WO2021047453A1

WO2021047453A1 - Image quality determination method, apparatus and device

Info

Publication number: WO2021047453A1
Application number: PCT/CN2020/113559
Authority: WO
Inventors: 钮毅
Original assignee: 上海高德威智能交通系统有限公司
Priority date: 2019-09-11
Filing date: 2020-09-04
Publication date: 2021-03-18
Also published as: CN112488985A

Abstract

Provided are an image quality determination method, apparatus and device. The method comprises: acquiring a target image to be measured; and inputting the target image into a trained image quality model to obtain image quality parameters of the target image, wherein the image quality model is obtained through training according to a captured image and a standard image of each target object in an image set, and the image quality of the standard image meets a set requirement. Subjective influence on determination of image quality is avoided, and the accuracy of an image quality determination result is improved.

Description

Image quality determination method, device and equipment

Technical field

The present disclosure relates to the field of image processing technology, and in particular to a method, device and equipment for determining image quality.

Background technique

Surveillance equipment will produce a large number of images with different postures, occlusions, illuminations, sizes, etc. Some of them have good image quality and some have poor image quality. In the application of these images, such as target recognition, if a large number of poor quality images are used, it will not only increase the burden of hardware resources such as transmission and calculation, but also easily lead to misleading human eyes or intelligent algorithm recognition; Using images with high imaging quality for recognition, a smaller number of images will produce a higher matching rate, greatly alleviating the hardware resource occupation, thereby improving the target recognition rate. Therefore, it is necessary to filter images based on image quality.

In the existing image quality determination method, the direct learning mechanism is adopted. The imaging quality of all images in the training image set is manually evaluated and scored according to the recognition requirements, and the quality scores of all images in the training image set are obtained. Images and image quality scores train a model for calculating image quality scores.

In the above method, the quality of the images in the training image set is determined by manual scoring, and the quality index is not easy to quantify. Because the evaluator is sensitive to different attributes of the image, the scoring result is easily and also affected by the subjective feelings of the evaluator, so the quality of the image in the training image set The scoring is extremely noisy, resulting in low accuracy in the determination of image quality.

Summary of the invention

In view of this, the present disclosure provides an image quality determination method, device, and equipment to avoid subjective influence when determining the image quality, and improve the accuracy of the determination result of the image quality.

The first aspect of the present disclosure provides an image quality determination method, including: obtaining an image to be detected; inputting the image to be detected into a trained image quality model to obtain image quality parameters of the image to be detected; wherein, the The image quality model is obtained by training based on images in a first predetermined image set. The images in the first predetermined image set include respective captured images and standard images of multiple target objects, and the image quality of the standard images satisfies the set Claim.

According to an embodiment of the present disclosure, the image quality model is trained in the following manner: the captured image and the standard image of any target object in the first predetermined image set are input into the trained image feature model to determine the image feature The model extracts the first feature from the input captured image, and extracts the second feature from the standard image and outputs the first feature and the second feature; the input is determined based on the first feature and the second feature output by the image feature model The image quality parameters of the captured images; the image quality model is trained by using each captured image and the image quality parameters of the captured images.

According to an embodiment of the present disclosure, the image feature model is obtained by training based on images in a second predetermined image set, and the images in the second predetermined image set include respective captured images and standard images of multiple target objects; The image feature model is trained in the following manner: establishing a first initial model and a second initial model; inputting a captured image of any target object in a second predetermined image set into the first initial model, so that the first The initial model extracts image features from the input captured image and outputs it to the second initial model. The second initial model predicts the label information of the target object based on the input image features; the prediction based on the output of the second initial model The label information and the standard label information of the target object optimize the first initial model; check whether the set first training end condition is currently met, and if so, determine that the currently optimized first initial model is the image feature model.

According to an embodiment of the present disclosure, determining the image quality parameter of the input captured image based on the first feature and the second feature output by the image feature model includes: calculating the similarity between the first feature and the second feature ; Use the similarity to determine the image quality parameters of the captured image.

According to an embodiment of the present disclosure, the determining the image quality parameter of the captured image by using the similarity includes: determining the similarity as the image quality parameter of the captured image; or, from a preset similarity interval A target similarity interval containing the similarity is determined from the correspondence with the image quality parameter, and the image quality parameter corresponding to the target similarity interval is determined as the image quality parameter of the captured image.

According to an embodiment of the present disclosure, the image quality model is trained in the following manner: establishing a third initial model; inputting any captured image in the first predetermined image set to the third initial model, so that all The third initial model predicts and outputs image quality parameters based on the input captured images; optimizes the third initial model based on the image quality parameters output by the third initial model and the image quality parameters of the input captured images; If the second training end condition is set, it is determined that the currently optimized third initial model is the image quality model.

A second aspect of the present disclosure provides an image quality determination device, which includes: a to-be-detected image acquisition module for acquiring the to-be-detected image; an image quality determination module for inputting the to-be-detected image into a trained image quality model, Obtain the image quality parameters of the image to be detected; wherein, the image quality model is obtained by training based on images in a first predetermined image set, and the images in the first predetermined image set include respective snapshots of multiple target objects An image and a standard image, and the image quality of the standard image meets the set requirements.

According to an embodiment of the present disclosure, the image quality model is trained by the following modules: a feature extraction module for inputting captured images and standard images of any target object in the first predetermined image set into the trained image features The image feature model is used to extract the first feature from the input captured image and the second feature from the standard image to output the first feature and the second feature; the captured image quality determination module is used to determine the quality of the captured image based on the image. The first feature and the second feature output by the feature model determine the image quality parameters of the input captured image; the image quality model training module is used to train the image quality model by using each captured image and the image quality parameters of the captured image.

According to an embodiment of the present disclosure, the image feature model is obtained by training based on images in a second predetermined image set, and the images in the second predetermined image set include respective captured images and standard images of multiple target objects; The image feature model is trained through the following modules: a model building module, used to establish a first initial model and a second initial model; a label information prediction module, used to input a captured image of any target object in the second predetermined image set to The first initial model, so that the first initial model extracts image features from the input captured image and outputs the image features to the second initial model, and the second initial model predicts the target object based on the input image features Label information; a model optimization module for optimizing the first initial model based on the predicted label information output by the second initial model and standard label information of the target object; a model determination module for checking whether the current set is satisfied If the first training end condition is yes, it is determined that the currently optimized first initial model is the image feature model.

According to an embodiment of the present disclosure, when the captured image quality determination module determines the image quality parameters of the input captured image based on the first feature and the second feature output by the image feature model, it is specifically used to: The similarity between the feature and the second feature; the similarity is used to determine the image quality parameter of the captured image.

According to an embodiment of the present disclosure, when the captured image quality determination module uses similarity to determine the image quality parameter of the captured image, it is specifically configured to: determine the similarity as the image quality parameter of the captured image; or Determine the target similarity interval containing the similarity from the correspondence between the preset similarity interval and the image quality parameter, and determine the image quality parameter corresponding to the target similarity interval as the image quality of the captured image parameter.

According to an embodiment of the present disclosure, the image quality model training module trains the image quality model in the following manner: establish a third initial model; input any captured image in the first predetermined image set to the third initial Model, so that the third initial model predicts and outputs image quality parameters based on the input captured images; optimizes the third initial model based on the image quality parameters output by the third initial model and the input captured images ; Check whether the set second training end condition is currently met, and if so, determine that the currently optimized third initial model is the image quality model.

A third aspect of the present disclosure provides an electronic device, including a processor and a machine-readable storage medium; the storage medium stores a program that can be called by the processor; wherein, when the processor executes the program, it is prompted to: The image to be detected; input the image to be detected into the trained image quality model to obtain the image quality parameters of the image to be detected; wherein the image quality model is obtained by training based on images in the first predetermined image set The images in the first predetermined image set include respective captured images and standard images of multiple target objects, and the image quality of the standard images meets a set requirement.

A fourth aspect of the present disclosure provides a machine-readable storage medium having machine-executable instructions stored thereon. When the machine-executable instructions are executed by a processor, the processor is prompted to: acquire an image to be detected; The image is input to the trained image quality model to obtain the image quality parameters of the image to be detected; wherein, the image quality model is obtained by training based on the images in the first predetermined image set, in the first predetermined image set The image includes the captured images and standard images of multiple target objects, and the image quality of the standard images meets the set requirements.

In the embodiments of the present disclosure, an image quality model can be trained based on the captured images and standard images of each target object in the image collection, and the image to be tested can be input to the image quality model, and the image quality parameters of the image to be tested can be obtained. Through manual scoring or manual participation in the evaluation, it is avoided that the artificial subjective influence brings a lot of noise to the determination of image quality, and the accuracy of the determination result of image quality is improved.

Description of the drawings

FIG. 1 is a schematic flowchart of an image quality determination method according to an embodiment of the present disclosure.

Fig. 2 is a structural block diagram of an image quality determining apparatus according to an embodiment of the present disclosure.

FIG. 3 is a schematic flowchart of training an image quality model according to an embodiment of the present disclosure.

Fig. 4 is a schematic flowchart of training an image feature model according to an embodiment of the present disclosure.

Fig. 5 is a structural block diagram of an electronic device according to an embodiment of the present disclosure.

detailed description

The exemplary embodiments will be described in detail here, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings indicate the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

The terms used in the present disclosure are only for the purpose of describing specific embodiments, and are not intended to limit the present disclosure. The singular forms of "a", "said" and "the" used in the present disclosure and appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and/or" as used herein refers to and includes any or all possible combinations of one or more associated listed items.

It should be understood that although the terms first, second, third, etc. may be used in this disclosure to describe various devices, the information should not be limited to these terms. These terms are only used to distinguish devices of the same type from each other. For example, without departing from the scope of the present disclosure, the first device may also be referred to as the second device, and similarly, the second device may also be referred to as the first device. Depending on the context, the word "if" as used herein can be interpreted as "when" or "when" or "in response to determination".

In order to make the description of the present disclosure clearer and concise, some technical terms in the present disclosure are explained below.

Neural network: A technology abstracted by imitating the structure of the brain. This technology connects a large number of simple functions in a complex manner to form a network system that can fit extremely complex functional relationships, generally including convolution/inverse Convolution operations, activation operations, pooling operations, and operations such as addition, subtraction, multiplication, and division, channel merging, and element rearrangement. Use specific data to train the network and adjust the connections in it, so that the neural network can learn to fit the mapping relationship between input and output.

In the embodiments of the present disclosure, image quality refers to the imaging quality of the target object in the image. The factors that affect the image quality include the posture, occlusion, size, illumination, blurriness, etc. of the target object in the image. The higher the imaging quality, the more beneficial it is. Confirm the identity of the target. The target object can refer to a human face, a vehicle, a pedestrian, a license plate, and so on.

Common image quality indicators are as follows:

1) Attitude: Different target objects can use different posture information to express posture, for example, the face can express the posture by yaw angle, pitch angle and rotation angle, and pedestrian, vehicle, license plate can use the deflection angle and perspective information to express posture, etc.;

2) Occlusion: Generally divided into fixed occlusion and non-fixed occlusion. For example, the fixed occlusion of the face involves the occlusion of personal objects such as hats, sunglasses, and masks, and the non-fixed occlusion involves the occlusion of hands, other people or objects;

3) Size: the size of the target object. For example, a human face is generally measured by the distance between the pupils of the eyes, a pedestrian is measured by the height, a vehicle is measured by the distance between the left and right rearview mirrors, and the license plate can be measured by the height of the character;

4) Illumination: Appearance brightness of the target object is suitable and uniform, and there is no overexposure, too dark, unevenness, etc. that cause the details and textures to be unclear;

5) Fuzziness: The key texture edges of the target object are required to be clear, for example, the face requires clear edges of the facial features; pedestrians require clear edges on the limbs, and clear textures of clothing and carry-on items; vehicles require license plate characters, body, lights, and windows The edges are sharp.

In the direct learning mechanism, the evaluator usually needs to measure these image quality indicators, such as whether to choose "occluded front face" or "complete partial face", or "slightly blurred front face" or clear partial face". The indicators are not easy to quantify from a subjective point of view, and it is difficult to score accurately.

The embodiments of the present disclosure can avoid subjective influence when determining the image quality, and improve the accuracy of the determination result of the image quality.

The image quality determination method of the embodiment of the present disclosure will be described in more detail below, but it should not be limited to this. In an embodiment, referring to FIG. 1, an image quality determination method may include the following steps:

A100: Obtain the image to be detected;

A200: Input the image to be detected into the trained image quality model to obtain the image quality parameters of the image to be detected;

Wherein, the image quality model is obtained by training based on images in a first predetermined image set, and the images in the first predetermined image set include respective captured images and standard images of multiple target objects, and the image quality of the standard image Meet the set requirements.

In the embodiments of the present disclosure, the execution subject of the image quality determination method is the electronic device, more specifically the processor of the electronic device. The electronic device may be a device with imaging function such as a camera, or a device capable of image processing, such as a computer, and the specific type is not limited, as long as it has data processing capabilities.

The image quality determination method of the embodiments of the present disclosure can be applied to multiple application scenarios, such as access control systems, bayonet systems, electronic passport systems, public security systems, transportation systems, bank self-service systems, information security systems, etc., which require target recognition. Scene. Of course, the specific scene is not limited, as long as the image quality needs to be determined.

In step A100, an image to be detected is acquired.

The image to be detected is the image whose image quality needs to be determined. The image to be detected may be an image collected from a surveillance scene, and the surveillance scene may be a scene that requires target recognition. In the surveillance scene, the camera continuously collects images. Therefore, generally, multiple images of the same target object can be obtained to obtain an image sequence, and each image in the image sequence can be used as the image to be detected.

Of course, the image to be detected can be an image collected in real time from video surveillance, or an image obtained through background retrieval.

In step A200, the image to be detected is input to the trained image quality model to obtain the image quality parameters of the image to be detected.

The image quality model can be pre-trained and stored locally in the electronic device or in an external device, and recalled when needed. After the image quality model is trained, the image quality model can be used to determine the image quality parameters of the target image.

The target image is input to the trained image quality model, and the image quality parameter of the target image is calculated and output from the image quality model. The image quality parameter is used as the result of determining the image quality and can characterize the image quality of the target image. The image quality parameter can be represented by a single value, or can be represented by a vector containing multiple values, and the embodiment of the present application does not limit the specific expression form of the image quality parameter.

Continuing to take the above image sequence as an example, after the image quality model is used to determine the image quality parameters of each target image in the image sequence, N target images with the best image quality parameters can be selected from all target images, and the value of N Can be greater than or equal to 1. The selected target image is the image with the highest imaging quality, which can be used to identify the target object in subsequent processing, which can avoid the unstable recognition result due to the interference of a large number of low-quality images, and can also avoid occupying too much hardware resources.

The image quality model is trained based on the captured images and standard images of each target object in the image collection. The image collection contains the captured images and standard images of multiple target objects. The captured image and standard image of a target object correspond to each other and both contain the target object. However, the image quality may be different, such as the posture, occlusion, etc. of the target object. Conditions such as size, lighting and/or blurriness may vary.

The image quality of the standard image meets the set requirements. The setting requirements can be determined according to needs. For example, a standard image is an image that meets the corresponding quality indicators in terms of attitude, occlusion, size, illumination, and blur; another example, the definition of the standard image reaches the set definition.

The standard image may be, for example, a document image of the target object. When the target object is a human face, the document image is the face image used on ID cards, visas, etc., and is usually collected in a fixed mode (for example, under a single background color); when the target object is a license plate, the document image It is the license plate image used on the license and other documents; when the target object is a vehicle, the license image is the image of the vehicle used on the license and other documents.

Or, the standard image may be an image with the highest definition selected from the image library of the target object. The image library saves multiple images of the target object collected in the surveillance scene. The image with the highest definition can be selected according to a certain selection method as the standard image, or it can be manually selected.

Of course, the source of the standard image is not limited to the above method, and it can also be generated in other ways, such as generated based on a standard template according to a certain standard.

The captured image may be an image containing the target object captured in the surveillance scene. The captured image has uncertainty in the pose, occlusion, size, illumination, and blur degree of the target object.

There can be multiple captured images for each target object, and the specific number is not limited. Appropriate and rich captured images can be prepared for each target object. The "rich" here means that the different postures, illumination, scale, clarity and other factors of the target object should be taken into consideration. Preferably, all the captured images can cover the various degrees of each influencing factor. Taking the light influencing factor as an example, there are five levels of captured images of underexposed, dim, appropriate, bright, and overexposed. Only when the captured image covers enough rich imaging material can the stability of the image quality model for subsequent training be ensured.

Using the standard image of the target object as a reference, the quality of the captured image of the target object can be determined, so that the image quality parameters of each captured image of the target object can be determined. Furthermore, the captured image and the image quality parameters of the captured image can be used to train the image quality model.

In the embodiments of the present disclosure, an image quality model can be trained based on the captured images and standard images of each target object in the image collection, and the target image to be measured is input to the image quality model, and the image quality parameters of the target image can be obtained. There is no need for manual scoring or manual participation in the evaluation, avoiding the artificial subjective influence that brings a lot of noise to the determination of the image quality, and improving the accuracy of the determination result of the image quality.

In one embodiment, the above method flow can be executed by the image quality determining apparatus 100. As shown in FIG. 2, the image quality determining apparatus 100 mainly includes two modules: a target image acquisition module 101 and an image quality determination module 102. The target image acquisition module 101 is used to perform the above step A100, and the image quality determination module 102 is used to perform the above step A200.

In an embodiment, referring to Fig. 3, the image quality model is trained in the following manner:

S100: Input the captured image and standard image of any target object in the first predetermined image set to the trained image feature model, so that the image feature model extracts the first feature and the second feature from the input captured image and standard image. Feature and output; the image quality of the standard image meets the set requirements;

S200: Determine image quality parameters of the input captured image based on the first feature and the second feature output by the image feature model;

S300: Use each captured image and the image quality parameters of the captured image to train the image quality model.

In step S100, the captured image and standard image of any target object in the first predetermined image set are input to the trained image feature model, so that the image feature model extracts the first feature and the standard image from the input captured image and standard image respectively. The second feature is output; the image quality of the standard image meets the set requirements. The image feature model is a pre-trained model, which can be stored in an electronic device or in an external device, and can be called when step S100 is executed. The image feature model can be used to extract features from the image and output the corresponding features. The features here can be represented by data in formats such as feature vectors.

When training the image feature model, the captured images of each target object can be used as a training sample set, and the global information of the target object in the captured image, or the local information of the target object, or both, can be used for training. How to train the image feature model is not a limitation. Of course, other images can also be used to train the image feature model, as long as the image feature model for feature extraction can be trained.

The captured image and standard image of each target object in the first predetermined image set can be input into the image feature model, and the input captured image can be extracted by the image feature model to obtain the first feature, and the input standard image can be characterized The second feature is extracted.

In step S200, the image quality parameter of the input captured image is determined based on the first feature and the second feature output by the image feature model. The first feature output by the image feature model is the first feature of the input captured image, and the second feature is the second feature of the standard image containing the target object corresponding to the input captured image.

The first feature can reflect the imaging condition of the target object in the captured image, and the second feature can reflect the imaging condition of the target object in the standard image.

Since the standard image is an image whose image quality meets the set requirements, it is regarded as a high-quality standard. The more similar the first feature of the captured image and the second feature of the standard image are, the more similar the captured image is to the standard image, the more it meets the high-quality standard, and the better the image quality. Therefore, the image quality parameter of the captured image can be determined based on the first feature and the second feature.

In this way, the image quality parameters of all captured images can be determined. The image quality parameter can characterize the image quality of the captured image. For example, the higher the image quality parameter, the higher the image quality of the captured image.

In step S300, the image quality model is trained by using each captured image and the image quality parameters of the captured image.

It is possible to use part or all of the captured images in the first predetermined image set as the training sample set, and use the image quality parameters of the captured images as the supervision information to realize the training of the image quality model. The specific training method is not limited, as long as the image quality model for determining the image quality can be trained by using each captured image and the image quality parameters of the captured image.

In one embodiment, referring to Fig. 4, the image feature model is trained in the following ways, including:

T100: Establish the first initial model and the second initial model;

T200: Input a captured image of any target object in the second predetermined image set to the first initial model, so that the first initial model extracts image features from the input captured image and outputs it to the second initial model Model, the second initial model predicts the label information of the target object based on the input image features;

T300: Optimize the first initial model based on the predicted label information output by the second initial model and the standard label information of the target object;

T400: Check whether the set first training end condition is currently met, and if so, determine that the currently optimized first initial model is the image feature model.

In step T100, a first initial model and a second initial model are established.

The first initial model and the second initial model can be cascaded together to form an end-to-end whole for joint training. The first initial model and the second initial model can be constructed using a CNN model, and the specific layer structure is not limited.

The first initial model can be a feature extractor, and the second initial model can be a classifier. Of course, the second initial model can be determined according to the task. For example, in a target classification task, the second initial model can be a classifier; in a character recognition task, the second initial model can be a decoder.

In step T200, the captured image of any target object in the second predetermined image set is input to the first initial model, so that the first initial model extracts image features from the input captured image and outputs the captured image to the first initial model. Two initial models, the second initial model predicts label information based on the input image features. The label information represents the prediction result of the target object, and can be numbers, characters, pictures, etc., and this embodiment does not limit the expression form of the label information.

The second predetermined image set is similar to the first predetermined image set, and includes respective captured images and standard images of multiple target objects. The first predetermined image set used to train the image quality model and the second predetermined image set used to train the image feature model may be the same or different.

The first initial model is responsible for extracting image features from captured images, and the second initial model is responsible for predicting label information based on image features. Through the cooperation of the first initial model and the second initial model, the prediction of the label information of each captured image in the image set is completed .

When the model parameters of the first initial model and the second initial model change, the prediction results may also change. Training is to change the model parameters to make the prediction results more accurate and closer to the required results.

In step T300, the first initial model is optimized based on the predicted label information output by the second initial model and the standard label information of the target object.

The standard label information of each target object can be calibrated in advance. Whenever the second initial model outputs a label information, the output label information is combined with the acquired target object (input target object in the captured image of the first initial model). The standard label information of) is compared, and the first initial model is optimized according to the comparison result, so as to reduce the difference between the subsequent predicted label information and the standard label information. Of course, while optimizing the first initial model, the second initial model can also be optimized.

As the first initial model and the second initial model are continuously optimized and the model parameters are changed, the predicted label information will gradually approach the standard label information of the target object in the input captured image. Through training, the cascaded first initial model and second initial model learn the mapping relationship between the captured images of each target object and the predicted label information of the target object.

The standard label information may be a standard image. Then, the first initial model extracts image features from the input captured image, and the second initial model calculates a predicted label based on the extracted image features; compares the predicted label with the standard image of the target object in the input captured image, and optimizes the first image based on the comparison result. An initial model and a second initial model to reduce the difference between the subsequent prediction results and the standard image.

The standard label information may also be label information representing the category of the target object, which is called "standard category label information". A label can correspond to a category in the classification task. For example, in the task of face recognition, the label can represent a specific person (person A, person B, person C, etc.); for example, in vehicle brand recognition, the label represents the manufacturer and its brand (Toyota Camry, Volkswagen, etc.) Passat, BYD Tang, etc.).

Then, the first initial model extracts image features from the input captured image, and the second initial model calculates the label information representing the predicted category based on the extracted image features, which is called "predicted category label information"; compare the predicted category label information with the input According to the standard category label information of the target object in the captured image, the first initial model and the second initial model are optimized according to the comparison result to reduce the difference between the predicted category label information and the standard category label information of subsequent predictions.

In step T400, it is checked whether the set first training end condition is currently met, and if so, it is determined that the currently optimized first initial model is the image feature model.

The first training end conditions can be multiple, for example, the number of training times can reach a specified number; or, the performance of the first initial model and the second initial model reach the set index; or, there is no image set that has not been input to the first A snapshot of the initial model, etc.

When the first training end condition is currently met, it is determined that the currently optimized first initial model is the image feature model, and the second initial model may no longer be used, otherwise, the training may continue.

Optionally, if the first training end condition is not currently met, continue the operation of selecting a captured image of the target object that is not input to the first initial model from the image set and inputting it to the first initial model. Through the above cycle, the required image feature model is obtained. Of course, the specific training method is not limited to this.

In this embodiment, the image feature model is trained by using the captured images of each target object, so that the image feature model can extract the first feature of each captured image more accurately than models trained on other images. And, if the image quality parameters of the captured images of the target objects are needed for training the image quality model, the captured images of the corresponding target objects can be used to train the image feature model, which is not limited by the scene and object type, and guarantees the quality of the trained images The stability and generalization ability of the model in various complex scenarios.

In one embodiment, in step S200, determining the image quality parameter of the input captured image based on the first feature and the second feature output by the image feature model includes:

S201: Calculate the similarity between the first feature and the second feature;

S202: Determine the image quality parameter of the captured image by using the similarity.

Any one of the SIFT algorithm, the SURF algorithm, the histogram matching algorithm, the mean hash algorithm, the Euclidean distance algorithm, and the cosine distance algorithm may be used to calculate the similarity between the first feature and the second feature.

For example, taking the cosine distance algorithm as an example, the first feature and the second feature are expressed in vector form, and the cosine value of the vector angle between the first feature and the second feature can be calculated, and the calculated cosine value is used as the first feature Similarity with the second feature.

After calculating the similarity between the first feature and the second feature, the similarity between the captured image and the standard image of the target object in the captured image can be determined. The higher the similarity, the higher the similarity between the captured image and the standard image. The more similar the images are, the less similar they are on the contrary. Therefore, taking the standard image as the high quality standard, the image quality parameters of the captured image can be determined according to the similarity, and the specific determination method is not limited.

In one embodiment, in step S202, the determining the image quality parameter of the captured image by using the similarity includes: determining the similarity as the image quality parameter of the captured image. In other words, the calculated similarity between the first feature and the second feature can be directly used as the image quality parameter of the captured image.

Alternatively, the target similarity interval containing the similarity is determined from the corresponding relationship between the preset similarity interval and the image quality parameter, and the image quality parameter corresponding to the target similarity interval is determined as the image of the captured image Quality parameters.

The similarity value range covering all possible similarity values can be divided into several similarity intervals in advance, and different similarity intervals correspond to different image quality parameters. The corresponding relationship between the preset similarity interval and the image quality parameter can be recorded in the electronic device, for example, can be saved in the form of a table.

After the similarity is calculated, the target similarity interval in which the similarity is located can be determined from the corresponding relationship between the preset similarity interval and the image quality parameter, and the image quality parameter corresponding to the target similarity interval is determined It is the image quality parameter of the captured image.

For example, the similarity value range is 0-100, and the range is divided into three similarity intervals 0-30, 31-60, 61-100, and the corresponding image quality parameters are numerical values representing low, medium, and high, respectively. The specific value is not limited. When the calculated similarity is 80, the target similarity interval is 61-100, the image quality parameter corresponding to the target similarity interval is a high value, and the high value is used as the image quality parameter of the captured image. Of course, the value range and the division granularity here are all examples, and are not specifically limited to this. For example, the division granularity can be finer.

By dividing the similarity value range into multiple files, the similarity is classified into one file each time, and the image quality parameters of the normalized file are used as the image quality parameters of the captured image. In this way, the output of the image quality model also includes These blocks of image quality parameters reduce the data processing complexity of the image quality model.

In this embodiment, the quantification of image quality parameters is realized by using the similarity between the first feature and the second feature, which is simpler than manual scoring and does not involve subjective factors.

In one embodiment, in step S300, the image quality model is trained by using each captured image and the image quality parameters of the captured image, including:

S301: Establish a third initial model;

S302: Input any captured image in the first predetermined image set to the third initial model, so that the third initial model predicts and outputs image quality parameters based on the input captured image;

S303: Optimizing the third initial model based on the image quality parameters output by the third initial model and the image quality parameters of the captured image input;

S304: Check whether the set second training end condition is currently met, and if so, determine that the currently optimized third initial model is the image quality model.

The third initial model is a predictor, which can be composed of any model that can implement regression tasks, such as logistic regression model, tree model, neural network model, etc., and the specifics are not limited.

In step S302, any captured image in the first predetermined image set is input to the third initial model, so that the third initial model predicts and outputs image quality parameters based on the input captured image.

The third initial model is responsible for predicting the image quality parameters of the captured image. The model parameters of the third initial model change, and the prediction results may also change. Training is to change the model parameters to make the prediction results more accurate and closer to the required results.

In step S303, the third initial model is optimized based on the image quality parameters output by the third initial model and the image quality parameters of the captured image input.

Whenever the third initial model outputs an image quality parameter, the output image quality parameter is compared with the image quality parameter of the input captured image. Here, the image quality parameter of the captured image input used for comparison refers to the known actual image quality parameter of the captured image. The third initial model is optimized according to the comparison result to reduce the difference between the subsequently predicted image quality parameter and the image quality parameter of the captured image.

As the third initial model is continuously optimized and the model parameters are changed, the predicted image quality parameters will gradually approach the image quality parameters of the input captured image. Through training, the third initial model learns the mapping relationship between each captured image and the image quality parameter of the captured image.

In step S304, it is checked whether the set second training end condition is currently met, and if so, it is determined that the currently optimized third initial model is the image quality model.

The second training end conditions can be multiple, for example, it can be that the number of training times reaches the specified number; or, the performance of the third initial model reaches the set index; or, there are no captured images in the image set that have not been input to the third initial model ,and many more.

When the second training end condition is currently met, it is determined that the currently optimized third initial model is the image quality model; otherwise, training can be continued.

Optionally, if the second training end condition is not currently met, continue the operation of selecting a captured image from the image set that has not been input to the third initial model and inputting it to the third initial model. Through the above cycle, the required image quality model can be obtained. Of course, the specific training method is not limited to this.

The present disclosure also provides an image quality determination device. Referring to FIG. 2, the image quality determination device 100 includes: a to-be-detected image acquisition module 101 for acquiring a to-be-detected image; an image quality determination module 102 for combining the to-be-detected image Input to the trained image quality model to obtain the image quality parameters of the image to be detected; wherein, the image quality model is obtained by training according to the images in the first predetermined image set, and the image quality parameters in the first predetermined image set The image includes respective captured images and standard images of multiple target objects, and the image quality of the standard images meets the set requirements.

In one embodiment, the image quality model is trained by the following modules: a feature extraction module, which is used to input the captured image and standard image of any target object in the first predetermined image set into the trained image feature model for The image feature model extracts the first feature from the input captured image and the second feature from the standard image and outputs the first feature and the second feature; the captured image quality determination module is used for the output based on the image feature model The first feature and the second feature determine the image quality parameters of the input captured image; the image quality model training module is used to train the image quality model by using each captured image and the image quality parameters of the captured image.

In one embodiment, the image feature model is obtained by training based on images in a second predetermined image set, and the images in the second predetermined image set include respective captured images and standard images of multiple target objects; The image feature model is trained through the following modules: a model building module, used to establish a first initial model and a second initial model; a label information prediction module, used to input a captured image of any target object in the second predetermined image set to the A first initial model, so that the first initial model extracts image features from the input captured image and outputs to the second initial model, and the second initial model predicts the label information of the target object based on the input image features Model optimization module, used to optimize the first initial model based on the predicted label information output by the second initial model and standard label information of the target object; model determination module, used to check whether the current set first Training end condition, if yes, it is determined that the currently optimized first initial model is the image feature model.

In one embodiment, when the captured image quality determination module determines the image quality parameters of the input captured image based on the first feature and the second feature output by the image feature model, it is specifically used to: calculate the first feature and the second feature. The similarity between the second features; the similarity is used to determine the image quality parameter of the captured image.

In one embodiment, when the captured image quality determination module uses similarity to determine the image quality parameter of the captured image, it is specifically configured to: determine the similarity as the image quality parameter of the captured image; or, from A target similarity interval containing the similarity is determined from the correspondence between the preset similarity interval and the image quality parameter, and the image quality parameter corresponding to the target similarity interval is determined as the image quality parameter of the captured image.

In one embodiment, the image quality model training module trains the image quality model in the following manner: establish a third initial model; input any captured image in the first predetermined image set to the third initial model, so that the The third initial model predicts and outputs image quality parameters based on the input captured image; optimizes the third initial model based on the image quality parameters output by the third initial model and the image quality parameters of the input captured image; If the second training end condition is determined, it is determined that the currently optimized third initial model is the image quality model.

For the implementation process of the functions and roles of each unit in the above-mentioned device, refer to the implementation process of the corresponding steps in the above-mentioned method for details, which will not be repeated here.

For the device embodiment, since it basically corresponds to the method embodiment, the relevant part can refer to the part of the description of the method embodiment. The device embodiments described above are merely illustrative, where the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units.

The present disclosure also provides an electronic device, including a processor and a machine-readable storage medium; the machine-readable storage medium stores a program that can be called by the processor; wherein, when the processor executes the program, the following The image quality determination method described in the foregoing embodiment.

The embodiments of the image quality determining apparatus of the present disclosure can be applied to electronic equipment. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory through the processor of the electronic device where it is located. From a hardware perspective, as shown in FIG. 5, FIG. 5 is a hardware structure diagram of an electronic device where the image quality determining apparatus 100 is shown according to an exemplary embodiment of the present disclosure, except for the processor 510, In addition to the memory 530, the network interface 520, and the non-volatile storage medium 540, the electronic device in which the apparatus 100 is located in the embodiment generally may include other hardware according to the actual function of the electronic device, which will not be repeated.

The present disclosure also provides a machine-readable storage medium with a program stored thereon, and when the program is executed by a processor, the method for determining image quality as described in any one of the foregoing embodiments is implemented.

The present disclosure may take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program codes. Machine-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be realized by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of machine-readable storage media include, but are not limited to: phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only Memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disc (DVD) or other optical storage , Magnetic cassette tape, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by computing devices.

The above are only the preferred embodiments of the present disclosure and are not intended to limit the present disclosure. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the present disclosure. Within the scope of protection.

Claims

An image quality determination method, characterized in that it comprises:

Obtain the image to be detected;

Input the image to be detected into a trained image quality model to obtain image quality parameters of the image to be detected;

Wherein, the image quality model is obtained by training based on images in a first predetermined image set, and the images in the first predetermined image set include respective captured images and standard images of multiple target objects, and the images of the standard images The quality meets the set requirements.
8. The image quality determination method according to claim 1, wherein the image quality model is trained in the following manner:

Input the captured image and standard image of any target object in the first predetermined image set to the trained image feature model, so that the image feature model extracts the first feature from the input captured image and the standard image. Extract the second feature and output the first feature and the second feature;

Determining the image quality parameters of the input captured image based on the first feature and the second feature output by the image feature model;

The image quality model is trained by using each captured image and the image quality parameters of the captured image.
The image quality determination method according to claim 2, wherein the image feature model is obtained by training based on images in a second predetermined image set, and the images in the second predetermined image set include multiple target objects Respective captured images and standard images;

The image feature model is trained in the following manner:

Establish the first initial model and the second initial model;

Inputting a captured image of any target object in the second predetermined image set to the first initial model, so that the first initial model extracts image features from the input captured image and output to the second initial model, The second initial model predicts the label information of the target object based on the input image features;

Optimizing the first initial model based on the predicted label information output by the second initial model and the standard label information of the target object;

It is checked whether the set first training end condition is currently met, and if so, it is determined that the currently optimized first initial model is the image feature model.
3. The image quality determination method of claim 2, wherein determining the image quality parameters of the input captured image based on the first feature and the second feature output by the image feature model comprises:

Calculating the similarity between the first feature and the second feature;

The image quality parameter of the captured image is determined by using the similarity.
5. The image quality determination method according to claim 4, wherein said determining the image quality parameter of the captured image by using the similarity comprises:

Determining the similarity as the image quality parameter of the captured image; or,

Determine the target similarity interval containing the similarity from the corresponding relationship between the preset similarity interval and the image quality parameter, and determine the image quality parameter corresponding to the target similarity interval as the image quality parameter of the captured image .
8. The image quality determination method according to claim 1, wherein the image quality model is trained in the following manner:

Establish the third initial model;

Inputting any captured image in the first predetermined image set to the third initial model, so that the third initial model predicts and outputs image quality parameters based on the input captured image;

Optimizing the third initial model based on the image quality parameters output by the third initial model and the image quality parameters of the captured image input;

It is checked whether the set second training end condition is currently met, and if so, it is determined that the currently optimized third initial model is the image quality model.
An image quality determining device, characterized in that it comprises:

The image acquisition module to be detected is used to acquire the image to be detected;

An image quality determination module, configured to input the image to be detected into a trained image quality model to obtain image quality parameters of the image to be detected;

Wherein, the image quality model is obtained by training based on images in a first predetermined image set, and the images in the first predetermined image set include respective captured images and standard images of multiple target objects, and the images of the standard images The quality meets the set requirements.
8. The image quality determining device according to claim 7, wherein the image quality model is trained through the following modules:

The feature extraction module is used to input the captured image and standard image of any target object in the first predetermined image set into the trained image feature model, so that the image feature model extracts the first feature from the input captured image respectively , And extracting the second feature from the standard image and outputting the first feature and the second feature;

The captured image quality determination module is configured to determine the image quality parameters of the input captured image based on the first feature and the second feature output by the image feature model;

The image quality model training module is used to train the image quality model by using each captured image and the image quality parameters of the captured image.
8. The image quality determining device according to claim 8, wherein the image feature model is obtained by training based on images in a second predetermined image set, and the images in the second predetermined image set include a plurality of target objects Respective captured images and standard images;

The image feature model is trained through the following modules:

The model establishment module is used to establish the first initial model and the second initial model;

The label information prediction module is used to input a captured image of any target object in the second predetermined image set to the first initial model, so that the first initial model extracts image features from the input captured image and outputs it to The second initial model, which predicts the label information of the target object based on the input image features;

A model optimization module, configured to optimize the first initial model based on the predicted label information output by the second initial model and the standard label information of the target object;

The model determination module is used to check whether the set first training end condition is currently met, and if so, determine that the currently optimized first initial model is the image feature model.
8. The image quality determining device according to claim 8, wherein when the captured image quality determining module determines the image quality parameters of the input captured image based on the first feature and the second feature output by the image feature model, the specific Used for:

Calculating the similarity between the first feature and the second feature;

The image quality parameter of the captured image is determined by using the similarity.
10. The image quality determining device according to claim 10, wherein when the captured image quality determining module uses the similarity to determine the image quality parameters of the captured image, it is specifically configured to:

Determining the similarity as the image quality parameter of the captured image; or,

Determine the target similarity interval containing the similarity from the corresponding relationship between the preset similarity interval and the image quality parameter, and determine the image quality parameter corresponding to the target similarity interval as the image quality parameter of the captured image .
8. The image quality determining device according to claim 8, wherein the image quality model training module trains the image quality model in the following manner:

Establish the third initial model;

Inputting any captured image in the first predetermined image set to the third initial model, so that the third initial model predicts and outputs image quality parameters based on the input captured image;

Optimizing the third initial model based on the image quality parameters output by the third initial model and the image quality parameters of the captured image input;

It is checked whether the set second training end condition is currently met, and if so, it is determined that the currently optimized third initial model is the image quality model.
An electronic device including a processor and a machine-readable storage medium; the storage medium stores a program that can be called by the processor; wherein, when the processor executes the program, it is prompted:

Obtain the image to be detected;

Input the image to be detected into a trained image quality model to obtain image quality parameters of the image to be detected;

Wherein, the image quality model is obtained by training based on images in a first predetermined image set, and the images in the first predetermined image set include respective captured images and standard images of multiple target objects, and the images of the standard images The quality meets the set requirements.
A machine-readable storage medium having machine-executable instructions stored thereon, and the machine-executable instructions are executed by a processor to cause the processor to:

Obtain the image to be detected;

Input the image to be detected into a trained image quality model to obtain image quality parameters of the image to be detected;

Wherein, the image quality model is obtained by training based on images in a first predetermined image set, and the images in the first predetermined image set include respective captured images and standard images of multiple target objects, and the images of the standard images The quality meets the set requirements.