of samples consisting of the repairing image and the original reference image to a training sample set of a judging network, wherein the training sample set also comprises a plurality of groups of samples consisting of two same original reference images; training the discrimination network model based on a training sample set of the discrimination network model so that the discrimination network model has the capability of giving a higher score to a restored image with high similarity to the original reference image and giving a lower score to a restored image with low similarity to the original reference image;
under the condition that the judging network model is not changed temporarily, inputting a distorted image to the repairing network model and obtaining a repairing image, inputting the repairing image and the original reference image to the judging network model and obtaining an output probability fed back by the judging network model, and adjusting the repairing network model based on a parameter set comprising the output probability so that the repairing image output by the repairing network model is close to the original reference image continuously;
and alternately executing the training process of distinguishing the network model and repairing the network model until the training end condition is reached.
Optionally, the adjusting the repair network model based on the parameter set including the output probability includes:
calculating a loss function for repairing the network model based on the parameter set;
adjusting the restoration network model according to the loss function of the restoration network model;
wherein the parameter set further comprises: repairing pixel loss between the image and the original reference image; and/or, simulating the human visual system to determine a perceptual loss between the restored image and the original reference image;
wherein, in case the parameter set comprises three parameters of pixel loss, perceptual loss and output probability, then the calculating a loss function of the repair network model based on the parameter set comprises: and determining the weighted sum of the three parameters of the pixel loss, the perception loss and the output probability as a loss function of the repair network model.
Optionally, before the obtaining the distorted image to be evaluated from the database, the method further includes: storing the preset evaluation network model;
the training process of the preset evaluation network model comprises the following steps:
under the condition that the preset restoration model is a restoration network model after training, training an evaluation network model by using the image difference data of the distorted image and the restoration image output by the restoration network model and the average subjective score corresponding to the distorted image, so that the evaluation network model learns that the image difference data is mapped to the average subjective score in the training process;
or the like, or, alternatively,
and under the condition that the preset restoration model is a trained variational self-encoder, training an evaluation network model by using image difference data of the distorted image and the restoration image output by the variational self-encoder and an average subjective score corresponding to the distorted image, so that the evaluation network model learns that the image difference data is mapped to the average subjective score in the training.
Optionally, when the application scene is face recognition, the obtaining of the distorted image to be evaluated from the database includes: acquiring a target video from a monitoring video stream database, determining a target face image from the target video, and taking the target face image as the distorted image to be evaluated; or, determining a target face image from a monitoring image database, and taking the target face image as the distorted image to be evaluated;
after the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if so, using the distorted image to be evaluated for face recognition operation; if not, discarding the distorted image to be evaluated;
under the condition that the application scene is a video conference, the step of obtaining the distorted image to be evaluated from the database comprises the following steps: acquiring a target video from a conference video stream database, acquiring a target image from the target video, and taking the target image as the distorted image to be evaluated;
after the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if yes, determining that the current communication quality is good, and starting or continuing the video conference; if not, executing a dynamic adjustment strategy to improve the communication quality;
under the condition that the application scene is video on demand or live broadcast, the step of obtaining the distorted image to be evaluated from the database comprises the following steps: acquiring a target video from a video stream database, acquiring a target image from the target video, and taking the target image as the distorted image to be evaluated;
after the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if yes, determining that the current communication quality is good, and starting or continuing video on demand or live broadcasting; if not, executing a dynamic adjustment strategy for improving the communication quality.
A no-reference image quality evaluation system comprising:
the processing equipment is used for acquiring a distorted image to be evaluated from a local database or a database of third-party equipment, and restoring the distorted image to be evaluated by using a preset restoration model to acquire a restored image; inputting image difference data of the distorted image to be evaluated and the restored image into a preset evaluation network model to obtain an image quality evaluation result of the distorted image to be evaluated, which is output by the preset evaluation network model;
and the image quality evaluation result of the distorted image to be evaluated is used as a judgment basis for subsequent processing operation.
Optionally, the processing device is further configured to use an image quality evaluation result of the distorted image to be evaluated as a judgment basis for a subsequent processing operation; alternatively, the first and second electrodes may be,
the processing device is further configured to send an image quality evaluation result of the distorted image to be evaluated to the third-party device, so as to be used as a judgment basis for subsequent processing operation of the third-party device.
A training method of a reference-free image quality evaluation model comprises the following steps:
initializing a restoration network model, judging a network model and evaluating the network model;
alternately training a restoration network model and a discrimination network model in an antagonistic mode, and obtaining a post-training restoration network model after a training end condition is reached;
and training the evaluation network model by using the image difference data of the distorted image and the restored image output by the restoration network model and the average subjective score corresponding to the distorted image, and obtaining the post-training evaluation network model after the training end condition is reached.
Through the technical means, the following beneficial effects can be realized:
the research shows that: when a human vision system sees an image, the human vision system can restore the distorted image instinctively to fill some detail information to obtain a restored image, and then the image quality evaluation is carried out on the restored image.
The method simulates a human visual system to evaluate the image quality, and does not directly use the distorted image after obtaining the distorted image to be evaluated, but repairs the distorted image to obtain a repaired image. Image difference data for the distorted image and the repaired image is then determined.
It is understood that the image difference data is larger if the distortion degree of the distorted image is larger, and the image difference data is smaller if the distortion degree of the distorted image is smaller. And inputting the image difference data into a preset evaluation network model so as to obtain an image quality evaluation result of the distorted image output by the preset evaluation network model.
According to the method, the distorted image is identified without manually selecting statistical characteristics or knowing the distortion type of the distorted image, and an accurate and effective image quality evaluation result can be obtained by simulating the processing operation of a human visual system.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a flowchart of a method for evaluating quality of a non-reference image according to an embodiment of the present invention;
FIG. 2 is a flowchart of a training method of a non-reference image quality evaluation model according to an embodiment of the present disclosure;
3a-3b are flowcharts illustrating still another training method for a non-reference image quality evaluation model disclosed in the embodiments of the present application;
fig. 4a-4b are schematic structural diagrams of a no-reference image quality evaluation system disclosed in an embodiment of the present application.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The method does not need to manually select statistical characteristics, can effectively deal with the distorted image with unknown distortion type, and obtains more accurate and effective image quality evaluation results. The present invention will be described in detail below.
The technical personnel of the invention find out through research that: the human vision system has a natural image processing function, and when seeing an image, the human vision system can restore the distorted image instinctively to fill some detail information to obtain a restored image, and then the image quality evaluation is carried out on the restored image.
Therefore, the invention provides a scheme for simulating the human visual system to perform image quality evaluation so as to obtain accurate and effective image quality evaluation results. The implementation of the present invention is explained in detail below.
The invention provides a no-reference image quality evaluation method. The no-reference image quality evaluation method may be applied to a processing device, and may include the following steps, see fig. 1:
step S100: and storing a preset restoration model for restoring the distorted image and a preset evaluation network model for evaluating the distorted image.
The step is optional operation, and the step can not be executed if the preset repairing model and the preset evaluation network model are stored in advance.
Step S101: and acquiring the distorted image to be evaluated from the database.
It will be appreciated that the processing device may retrieve the distorted image to be evaluated from a local database or a database of a third party device. Obtaining the distorted image to be evaluated from the database can have two situations:
in the first case: and acquiring a distorted image to be evaluated from the video stream.
When the video stream is stored in the database, a target image meeting the screening condition may be determined from the video stream according to the screening condition corresponding to the application scene, and the target image may be used as a distorted image to be evaluated.
In the second case: and acquiring a distorted image to be evaluated from a plurality of images.
When a plurality of images are stored in the database, a target image that meets the screening condition may be determined from the plurality of images according to the screening condition corresponding to the application scene, and the target image may be used as a distorted image to be evaluated.
The filtering conditions in the two cases may be set according to actual application scenarios, which is not limited to this.
Step S102: and repairing the distorted image to be evaluated by using a preset repairing model to obtain a repaired image.
This step can be divided into two implementations:
the first implementation mode comprises the following steps: and presetting the repairing model as a trained repairing network model.
And under the condition that the preset restoration model is a post-training restoration network model, inputting the distorted image to be evaluated to the post-training restoration network model to obtain a restoration image output by the post-training restoration network model. The trained restoration network model is a data model designed for improving the accuracy of the restored image, the input of the trained restoration network model is a distorted image, the output of the trained restoration network model is a restored image, and the similarity between the restored image and an original reference image is high.
For details on the post-training repair network model, see the embodiment shown in fig. 2. And will not be described in detail herein.
The second implementation mode comprises the following steps: and presetting the restoration model as a variational self-encoder after training.
And under the condition that the preset restoration model is the trained variational self-encoder, inputting the distorted image to be evaluated to the trained variational self-encoder to obtain a restoration image output by the trained variational self-encoder. The variational self-encoder may be trained in advance based on the distorted image and the original reference image, thereby obtaining a post-training variational self-encoder. The post-training variational self-encoder can repair the distorted image to obtain a repaired image.
It can be understood that the repairing effect of the network model after training in the first implementation is better than that of the variational self-encoder after training in the second implementation.
Step S103: inputting image difference data of the distorted image to be evaluated and the restored image into a preset evaluation network model to obtain an image quality evaluation result of the distorted image to be evaluated, which is output by the preset evaluation network model; and the image quality evaluation result of the distorted image to be evaluated is used as a judgment basis for subsequent processing operation.
The training process of the preset evaluation network model in this step corresponds to step S102, and may also include two implementation manners:
under the condition that the preset restoration model is a restoration network model after training, training an evaluation network model by using image difference data of a distorted image and a restoration image output by the restoration network model and an average subjective score corresponding to the distorted image, so that the evaluation result of the restoration image output by the evaluation network model to the restoration network model is continuously close to the average subjective score;
correspondingly, the evaluation network model trained in the mode can be used as a preset evaluation network model, image difference data is received, and an image quality evaluation result is output after calculation.
Or the like, or, alternatively,
and under the condition that the preset restoration model is the trained variational self-encoder, training an evaluation network model by using image difference data of the distorted image and the restoration image output by the variational self-encoder and an average subjective score corresponding to the distorted image, so that the evaluation result of the restoration image output by the variational self-encoder by the evaluation network model is continuously close to the average subjective score.
Correspondingly, the evaluation network model trained in the mode can be used as a preset evaluation network model, image difference data is received, and an image quality evaluation result is output after calculation.
It can be understood that, in the case that the distortion degree of the distorted image to be evaluated is larger, the more the distorted image to be evaluated is repaired, the larger the image difference data between the distorted image to be evaluated and the repaired image is, the smaller the corresponding image quality evaluation result is.
Under the condition that the distortion degree of the distorted image to be evaluated is smaller, the repairing of the distorted image to be evaluated is less, the image difference data of the distorted image to be evaluated and the repaired image is smaller, and the corresponding image quality evaluation result is larger.
And inputting image difference data to a preset evaluation network model, and outputting an image quality evaluation result of the distorted image to be evaluated after calculation of the preset evaluation network model.
In the actual application process, according to different calculation modes of the preset evaluation network model, the preset evaluation network model may include other input parameters, such as a distorted image to be evaluated, besides the image difference data, and details are not repeated herein.
The post-processing device may also perform the following operations at step S103:
step S1041: the processing equipment uses the image quality evaluation result of the distorted image to be evaluated as a judgment basis of subsequent processing operation; alternatively, the first and second electrodes may be,
step S1042: and sending the image quality evaluation result of the distorted image to be evaluated to third-party equipment to be used as a judgment basis for subsequent processing operation of the third-party equipment.
The processing device may adopt step S1041 or step S1042, which may be determined according to an actual application scenario and is not limited herein.
Through the technical means, the following beneficial effects can be realized:
the research shows that: when a human vision system sees an image, the human vision system can repair the distorted image instinctively to fill some detail information to obtain a repaired image, and then image quality evaluation is carried out on the repaired image to obtain an accurate and effective image quality evaluation result.
The method simulates a human visual system to evaluate the image quality, and does not directly use the distorted image after obtaining the distorted image to be evaluated, but repairs the distorted image to obtain a repaired image. Image difference data for the distorted image and the repaired image is then determined.
It is understood that the image difference data is large if the distortion degree of the distorted image is large (i.e., the image quality evaluation is low), and the image difference data is small if the distortion degree of the distorted image is small (i.e., the image quality evaluation is low). And inputting the image difference data into a preset evaluation network model so as to obtain an image quality evaluation result of the distorted image output by the preset evaluation network model.
According to the method, the distorted image is repaired by simulating a human visual system without manually selecting statistical characteristics to identify the distorted image or knowing the distortion type of the distorted image, and an accurate and effective image quality evaluation result is obtained based on the image difference data of the repaired image and the distorted image.
Because the trained restoration network model in the first implementation manner of step 102 is a data model designed to improve the accuracy of the restored image, the input of the trained restoration network model is a distorted image, the output of the trained restoration network model is a restored image, and the similarity between the restored image and the original reference image is high. Namely, the trained repairing network model has a high repairing effect. Therefore, a preferred implementation manner of the reference-free image quality evaluation method is provided:
on the basis of the embodiment of fig. 1, the trained repairing network model and the evaluation network model obtained based on the repairing network model are stored in step S100 in advance; in step S102, a trained restoration network model is adopted to restore the distorted image to be evaluated and obtain a restored image; in step S103, based on the evaluation network model obtained by repairing the network model, the image difference data is received and the image quality evaluation result of the image to be evaluated is output. Steps S1041 and S1042 correspond to fig. 1.
The following describes a training process of the post-training restoration network model and the acquisition of the evaluation network model based on the restoration network model to explain the principle that the post-training restoration network model and the evaluation network model can realize more accurate image quality evaluation results.
Referring to fig. 2, the present invention provides a training method for a reference-free image quality evaluation model, including:
step S200: a distorted image set and a reference image set are constructed.
Distorted image set B
distIncluding a plurality of distorted images I
distMean subjective score S corresponding to distorted image, reference image set B
refComprising a plurality of images I with distortion
distCorresponding original reference picture I
ref。
Step S201: initializing a repair network model, judging a network model and evaluating the network model.
Initializing a repair network model R
θAnd the restoration network model is used for restoring the distorted image to obtain a restoration image, the input of the restoration image is the distorted image, and the output of the restoration image is the restoration image.
Initialization discrimination network model
Wherein the discrimination network model is used for judging whether the restored image can be regarded as an original reference image. To quantify this determination, the output of the discrimination network is a probability that the restored image (first image) can be considered as the original reference image (second image).
In other words, the purpose of training the discriminant network model is to make the discriminant network model learn to distinguish whether the first image is the original reference image or the restored image, so as to assign the highest probability to the original reference image, assign different scores to the restoration process based on the restored image, and after training for several times, the discriminant network model has the capability of assigning a higher score to the restored image with high similarity to the original reference image and assigning a lower score to the restored image with low similarity to the original reference image.
Initializing evaluation network model E
ωWherein, the input of the evaluation network model comprises a distorted image and a repaired image obtained by repairing the distorted image by the network modelThe difference data is output as an image quality evaluation result of the distorted image. The evaluation network model is used for scoring the distortion degree of the distorted image.
Step S202: and training the repairing network model and the judging network model alternately in a countermeasure mode so that the repairing image output by the repairing network model is continuously close to the original reference image.
The basic principle of the countermeasure mode is as follows: the method comprises the steps of restoring a distorted image of a network model to obtain a restored image, training a judgment network model to judge the restoration effect of the restored network model, and improving the restored network model based on the restoration effect determined by the judgment network model so that the restored network model can restore the distorted image more perfectly.
Step S202 may perform the training operation by adopting the following steps:
s1: under the condition that a repairing network model is not changed temporarily, obtaining a repairing image output after the repairing network model repairs a distorted image and an original reference image corresponding to the distorted image, adding the group of samples consisting of the repairing image and the original reference image to a training sample set of a judging network, wherein the training sample set also comprises a plurality of groups of samples consisting of two same original reference images; the discriminant network model is trained based on a training sample set of the discriminant network model, so that the discriminant network model has the capability of giving a higher score to a restored image with high similarity to the original reference image and giving a lower score to a restored image with low similarity to the original reference image.
S2: under the condition that the judging network model is not changed temporarily, inputting a distorted image to the repairing network model and obtaining a repairing image, inputting the repairing image and the original reference image to the judging network model and obtaining an output probability fed back by the judging network model, and adjusting the repairing network model based on a parameter set comprising the output probability so that the repairing image output by the repairing network model is close to the original reference image continuously.
Wherein said adapting the repair network model based on the set of parameters including the output probabilities comprises: calculating a loss function for repairing the network model based on the parameter set; adjusting the restoration network model according to the loss function of the restoration network model;
wherein the parameter set further comprises: repairing pixel loss between the image and the original reference image; and/or, simulating the human visual system to determine a perceptual loss between the restored image and the original reference image;
wherein, in case the parameter set comprises three parameters of pixel loss, perceptual loss and output probability, then the calculating a loss function of the repair network model based on the parameter set comprises: and determining the weighted sum of the three parameters of the pixel loss, the perception loss and the output probability as a loss function of the repair network model.
S3: and alternately executing the training process of distinguishing the network model and repairing the network model until the training end condition is reached.
Step S203: and training the evaluation network model by using the image difference data of the distorted image and the repaired image output by the repair network model and the average subjective score corresponding to the distorted image, so that the evaluation network model learns the mapping from the image difference data to the average subjective score in the training.
For specific implementation of steps S1-S3, see the specific embodiment of the training process shown in fig. 3.
Step S203: and training the evaluation network model by using the image difference data of the distorted image and the repaired image output by the repair network model and the average subjective score corresponding to the distorted image, so that the evaluation result of the evaluation network model on the repaired image output by the repair network model is continuously close to the average subjective score.
Referring to fig. 3a and 3b, a detailed description of a specific training process for repairing the network model and evaluating the network model after training is given below by way of a specific example:
step S301: and determining a training sample set for repairing the network model and a training sample set for evaluating the network model.
From distorted image sets B
distThe number of m samples is randomly drawn,
and finding out the original reference image corresponding to each sample
One distorted image and the corresponding original reference image are a group of samples, and the groups of samples can form a training sample set of the repairing network model.
The training sample set for evaluating the network model comprises m distorted images in the training sample set for repairing the network model, and the average subjective score S of the m distorted images
i,i＝1,2,…,m。
Step S302: and determining a training sample set of the discriminant network model.
From the original reference picture set B
refIn the process of randomly taking m samples,
two identical original reference pictures are a set of samples.
Under the condition that the restoration network model is not changed temporarily, acquiring a restoration image output after the restoration network model restores the distorted image and an original reference image corresponding to the distorted image, and acquiring a plurality of groups of samples consisting of the restoration image and the original reference image.
And constructing a training sample set of the discriminant network model, wherein the training sample set comprises a plurality of groups of samples consisting of two same original reference images and a plurality of groups of samples consisting of the repaired images and the original reference images.
Step S303: and training the discriminant network model based on the training sample set of the discriminant network model.
One or more groups of samples are randomly selected from the training sample set and input into the discriminant network model. Then, the discriminant network model calculates a loss function to adjust trainable parameters of the discriminant network model using the loss function so that the discriminant network model may be more complete.
Loss function of the discriminant network model
The method comprises the following steps:
loss function in case of discriminating two identical original reference images as input of network model
Is a first loss function
Loss function in case of discriminating input of network model as restored image and original reference image
As a second loss function
Wherein theta is a trainable parameter of the repairing network model,
to discriminate trainable parameters of the network model, B
distRepresenting a randomly decimated sample set of distorted images, R
θA model of the repair network is represented,
representation of a discriminating network model, I
refRepresenting the original reference picture, I
distRepresenting a distorted image, B
refRepresenting an original reference image sample set which is randomly extracted, wherein m is the number of samples of the sample set; i, j each represents any sample in the m sample sets.
Discriminating network model D in updating
φThe parameter can be expressed by the following formula phi: -phi- α▽
φL
DWherein, ▽
φL
DLoss function L representing a discriminative network model
DFor theThe gradient of the parameter φ, α, represents the learning rate, although other ways of adjusting the trainable parameters of the discriminative network model may be used, and are not further described herein.
Step S304: and training the restoration network model based on the training sample set of the restoration network model.
And randomly selecting a group of samples in the training sample set, and inputting the group of samples into the repairing network model. The repair network model then calculates a loss function to use the loss function to adjust trainable parameters of the repair network model so that the repair network model may be more complete.
Under the condition that the loss function of the repairing network model comprises pixel loss, perception loss and the output probability of the judging network model, the loss function of the repairing network model comprises the following steps:
λ
1、λ
2and λ
3The weights, respectively, may be determined on a case-by-case basis.
Wherein the pixel loss L between the restored image and the original reference image
pixThe method can comprise the following steps:
wherein simulating the human visual system determines the perceptual loss L between the restored image and the original reference image
perThe method comprises the following steps:
wherein, the perception loss adopts the perception characteristic vector of the image extracted from the preset image content identification network model,
a feature tensor, H, representing the output of a certain intermediate layer of the network model of the content of the preset recognition image
j，W
j，C
jRespectively representing the height, width and number of layers of the intermediate layer feature tensor;
wherein, the output probability of the network model to the repair network model repair result is judged
The method can comprise the following steps:
updating a repair network model R
θThe network can train the parameter theta and adjust the formula as theta- α▽
θL
RWherein, ▽
θL
RLoss function L representing a repair network model
Rα represents the learning rate for the gradient of the parameter θ, although other ways of adjusting the trainable parameters of the discriminative network model are possible and are not further described herein.
Step S305: and training the evaluation network model based on the training sample set of the evaluation network model.
On the basis of step S4, image difference data of the distorted image and the restored image is determined
And determining the average subjective score corresponding to the distorted image from the training sample set.
Input device
And mean subjective score s
iAnd (4) evaluating the network model. The evaluation network model then calculates a loss function to use the loss function to adjust trainable parameters of the evaluation network model so that the evaluation network model may be more sophisticated.
Evaluating a loss function L of a network model
E(ω) may include:
after calculating the loss function of the evaluation network model, updating the evaluation network model E
ωTrainable parameter ω - α▽
ωL
EWherein, ▽
ωL
ELoss function L representing an evaluation network model
Eα represents the learning rate for the gradient of parameter ω other ways of adjusting the trainable parameters of the discriminative network model are of course possible and are not further enumerated here.
Step S306: and repeating the steps S301-S305 until the training end condition is met, and obtaining the repaired network model and the evaluation network model after training.
It is understood that the process of training the evaluation network model may be performed after the end of the training of the repair network model, or may be performed in each cycle of training the repair network model.
Through the training process, the following steps are known:
in the process of training the restoration network model, the restoration network model and the judgment network model are alternately trained in a confrontation mode, namely, the judgment network model is utilized to assist in training the restoration network model, so that the restoration network model obtained by training is more accurate and efficient.
The evaluation network model is used in the training process and also used in the repair network model, so that the trained evaluation network model is more accurate and efficient after the repair network model is more accurate and efficient.
Referring to fig. 4a, the present invention provides a no-reference image quality evaluation system, which may include:
the processing device needs to obtain a preset restoration model and a preset evaluation network model in advance, and the following two modes can be adopted:
the first mode is as follows:
the processing equipment executes training operation, obtains a preset restoration model required for restoring the distorted image, evaluates a preset evaluation network model of the restored image, and stores the preset restoration model and the preset evaluation network model.
The second mode is as follows:
the method comprises the steps that training equipment executes training operation, a preset repairing model required for repairing a distorted image is obtained, a preset evaluation network model used for evaluating the repaired image is obtained, the preset repairing model and the preset evaluation network model are sent to processing equipment, and the processing equipment is used for receiving and storing the preset repairing model and the preset evaluation network model.
The online operation process comprises the following steps:
a plurality of clients for generating video streams or images during use and transmitting the video streams or images to the image processing device;
the processing device is used for receiving and storing a video stream or an image and storing the video stream or the image into a local database, acquiring a distorted image to be evaluated from the local database, restoring the distorted image to be evaluated by using a preset restoration model to obtain a restored image, inputting image difference data of the distorted image to be evaluated and the restored image to a preset evaluation network model, and acquiring an image quality evaluation result of the distorted image to be evaluated, which is output by the preset evaluation network model;
and the processing equipment is also used for using the image quality evaluation result of the distorted image to be evaluated as a judgment basis of subsequent processing operation.
Referring to fig. 4b, the present invention provides a no-reference image quality evaluation system, which may include:
the processing device needs to obtain a preset restoration model and a preset evaluation network model in advance, and the following two modes can be adopted:
the first mode is as follows:
the processing equipment executes training operation, obtains a preset restoration model required for restoring the distorted image, evaluates a preset evaluation network model of the restored image, and stores the preset restoration model and the preset evaluation network model.
The second mode is as follows:
the method comprises the steps that training equipment executes training operation, a preset repairing model required for repairing a distorted image is obtained, a preset evaluation network model used for evaluating the repaired image is obtained, the preset repairing model and the preset evaluation network model are sent to processing equipment, and the processing equipment is used for receiving and storing the preset repairing model and the preset evaluation network model.
The online operation process comprises the following steps:
a plurality of clients, which are used for generating video streams or images in the using process and sending the video streams or the images to third-party equipment;
the third-party equipment is used for receiving and storing the video stream or the image into the local database, acquiring the distorted image to be evaluated from the local database and sending the distorted image to be evaluated to the processing equipment (equivalent to the processing equipment); and the evaluation module is also used for receiving the image quality evaluation result of the distorted image to be evaluated, so as to be used as a judgment basis for subsequent processing operation.
And the processing device is used for restoring the distorted image to be evaluated to obtain a restored image, inputting image difference data of the distorted image to be evaluated and the restored image to a preset evaluation network model, obtaining an image quality evaluation result of the distorted image to be evaluated, which is output by the preset evaluation network model, and sending the image quality evaluation result of the distorted image to be evaluated to a third-party device.
The invention can be used in quality detection schemes involving video streams and images, and is further explained by taking the following three application scenarios as examples:
under the condition that the application scene is face recognition, the step of obtaining the distorted image to be evaluated from the database comprises the following steps: acquiring a target video from a monitoring video stream database, determining a target face image from the target video, and taking the target face image as the distorted image to be evaluated; or, determining a target face image from a monitoring image database, and taking the target face image as the distorted image to be evaluated;
after the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if so, using the distorted image to be evaluated for face recognition operation; and if not, discarding the distorted image to be evaluated.
Under the condition that the application scene is a video conference, the step of obtaining the distorted image to be evaluated from the database comprises the following steps: and acquiring a target video from a conference video stream database, acquiring a target image from the target video, and taking the target image as the distorted image to be evaluated.
After the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if yes, determining that the current communication quality is good, and starting or continuing the video conference; if not, executing a dynamic adjustment strategy for improving the communication quality.
Under the condition that the application scene is video on demand or live broadcast, the step of obtaining the distorted image to be evaluated from the database comprises the following steps: acquiring a target video from a video stream database, acquiring a target image from the target video, and taking the target image as the distorted image to be evaluated;
after the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if yes, determining that the current communication quality is good, and starting or continuing video on demand or live broadcasting; if not, executing a dynamic adjustment strategy for improving the communication quality.
According to the technical characteristics, the invention has the following beneficial effects:
the research shows that: when a human vision system sees an image, the human vision system can restore the distorted image instinctively to fill some detail information to obtain a restored image, and then the image quality evaluation is carried out on the restored image.
The method simulates a human visual system to evaluate the image quality, and does not directly use the distorted image after obtaining the distorted image to be evaluated, but repairs the distorted image to obtain a repaired image. Image difference data for the distorted image and the repaired image is then determined.
It is understood that the image difference data is larger if the distortion degree of the distorted image is larger, and the image difference data is smaller if the distortion degree of the distorted image is smaller. And inputting the distorted image and the image difference data into a preset evaluation network model so as to obtain an image quality evaluation result of the distorted image output by the preset evaluation network model.
According to the method, the distorted image is identified without manually selecting statistical characteristics or knowing the distortion type of the distorted image, and an accurate and effective image quality evaluation result can be obtained by simulating the processing operation of a human visual system.
Specific views of the non-reference image quality evaluation system shown in fig. 4a-4b can be detailed in the embodiments shown in fig. 1 to 3, and are not described herein again.
The functions described in the method of the present embodiment, if implemented in the form of software functional units and sold or used as independent products, may be stored in a storage medium readable by a computing device. Based on such understanding, part of the contribution of the embodiments of the present invention to the prior art or part of the technical solution may be embodied in the form of a software product stored in a storage medium and including instructions for causing a computing device (which may be a personal computer, a processing device, a mobile computing device, a network device, or the like) to execute all or part of the steps of the method described in the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.
The embodiments are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same or similar parts among the embodiments are referred to each other.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (10)
1. A no-reference image quality evaluation method is characterized by comprising the following steps:
acquiring a distorted image to be evaluated from a database;
restoring the distorted image to be evaluated by using a preset restoration model to obtain a restored image;
inputting image difference data of the distorted image to be evaluated and the restored image into a preset evaluation network model to obtain an image quality evaluation result of the distorted image to be evaluated, which is output by the preset evaluation network model;
and the image quality evaluation result of the distorted image to be evaluated is used as a judgment basis for subsequent processing operation.
2. The method according to claim 1, wherein the repairing the distorted image to be evaluated to obtain a repaired image comprises:
under the condition that the preset restoration model is a post-training restoration network model, inputting the distorted image to be evaluated to the post-training restoration network model to obtain a restoration image output by the post-training restoration network model; or the like, or, alternatively,
and under the condition that the preset restoration model is the trained variational self-encoder, inputting the distorted image to be evaluated to the trained variational self-encoder to obtain a restoration image output by the trained variational self-encoder.
3. The method of claim 2,
under the condition that the preset restoration model is a trained restoration network model, before obtaining a distorted image to be evaluated from a database, the method further comprises the following steps: storing the trained restoration network model;
under the condition that the preset restoration model is a trained variational self-encoder, before obtaining a distorted image to be evaluated from a database, the method further comprises the following steps: storing the trained variational self-encoder;
the training process of the trained restoration network model comprises the following steps:
initializing a repair network model and a discrimination network model; wherein the discrimination network model is used for assisting in training the restoration network model;
alternately training a repairing network model and a judging network model in a confrontation mode so that a repairing image output by the repairing network model is continuously close to an original reference image;
and obtaining a post-training repair network model after the training end condition is reached.
4. The method of claim 3, wherein the training the repair network model and the discriminant network model alternately in a countermeasure mode comprises:
under the condition that a repairing network model is not changed temporarily, obtaining a repairing image output after the repairing network model repairs a distorted image and an original reference image corresponding to the distorted image, adding the group of samples consisting of the repairing image and the original reference image to a training sample set of a judging network, wherein the training sample set also comprises a plurality of groups of samples consisting of two same original reference images; training the discrimination network model based on a training sample set of the discrimination network model so that the discrimination network model has the capability of giving a higher score to a restored image with high similarity to the original reference image and giving a lower score to a restored image with low similarity to the original reference image;
under the condition that the judging network model is not changed temporarily, inputting a distorted image to the repairing network model and obtaining a repairing image, inputting the repairing image and the original reference image to the judging network model and obtaining an output probability fed back by the judging network model, and adjusting the repairing network model based on a parameter set comprising the output probability so that the repairing image output by the repairing network model is close to the original reference image continuously;
and alternately executing the training process of distinguishing the network model and repairing the network model until the training end condition is reached.
5. The method of claim 4, wherein adjusting the repair network model based on the set of parameters including the output probability comprises:
calculating a loss function for repairing the network model based on the parameter set;
adjusting the restoration network model according to the loss function of the restoration network model;
wherein the parameter set further comprises: repairing pixel loss between the image and the original reference image; and/or, simulating the human visual system to determine a perceptual loss between the restored image and the original reference image;
wherein, in case the parameter set comprises three parameters of pixel loss, perceptual loss and output probability, then the calculating a loss function of the repair network model based on the parameter set comprises: and determining the weighted sum of the three parameters of the pixel loss, the perception loss and the output probability as a loss function of the repair network model.
6. The method of claim 2, further comprising, prior to said obtaining the distorted image to be evaluated from the database: storing the preset evaluation network model;
the training process of the preset evaluation network model comprises the following steps:
under the condition that the preset restoration model is a restoration network model after training, training an evaluation network model by using the image difference data of the distorted image and the restoration image output by the restoration network model and the average subjective score corresponding to the distorted image, so that the evaluation network model learns that the image difference data is mapped to the average subjective score in the training process;
or the like, or, alternatively,
and under the condition that the preset restoration model is a trained variational self-encoder, training an evaluation network model by using image difference data of the distorted image and the restoration image output by the variational self-encoder and an average subjective score corresponding to the distorted image, so that the evaluation network model learns that the image difference data is mapped to the average subjective score in the training.
7. The method of claim 1,
under the condition that the application scene is face recognition, the step of obtaining the distorted image to be evaluated from the database comprises the following steps: acquiring a target video from a monitoring video stream database, determining a target face image from the target video, and taking the target face image as the distorted image to be evaluated; or, determining a target face image from a monitoring image database, and taking the target face image as the distorted image to be evaluated;
after the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if so, using the distorted image to be evaluated for face recognition operation; if not, discarding the distorted image to be evaluated;
under the condition that the application scene is a video conference, the step of obtaining the distorted image to be evaluated from the database comprises the following steps: acquiring a target video from a conference video stream database, acquiring a target image from the target video, and taking the target image as the distorted image to be evaluated;
after the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if yes, determining that the current communication quality is good, and starting or continuing the video conference; if not, executing a dynamic adjustment strategy to improve the communication quality;
under the condition that the application scene is video on demand or live broadcast, the step of obtaining the distorted image to be evaluated from the database comprises the following steps: acquiring a target video from a video stream database, acquiring a target image from the target video, and taking the target image as the distorted image to be evaluated;
after the obtaining of the image quality evaluation result of the distorted image to be evaluated output by the preset evaluation network model, the method further includes:
judging whether the image quality evaluation result of the distorted image to be evaluated is larger than a preset threshold value or not; if yes, determining that the current communication quality is good, and starting or continuing video on demand or live broadcasting; if not, executing a dynamic adjustment strategy for improving the communication quality.
8. A no-reference image quality evaluation system, comprising:
the processing equipment is used for acquiring a distorted image to be evaluated from a local database or a database of third-party equipment, and restoring the distorted image to be evaluated by using a preset restoration model to acquire a restored image; inputting image difference data of the distorted image to be evaluated and the restored image into a preset evaluation network model to obtain an image quality evaluation result of the distorted image to be evaluated, which is output by the preset evaluation network model;
and the image quality evaluation result of the distorted image to be evaluated is used as a judgment basis for subsequent processing operation.
9. The system of claim 8,
the processing equipment is also used for using the image quality evaluation result of the distorted image to be evaluated as a judgment basis of subsequent processing operation; alternatively, the first and second electrodes may be,
the processing device is further configured to send an image quality evaluation result of the distorted image to be evaluated to the third-party device, so as to be used as a judgment basis for subsequent processing operation of the third-party device.
10. A training method of a reference-free image quality evaluation model is characterized by comprising the following steps:
initializing a restoration network model, judging a network model and evaluating the network model;
alternately training a restoration network model and a discrimination network model in an antagonistic mode, and obtaining a post-training restoration network model after a training end condition is reached;
and training the evaluation network model by using the image difference data of the distorted image and the restored image output by the restoration network model and the average subjective score corresponding to the distorted image, and obtaining the post-training evaluation network model after the training end condition is reached.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911023113.2A CN110782445A (en) | 2019-10-25 | 2019-10-25 | No-reference image quality evaluation method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911023113.2A CN110782445A (en) | 2019-10-25 | 2019-10-25 | No-reference image quality evaluation method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110782445A true CN110782445A (en) | 2020-02-11 |
Family
ID=69386455
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911023113.2A Pending CN110782445A (en) | 2019-10-25 | 2019-10-25 | No-reference image quality evaluation method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110782445A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111340140A (en) * | 2020-03-30 | 2020-06-26 | 北京金山云网络技术有限公司 | Image data set acquisition method and device, electronic equipment and storage medium |
| CN111640099A (en) * | 2020-05-29 | 2020-09-08 | 北京金山云网络技术有限公司 | Method and device for determining image quality, electronic equipment and storage medium |
| CN111784580A (en) * | 2020-06-30 | 2020-10-16 | 北京金山云网络技术有限公司 | Super-resolution method and device for image and server |
| CN112288627A (en) * | 2020-10-23 | 2021-01-29 | 武汉大学 | Recognition-oriented low-resolution face image super-resolution method |
| CN112614109A (en) * | 2020-12-24 | 2021-04-06 | 四川云从天府人工智能科技有限公司 | Image quality evaluation method, device and computer readable storage medium |
| CN113505854A (en) * | 2021-07-29 | 2021-10-15 | 济南博观智能科技有限公司 | Method, device, equipment and medium for constructing facial image quality evaluation model |
| CN113658091A (en) * | 2020-05-12 | 2021-11-16 | Tcl科技集团股份有限公司 | Image evaluation method, storage medium and terminal equipment |
| CN114913468A (en) * | 2022-06-16 | 2022-08-16 | 阿里巴巴（中国）有限公司 | Object repairing method, repair evaluating method, electronic device, and storage medium |
| CN115423809A (en) * | 2022-11-04 | 2022-12-02 | 江西电信信息产业有限公司 | Image quality evaluation method and device, readable storage medium and electronic equipment |
| CN115423725A (en) * | 2022-11-04 | 2022-12-02 | 北京数慧时空信息技术有限公司 | Remote sensing image restoration method and device |
| CN112614109B (en) * | 2020-12-24 | 2024-06-07 | 四川云从天府人工智能科技有限公司 | Image quality evaluation method, apparatus and computer readable storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108090902A (en) * | 2017-12-30 | 2018-05-29 | 中国传媒大学 | A kind of non-reference picture assessment method for encoding quality based on multiple dimensioned generation confrontation network |
| CN108648188A (en) * | 2018-05-15 | 2018-10-12 | 南京邮电大学 | A kind of non-reference picture quality appraisement method based on generation confrontation network |
| CN109118470A (en) * | 2018-06-26 | 2019-01-01 | 腾讯科技（深圳）有限公司 | A kind of image quality evaluating method, device, terminal and server |
| CN110120034A (en) * | 2019-04-16 | 2019-08-13 | 西安理工大学 | A kind of image quality evaluating method relevant to visual perception |
-
2019
- 2019-10-25 CN CN201911023113.2A patent/CN110782445A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108090902A (en) * | 2017-12-30 | 2018-05-29 | 中国传媒大学 | A kind of non-reference picture assessment method for encoding quality based on multiple dimensioned generation confrontation network |
| CN108648188A (en) * | 2018-05-15 | 2018-10-12 | 南京邮电大学 | A kind of non-reference picture quality appraisement method based on generation confrontation network |
| CN109118470A (en) * | 2018-06-26 | 2019-01-01 | 腾讯科技（深圳）有限公司 | A kind of image quality evaluating method, device, terminal and server |
| CN110120034A (en) * | 2019-04-16 | 2019-08-13 | 西安理工大学 | A kind of image quality evaluating method relevant to visual perception |
Non-Patent Citations (2)
| Title |
|---|
| KWAN-YEE LIN, ET AL.: "Hallucinated-IQA: No-Reference Image Quality Assessment via Adversarial Learning", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 叶雨琪 等: "基于多特征的彩色唐卡修复图像无参考质量评价方法", 《激光与光电子学进展》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111340140A (en) * | 2020-03-30 | 2020-06-26 | 北京金山云网络技术有限公司 | Image data set acquisition method and device, electronic equipment and storage medium |
| CN113658091A (en) * | 2020-05-12 | 2021-11-16 | Tcl科技集团股份有限公司 | Image evaluation method, storage medium and terminal equipment |
| CN111640099A (en) * | 2020-05-29 | 2020-09-08 | 北京金山云网络技术有限公司 | Method and device for determining image quality, electronic equipment and storage medium |
| CN111784580A (en) * | 2020-06-30 | 2020-10-16 | 北京金山云网络技术有限公司 | Super-resolution method and device for image and server |
| CN112288627A (en) * | 2020-10-23 | 2021-01-29 | 武汉大学 | Recognition-oriented low-resolution face image super-resolution method |
| CN112288627B (en) * | 2020-10-23 | 2022-07-05 | 武汉大学 | Recognition-oriented low-resolution face image super-resolution method |
| CN112614109A (en) * | 2020-12-24 | 2021-04-06 | 四川云从天府人工智能科技有限公司 | Image quality evaluation method, device and computer readable storage medium |
| CN112614109B (en) * | 2020-12-24 | 2024-06-07 | 四川云从天府人工智能科技有限公司 | Image quality evaluation method, apparatus and computer readable storage medium |
| CN113505854A (en) * | 2021-07-29 | 2021-10-15 | 济南博观智能科技有限公司 | Method, device, equipment and medium for constructing facial image quality evaluation model |
| CN113505854B (en) * | 2021-07-29 | 2023-08-15 | 济南博观智能科技有限公司 | Face image quality evaluation model construction method, device, equipment and medium |
| CN114913468A (en) * | 2022-06-16 | 2022-08-16 | 阿里巴巴（中国）有限公司 | Object repairing method, repair evaluating method, electronic device, and storage medium |
| CN115423809A (en) * | 2022-11-04 | 2022-12-02 | 江西电信信息产业有限公司 | Image quality evaluation method and device, readable storage medium and electronic equipment |
| CN115423725A (en) * | 2022-11-04 | 2022-12-02 | 北京数慧时空信息技术有限公司 | Remote sensing image restoration method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110782445A (en) | No-reference image quality evaluation method and system | |
| CN108428227B (en) | No-reference image quality evaluation method based on full convolution neural network | |
| CN111163338B (en) | Video definition evaluation model training method, video recommendation method and related device | |
| CN110210492B (en) | Stereo image visual saliency detection method based on deep learning | |
| CN113762107B (en) | Object state evaluation method, device, electronic equipment and readable storage medium | |
| Wang et al. | No-reference synthetic image quality assessment with convolutional neural network and local image saliency | |
| CN111047543A (en) | Image enhancement method, device and storage medium | |
| CN109685772B (en) | No-reference stereo image quality evaluation method based on registration distortion representation | |
| CN108875623B (en) | Face recognition method based on image feature fusion contrast technology | |
| CN109754390B (en) | No-reference image quality evaluation method based on mixed visual features | |
| CN114066812B (en) | No-reference image quality evaluation method based on spatial attention mechanism | |
| CN111179202A (en) | Single image defogging enhancement method and system based on generation countermeasure network | |
| Chen et al. | Subjective and objective quality evaluation of sonar images for underwater acoustic transmission | |
| CN111369548A (en) | No-reference video quality evaluation method and device based on generation countermeasure network | |
| CN108769543B (en) | Method and device for determining exposure time | |
| Bohr et al. | A no reference image blur detection using cumulative probability blur detection (cpbd) metric | |
| Gaata et al. | No-reference quality metric for watermarked images based on combining of objective metrics using neural network | |
| CN107657229B (en) | content classification-based video ambiguity detection human eye vision correction method | |
| Gaata et al. | No-reference quality metric based on fuzzy neural network for subjective image watermarking evaluation | |
| Li et al. | A novel spatial pooling strategy for image quality assessment | |
| CN111539263B (en) | Video face recognition method based on aggregation countermeasure network | |
| CN111310528B (en) | Image detection method, identity verification method, payment method and payment device | |
| Hong et al. | Saliency-based feature learning for no-reference image quality assessment | |
| CN110796106A (en) | Method for establishing portrait quality evaluation model and recognizing portrait from video | |
| CN112734733B (en) | Non-reference image quality monitoring method based on channel recombination and feature fusion |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200211 |
|
| RJ01 | Rejection of invention patent application after publication |