WO2022116104A1 - 图像处理方法、装置、设备及存储介质 - Google Patents

图像处理方法、装置、设备及存储介质 Download PDF

Info

Publication number
WO2022116104A1
WO2022116104A1 PCT/CN2020/133674 CN2020133674W WO2022116104A1 WO 2022116104 A1 WO2022116104 A1 WO 2022116104A1 CN 2020133674 W CN2020133674 W CN 2020133674W WO 2022116104 A1 WO2022116104 A1 WO 2022116104A1
Authority
WO
WIPO (PCT)
Prior art keywords
weight
image
target object
category
detected
Prior art date
Application number
PCT/CN2020/133674
Other languages
English (en)
French (fr)
Inventor
罗达新
林永兵
赵胜男
马莎
Original Assignee
华为技术有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to CN202080004518.6A priority Critical patent/CN112703532B/zh
Priority to PCT/CN2020/133674 priority patent/WO2022116104A1/zh
Publication of WO2022116104A1 publication Critical patent/WO2022116104A1/zh

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
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion

Definitions

  • the present application relates to the technical field of artificial intelligence, and in particular, to an image processing method, apparatus, device, and storage medium.
  • the difference or distance between the input picture and the reference picture is usually calculated, that is, the full reference image quality assessment (FR-IQA).
  • the reference picture is generally an uncompressed original picture. The larger the distance, the worse the quality of the input picture.
  • the content of the calculation is irrelevant to the task, important information in the picture is discarded, and the accuracy of the picture quality evaluation is poor.
  • Embodiments of the present application provide an image processing method, apparatus, device, and storage medium, which can improve the accuracy of image quality assessment.
  • the technical solution is as follows:
  • an image processing method includes:
  • the category weight of the target object based on the category information of the target object, where the category weight is used to reflect the relative importance of the category to which the target object belongs in the image to be detected;
  • the comprehensive weight of the target object is obtained
  • An image quality score of the image to be detected is determined based on the comprehensive weight of each object in the image to be detected.
  • the position weight reflecting the importance of the position of the target object is obtained, and then the category weight reflecting the importance of the target object category is obtained by identifying the category of the target object, and the position of the target object is further determined.
  • the weight is combined with the category weight to obtain a comprehensive weight reflecting the comprehensive importance of the target object.
  • obtaining the position weight of the target object based on the position information of the target object in the to-be-detected image includes:
  • position information of the target object in the image to be detected the position information including the position of the target object in the physical space described by the image to be detected
  • the position weight of the target object is obtained, and the position weight relationship reflects the mathematical correspondence between the position and the weight.
  • the target object includes target pixels
  • the acquiring position information of the target object in the to-be-detected image includes:
  • image transformation processing is performed on the front view or the side view to obtain a top view corresponding to the front view or the side view;
  • the position coordinates of the target pixel in the top view in the front view or the side view are acquired.
  • the obtaining the position weight of the target object according to the position weight relationship and the position information of the target object includes:
  • the position weight of the target pixel is obtained, and the position weight curve reflects the pixel position by describing the mathematical relationship between the pixel position and the weight. relative importance.
  • the obtaining the category weight of the target object based on the category information of the target object includes:
  • the category weight of the target object is obtained, and the category weight relationship reflects the mathematical correspondence between categories and weights.
  • the identifying the category information of the target object in the image to be detected includes:
  • the obtaining the category weight of the target object according to the category weight relationship and the category information of the target object includes:
  • the category weight data structure is searched to obtain the category weight of the target object.
  • the method further includes:
  • the state weight of the target object In the case that the motion state of the entity meets the conditions, determine the state weight of the target object, where the state weight is used to represent the relative importance of the target object in the to-be-detected image under different motion states;
  • the state category weight of the target object is obtained, and the state category weight is used to represent the relative importance of different categories of target objects in different motion states degree;
  • the comprehensive weight of the target object is obtained according to the position weight and the status category weight of the target object;
  • the step of obtaining the comprehensive weight of the target object according to the position weight and category weight of the target object is performed.
  • the importance of the target object can be more accurately reflected by the weight, and the weight not only reflects the position information and category information of the target object, but also reflects the motion state of the target object.
  • the comprehensive weight of the target object is obtained according to the position weight and category weight of the target object, including:
  • a weighted summation process is performed on the position weight and the category weight to obtain the comprehensive weight of the target pixel.
  • the calculation method of the comprehensive weight is made more scientific, so that the importance of the target pixel can be more objectively reflected.
  • the target object includes target pixels, and the to-be-detected image is a compressed image
  • the determining the image quality score of the image to be detected based on the combined weight of each object in the image to be detected includes:
  • the reference image refers to the uncompressed original image of the image to be detected
  • the weighted similarity between the image to be detected and the reference image is obtained based on the summed weight of each pixel and the pixel difference value on each pixel, where the weighted similarity refers to the similarity based on the summed
  • the peak signal-to-noise ratio and/or the structural similarity calculated by the weight, the peak signal-to-noise ratio and/or the structural similarity are used to reflect the degree of difference between the image to be detected and the reference image;
  • an image quality score of the image to be detected is obtained.
  • the image quality can be more accurately evaluated, and the weight index can be used reasonably.
  • the method further includes:
  • the image processing algorithm is adjusted.
  • the image quality can be used to improve the target effect desired by the image processing algorithm, and the image can be processed without losing important information in the image as much as possible.
  • the target object includes any one of the following: target pixel, target image block, and target entity.
  • an image processing apparatus includes:
  • an image acquisition module to be tested used to acquire the image to be tested
  • a position weight determination module configured to obtain the position weight of the target object based on the position information of the target object in the to-be-detected image, where the position weight is used to reflect the position occupied by the target object in the to-be-detected image the relative importance of
  • a category weight determination module configured to obtain a category weight of the target object based on category information of the target object, where the category weight is used to reflect the relative importance of the category to which the target object belongs in the image to be detected;
  • a comprehensive weight determination module configured to obtain the comprehensive weight of the target object according to the position weight and category weight of the target object
  • An image quality detection module configured to determine an image quality score of the to-be-detected image based on the comprehensive weight of each object in the to-be-detected image.
  • a computer device includes a processor and a memory, the memory stores a computer program, and the computer program is loaded and executed by the processor to realize the above image processing method.
  • a computer-readable storage medium where a computer program is stored in the storage medium, and the computer program is loaded and executed by a processor to implement the above image processing method.
  • a computer program product or computer program where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above-mentioned image processing method.
  • the position weight reflecting the importance of the position of the target object is obtained, and then the category weight reflecting the importance of the target object category is obtained by identifying the category of the target object, and the position of the target object is further determined.
  • the weight is combined with the category weight to obtain a comprehensive weight reflecting the comprehensive importance of the target object.
  • the embodiment of the present application designs a position weight model, which can accurately calculate the position weight of each pixel in the image; and through the semantic segmentation model, the image to be detected is divided into different regions according to categories, and the category of each region is calculated.
  • the weights ensure the accuracy of determining the category weights; finally, the position weights and category weights are integrated, and the final image quality scores are calculated by combining traditional PSNR/SSIM methods, which further ensures the reliability of the obtained image quality scores.
  • FIG. 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application.
  • FIG. 2 is a flowchart of an image processing method provided by an embodiment of the present application.
  • Figure 3 Figure 4, Figure 5, Figure 6 exemplarily show schematic diagrams of several images to be detected
  • FIG. 7 exemplarily shows a schematic diagram reflecting the category-weight relationship
  • FIG. 8 exemplarily shows a schematic diagram of a driving image
  • FIG. 9 is a flowchart of an image processing method provided by another embodiment of the present application.
  • FIG. 10 exemplarily shows a schematic diagram of an image processing flow provided by an embodiment of the present application.
  • FIG. 11 is a flowchart of an image processing method provided by an embodiment of the present application.
  • FIG. 12 exemplarily shows a schematic diagram of a front view and a top view in an automatic driving task
  • FIG. 13 is a block diagram of an image processing apparatus provided by an embodiment of the present application.
  • FIG. 14 is a block diagram of an image processing apparatus provided by another embodiment of the present application.
  • FIG. 15 is a structural block diagram of a computer device provided by an embodiment of the present application.
  • Image Quality Assessment is one of the basic technologies in image processing. It mainly analyzes and studies the characteristics of the image, and then evaluates the quality of the image (the degree of image distortion). Image quality evaluation plays an important role in algorithm analysis and comparison and system performance evaluation in image processing systems.
  • the objective evaluation of image quality can be divided into three types: full reference (Full-Reference, FR), partial reference (Reduced-Reference, RR) and no reference (No-Reference, NR).
  • Full Reference Image Quality Assessment refers to comparing the difference between the image to be evaluated and the reference image and analyzing the distortion degree of the image to be evaluated when an ideal image is selected as the reference image. Thereby, the quality evaluation of the image to be evaluated is obtained.
  • the commonly used objective evaluation of full reference image quality is mainly based on three aspects: pixel statistics, information theory, and structural information.
  • PSNR Peak-Signal to Noise Ratio
  • MSE Mean Square Error
  • the main function of human vision is to extract the structural information in the background, and the human visual system can achieve this goal highly adaptively, so the measurement of the structural distortion of the image should be the best approximation of the image perceptual quality.
  • SSIM Structure Similarity
  • the structural similarity index defines the structural information as independent of brightness and contrast from the perspective of image composition, reflects the properties of the object structure in the scene, and models the distortion as brightness, contrast and structure. a combination of different factors.
  • the mean is used as an estimate of brightness, the standard deviation is used as an estimate of contrast, and the covariance is used as a measure of structural similarity. Its calculation method is as follows. where x, y represent the two images, ⁇ and ⁇ are the mean and variance, respectively, and c is a constant.
  • No reference image quality assessment (Non Reference Image Quality Assessment, NR-IQA), the no reference method is also called the first evaluation method, because the general ideal image is difficult to obtain, so this kind of quality evaluation method that is completely out of the dependence on the ideal reference image
  • No-reference methods are generally based on image statistical properties.
  • Mean refers to the average value of image pixels, which reflects the average brightness of the image. The higher the average brightness, the better the image quality.
  • Standard deviation refers to the degree of dispersion of image pixel gray values relative to the mean. The larger the standard deviation, the more dispersed the gray levels in the image, and the better the image quality.
  • the average gradient can reflect the detail contrast and texture transformation in the image, and it reflects the clarity of the image to a certain extent.
  • Entropy refers to the average amount of information of an image. It measures the amount of information in an image from the perspective of information theory. The greater the information entropy in an image, the more information the image contains.
  • the no-reference image quality evaluation method first makes a certain assumption about the characteristics of the ideal image, then establishes a corresponding mathematical analysis model for the assumption, and finally calculates the performance characteristics of the image to be evaluated under the model, so as to obtain the image quality. Quality evaluation results.
  • High-dynamic range images can provide more dynamic range and image details than ordinary images. According to different exposure times, LDR (Low-Dynamic Range, low dynamic range images), And use the LDR images with the best detail corresponding to each exposure time to synthesize the final HDR image. It can better reflect the visual effect in the real environment.
  • PSNR Peak Signal to Noise Ratio
  • I represents the input image
  • K represents the reference image
  • MAX represents the maximum allowed image value. It is an objective standard for evaluating images. It has limitations and is generally used for an engineering project between the maximum signal and background noise. Peak in Chinese means apex. And Ratio means ratio or proportion. The whole meaning is to reach the peak signal of the noise ratio, PSNR is generally used for an engineering project between the maximum signal and the background noise. Usually after image compression, the output image will be different from the original image to some extent. In order to measure the image quality after processing, we usually refer to the PSNR value to measure whether a processing program is satisfactory. It is the logarithm of the mean square error between the original image and the processed image relative to (2 n -1) 2 (the square of the maximum signal value, where n is the number of bits per sample), and its unit is dB.
  • FIG. 1 shows a schematic diagram of an implementation environment provided by an embodiment of the present application.
  • the implementation environment of this solution can be realized as an automatic driving task system.
  • the implementation environment may include: a terminal 10 and a server 20 .
  • the terminal 10 can be an electronic device such as a mobile phone, a tablet computer, a multimedia playback device, a PC (Personal Computer), etc., or an intelligent vehicle-mounted terminal loaded in an unmanned vehicle or an ordinary vehicle, or any device involving image processing. terminal.
  • the terminal 10 may be configured with or connected to a camera, and images are collected through the camera.
  • the server 20 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.
  • the server 20 may be a background server that provides services for the above-mentioned terminal 10 .
  • Communication between the terminal 10 and the server 20 may be performed through a network, which is not limited in this application.
  • the execution body of each step may be the server 20 or the terminal 10 , and may also be executed by the terminal 10 and the server 20 interactively and cooperatively.
  • the execution subject of each step for introduction and description, but this does not constitute a limitation.
  • the image quality assessment method provided in this application is mainly applied to autonomous driving scenarios.
  • FIG. 2 shows a flowchart of an image processing method provided by an embodiment of the present application.
  • the method may include the following steps (210-250):
  • Step 210 acquiring an image to be detected.
  • the image to be detected refers to an image of the image quality to be detected. This embodiment of the present application does not limit the data format of the image.
  • the image to be detected is a compressed image.
  • FIG. 3 it exemplarily shows a schematic diagram of an image to be detected.
  • Fig. 3(a) is the original uncompressed image
  • Fig. 3(b) and Fig. 3(c) are the compressed images obtained after compressing Fig. 3(a) respectively
  • Fig. 3(b) and Fig. 3(c) ) are compressed differently.
  • the image quality of FIG. 3(b) and FIG. 3(c) can be evaluated by the method provided in the embodiment of the present application, and the specific method is referred to later.
  • the image to be detected is a lane image collected during the driving of the vehicle.
  • the so-called autonomous driving scene that is, the image to be detected is captured by the front-view, rear-view or side-view camera of the autonomous vehicle, and the image usually includes roads, pedestrians, vehicles, road signs, etc.
  • FIG. 4 it exemplarily shows a schematic diagram of an image to be detected.
  • Figure 4 is a lane image collected while the vehicle is running.
  • the image to be detected is a dehaze image.
  • FIG. 5 it exemplarily shows a schematic diagram of an image to be detected.
  • Fig. 5(a) is the original image without dehazing
  • Fig. 5(b), Fig. 5(c), and Fig. 5(d) are obtained after dehazing image processing of Fig. 5(a) respectively.
  • the dehazing effects of Figure 5(b), Figure 5(c), and Figure 5(d) are different.
  • the image quality of FIG. 5(b), FIG. 5(c), and FIG. 5(d) can be evaluated by the method provided in the embodiment of the present application, and the specific method is referred to later.
  • the image to be detected is a video surveillance image collected by a terminal, and the image quality of the video surveillance image can be evaluated by the method provided in the embodiment of the present application, and the specific method is described later.
  • Step 220 Obtain the position weight of the target object based on the position information of the target object in the image to be detected.
  • a target object is any object used to detect image quality.
  • the target object includes any one of the following: target pixel, target image block, and target entity.
  • the target pixel refers to any pixel in the image to be detected.
  • the target image block refers to any image block in the image to be detected, and the image block may be a regular-shaped image block or an irregular-shaped image block.
  • the target entity refers to any entity in the image to be detected, and the above entities include human bodies and objects, such as pedestrians, vehicles, trees, roads and other entities.
  • the weight information of the target object can be obtained.
  • the attribute information includes various data information describing the target object.
  • the attribute information includes location information.
  • the attribute information includes category information.
  • the weight information reflects the relative importance of the target object in the image to be detected.
  • the position information is used to reflect the position of the target object, which can be either the position of the target object in the image coordinate system of the image to be detected, or the position of the target object in the physical space described by the image to be detected. Not limited.
  • the position weight is used to reflect the relative importance of the position occupied by the target object in the image to be detected.
  • the position weight refers to the importance of the position of each pixel in the image in the corresponding actual physical space.
  • the closer the pixel is to the ego vehicle the higher the position weight. For example, the position directly in front of the ego vehicle has the highest weight, while the position in the distant sky has the lowest weight.
  • step 220 includes the following sub-steps:
  • Step 221 Obtain position information of the target object in the image to be detected.
  • the location information includes the location of the target object in the physical space described by the image to be detected.
  • Step 222 Obtain the position weight of the target object according to the position weight relationship and the position information of the target object.
  • the position weight relationship reflects the mathematical correspondence between the position and the weight.
  • the position-weight relationship is reflected by setting the function of position and weight.
  • Step 230 Obtain the category weight of the target object based on the category information of the target object.
  • the category weight is used to reflect the relative importance of the category to which the target object belongs in the image to be detected.
  • FIG. 6 it exemplarily shows a schematic diagram of an image to be detected.
  • Fig. 6(a) is the driving image of the vehicle before compression
  • Fig. 6(b) is the driving image of the vehicle after compression, in which the tree 1 in Fig. 6(a) and the tree 2 in Fig. 6(b) are the same tree.
  • the compression algorithm causes blurring of vehicles in the original image
  • the blur caused by image compression in the area where the trees are located has little effect.
  • all areas of the image and The importance of the pixels is the same, which does not conform to the specific task scene. In this way, the impact caused by the trees will be reflected, but this impact is not concerned in this application.
  • the level of concern for the image quality of trees Therefore, by adding different weights to different objects or regions, different image contents can be focused on differently.
  • step 230 includes the following sub-steps:
  • Step 231 Identify the category information of the target object in the image to be detected.
  • the category information includes the category of the entity to which the target object belongs in the image to be detected
  • Semantic segmentation is performed on the image to be detected to obtain category information of each object in the image to be detected.
  • the neural network-based semantic segmentation algorithm classifies each object in the image and outputs the object category corresponding to each pixel.
  • object category corresponding to each pixel.
  • Step 232 Obtain the category weight of the target object according to the category weight relationship and the category information of the target object.
  • the category weight relationship reflects the mathematical correspondence between categories and weights.
  • search is performed in the category weight data structure to obtain the category weight of the target object.
  • the weight coefficients of each category are preset.
  • FIG. 7 it exemplarily shows a schematic diagram reflecting the category-weight relationship.
  • the horizontal axis represents the category
  • the vertical axis represents the weight value. It can be seen from the broken line that different target categories have different weights.
  • different category weights are given to different objects. This weight is assigned to the corresponding category area on the semantic segmentation map. For example, if the semantic segmentation map includes the vehicle area and the tree area, the category weight of the vehicle area is higher than that of the tree area.
  • Step 240 Obtain the comprehensive weight of the target object according to the position weight and the category weight of the target object.
  • the comprehensive weight refers to a data index that integrates various factors to reflect the relative importance of the target pixel.
  • a weighted summation process is performed on the position weight and the category weight to obtain the comprehensive weight of the target pixel.
  • the composite weight of a pixel can be calculated by the following formula:
  • m and n are weighting coefficients, which can be flexibly adjusted according to the task.
  • W ij is the comprehensive weight of the target pixel
  • W d is the position weight of the target pixel
  • W object is the class weight of the target pixel.
  • FIG. 8 it exemplarily shows a schematic diagram of a driving image.
  • the technical solution provided in this application has different degrees of attention to objects at different distances from the center of the vehicle, for example, the degree of attention to nearby vehicles is greater than that of distant vehicles. Weights.
  • the importance of different objects or areas is different, and the importance of different types of entities is different.
  • vehicles and pedestrians are more important than trees and lamp posts.
  • the category weight of the vehicle 50 ahead in the figure is more than The category weight of trees 40; for example, road>roadside, ground>sky, for example, the weight of the ground area 10 is greater than the weight of the height area 20 in the figure.
  • Step 250 Determine the image quality score of the image to be detected based on the comprehensive weight of each object in the image to be detected.
  • the above image quality score refers to a quantitative data index for measuring image quality.
  • a relative or absolute value can be output as a measure of image quality, which is conducive to quantitative evaluation of image compression algorithms, image generation algorithms and image enhancement algorithms.
  • image quality detection processing is performed on the image to be detected, and an image quality score of the image to be detected is obtained.
  • image quality detection processing is performed on the image to be detected to obtain an image quality score of the image to be detected.
  • image quality detection processing is performed on the image to be detected to obtain an image quality score of the image to be detected.
  • the image quality score is used to evaluate the image compression algorithm
  • the image quality score is used to evaluate the preservation degree of the key information of the image before and after image compression.
  • the higher the image quality score the better the compression effect, and the less key information is lost in the compressed image.
  • step 250 when calculating data indicators such as PSNR/SSIM, different pixels may be multiplied by different comprehensive weights to obtain a final quality score.
  • steps 251-253 are an implementation of step 250, which can calculate PSNR/SSIM according to weights:
  • Step 251 Obtain the pixel difference value of each pixel between the image to be detected and the reference image.
  • the reference image refers to the uncompressed original image of the image to be detected.
  • Step 252 obtaining the weighted similarity between the image to be detected and the reference image based on the summed weight of each pixel and the pixel difference value on each pixel.
  • the weighted similarity refers to the peak signal-to-noise ratio and/or the structural similarity calculated based on the summed weight, and the peak signal-to-noise ratio and/or the structural similarity are used to reflect the degree of difference between the image to be detected and the reference image.
  • the peak signal-to-noise ratio and/or the structural similarity between the image to be detected and the reference image are obtained.
  • W ij is the comprehensive weight of the pixel in the i-th row and the j-th column.
  • Step 253 based on the weighted similarity, obtain an image quality score of the image to be detected.
  • the weighted similarity is used as the image quality score of the image to be detected.
  • the image quality score is a value obtained by performing normalization processing on the weighted similarity, such as normalization processing, and the value interval of the image quality score is within [0, 1], which is convenient for comparison.
  • the technical solutions provided by the embodiments of the present application obtain the position weight reflecting the importance of the position of the target object by detecting the position of the target object in the image to be detected, and then obtain the target object reflecting the category of the target object by identifying the category of the target object.
  • the category weight of category importance further combines the position weight of the target object with the category weight to obtain a comprehensive weight reflecting the comprehensive importance of the target object. Detecting the image quality of an image makes the image quality evaluation process more reasonable, the obtained image quality score reflecting the image quality is more accurate, and the accuracy of the image quality evaluation is improved.
  • the embodiment of the present application designs a position weight model, which can accurately calculate the position weight of each pixel in the image; and through the semantic segmentation model, the image to be detected is divided into different regions according to categories, and the category of each region is calculated.
  • the weights ensure the accuracy of determining the category weights; finally, the position weights and category weights are integrated, and the final image quality scores are calculated by combining traditional PSNR/SSIM methods, which further ensures the reliability of the obtained image quality scores.
  • the image quality score has a variety of uses, which are described below.
  • Different image processing processes can obtain images of different quality.
  • These image processing processes include camera systems, image signal processing, storage, transmission, compression, etc.
  • Image Quality Assessment is an empirical way to measure image quality, including subjective and objective methods.
  • the subjective method is based on the perception of objects or attributes in the image by human vision, and the judgment of the quality of the image is obtained; the objective method is based on the preset calculation model to obtain quantitative values.
  • the technical solution provided by the present application mainly relates to an objective image quality assessment method, that is, designing a calculation model for calculating the image quality score.
  • Image quality assessment can be applied to many different fields, and each field uses image quality assessment for different methods and purposes. For example, evaluation agencies are used to evaluate the pros and cons of pictures taken by different cameras; in the process of machine learning model training, it is used to guide the convergence direction of algorithms, etc.
  • the image quality score output by the image quality evaluation method provided in the embodiment of the present application can be used to measure whether the image is suitable for the automatic driving task.
  • the image processing algorithm is adjusted based on the image quality score. For example, make adjustments to image compression algorithms.
  • the similarity of two pictures is measured by the peak signal-to-noise ratio or the structural similarity measure. Whether it is the peak signal-to-noise ratio or the structural similarity measure, the solution needs to calculate the difference between the input picture and the reference picture.
  • the difference or distance between them is the full reference image quality assessment (FR-IQA).
  • the reference picture is generally an uncompressed original picture. The larger the distance, the worse the quality of the input picture.
  • the content of the calculation cannot represent the real situation.
  • the formula contains too much content unrelated to the scene or task, such as PSNR calculates the MSE of all pixels, but in fact, not all pixels will have the same value impact on the task; on the other hand, the calculation of the formula The process discards the information of some images, which leads to the deviation of the results from the actual situation.
  • PSNR calculates the MSE of a single pixel, but in fact, for the image, there is a close relationship between the pixels, and the size and shape of the object are reflected in these The relationship between pixels, so the calculation of PSNR discards these important information.
  • the IQA evaluation method provided in this application is suitable for automatic driving tasks.
  • This method can evaluate the quality of input images in the field of automatic driving, and the algorithm is simple and easy to implement.
  • IQA is usually used as the evaluation of image processing algorithms (such as image compression) or the loss function of neural networks, and it is necessary to avoid the algorithm being too complicated or difficult to implement. Therefore, the present application is designed to be a simple, efficient and general IQA algorithm, which is suitable for a variety of application scenarios, and the weights can be implemented according to specific application scenarios.
  • the weight value may also be added according to the following information, such as vehicle speed information, pedestrian movement speed and direction, and the distance from the pedestrian to the lane, and other data indicators reflecting the movement state of an object. Please refer to FIG. 9 , the above step 210 further includes the following steps:
  • Step 260 Detect the motion state of the entity to which the target object belongs in the image to be detected.
  • the motion state of the entity to which the target object belongs in the to-be-detected image is obtained by selecting the previous frames or previous frames of the to-be-detected image and compared with the to-be-detected image. For example, if a pedestrian is in a moving state of walking, the position information of the pedestrian will change, and the current pedestrian's moving state can be determined by comparing it with the previous position of the pedestrian.
  • Step 270 Determine whether the motion state of the entity to which the target object belongs in the image to be detected meets the conditions. If yes, go to step 280; otherwise, go to step 240.
  • Step 280 Determine the state weight of the target object.
  • the state weight of the target object is determined, and the state weight is used to represent the relative importance of the target object in the image to be detected under different motion states.
  • Step 290 Combine the state weight of the target object with the category weight of the target object to obtain the state category weight of the target object, and the state category weight is used to represent the relative importance of different categories of target objects in different motion states.
  • the state class weight can be expressed as W object , which can be calculated by the following formula:
  • W o represents the preset category weight for the category to which the target object belongs
  • W enhance represents the state weight increased according to the motion state of the target object, for example, the weight coefficient increased according to the vehicle or pedestrian status.
  • step 240 can be replaced by the following step 291:
  • Step 291 Obtain the comprehensive weight of the target object according to the position weight and the state category weight of the target object.
  • the comprehensive weight of the target object is obtained according to the position weight of the target object and the weight of the state category.
  • the step of obtaining the comprehensive weight of the target object according to the position weight and the category weight of the target object is performed.
  • FIG. 10 it exemplarily shows a schematic diagram of the image processing flow provided by the embodiment of the present application, which is used to evaluate whether the compression algorithm has an impact on the automatic driving task.
  • the figure includes the image input link, the position weight calculation link, the semantic segmentation link, the category weight calculation link, the global image quality score calculation link and the weighted image quality score calculation link in the image processing flow.
  • the image input link is used to obtain the image to be evaluated and the reference image.
  • the position weight calculation link is used to calculate the position weight of each pixel of the image.
  • the semantic segmentation model is used to obtain the categories of different pixels in the image.
  • the category weight value corresponding to each pixel is obtained according to the category result output by the semantic segmentation model and the preset weights of different categories.
  • the global image quality score calculation link uses the input image and the reference image to calculate the global image quality score. Available methods include PSNR and SSIM.
  • the weighted image quality score calculation link combines the position weight and the category weight to obtain the comprehensive weight of each pixel, and calculates the final image quality score in combination with the global image quality score.
  • the technical solutions provided by the embodiments of the present application in addition to obtaining the position weight of the target object reflecting the importance of the position of the target object and the category weight reflecting the importance of the category of the target object, also obtain the target object according to the motion state of the target object. Increase the state weight, and then combine the three weights reasonably to obtain a comprehensive weight that reflects the comprehensive importance of the target object from multiple aspects, calculate the weight of the target object more scientifically and reliably, and further ensure the accuracy of image quality assessment.
  • FIG. 11 shows a flowchart of an image processing method provided by an embodiment of the present application.
  • the method may include the following steps (1-14):
  • Step 1 acquiring an image to be detected.
  • Step 2 detecting whether the image to be detected is a top view. If yes, go to step 3; otherwise, go to step 5.
  • Step 3 Perform image transformation processing on the front view or the side view to obtain a top view corresponding to the front view or the side view.
  • image transformation processing is performed on the front view or the side view to obtain a top view corresponding to the front view or the side view.
  • Step 4 Based on the pixel mapping relationship between the front view or the side view and the top view, the position coordinates of the target pixel in the front view or the side view in the top view are obtained.
  • Step 5 Obtain the position weight of the target pixel according to the position weight curve and the position coordinates of the target pixel in the top view.
  • the position weight curve reflects the relative importance of pixel positions by describing the mathematical relationship between pixel positions and weights.
  • the above process can be simply understood as: if the input image is a front view or a side view, convert it to a top view; in a possible design, the following formula is used to calculate the position weights corresponding to different pixels in the top view.
  • a and b are parameters related to the curve shape.
  • P 0 (x0, y0) is the center point of the ego vehicle.
  • the above formula can be applied to the above plan view.
  • the position weight of each pixel in the image to be detected is equal to the position weight of the corresponding pixel in the top view of each pixel in the image to be detected.
  • FIG. 12 it exemplarily shows a schematic diagram of a front view and a top view in an automatic driving task.
  • Fig. 12(a) is a front view of the road in front of the current driving vehicle
  • Fig. 12(b) is a top view obtained by performing image transformation from the front view shown in Fig. 12(a).
  • the position weight value of the pixel is obtained based on the transformation relationship between the front view and the top view to obtain the position weight value of each pixel in the front view.
  • the pixel that is closer to the center n of the vehicle has a higher weight, and its corresponding position in the front view is directly in front of the vehicle.
  • the distance between the center n and the center n-1 of the other car is calculated.
  • the image quality in this area needs to be guaranteed.
  • the quality of the distant image can be lower.
  • the purpose of setting the weight value is to focus on the image quality, and more Focusing on the image quality of the area near the vehicle makes the image quality evaluation more reasonable.
  • Step 6 Perform semantic segmentation processing on the image to be detected to obtain category information of each pixel in the image to be detected.
  • Step 7 Based on the category information of the target pixel, search in the category weight data structure to obtain the category weight of the target pixel.
  • Step 8 Detect the motion state of the entity to which the target object belongs in the image to be detected.
  • Step 9 Determine whether the motion state of the entity to which the target object belongs in the image to be detected meets the conditions. If yes, go to step 10; otherwise, go to step 13.
  • Step 10 Determine the state weight of the target object.
  • Step 11 Combine the state weight of the target object with the category weight of the target object to obtain the state category weight of the target object.
  • Step 12 Obtain the comprehensive weight of the target object according to the position weight and the state category weight of the target object.
  • Step 13 Obtain the comprehensive weight of the target object according to the position weight and the category weight of the target object.
  • Step 14 Determine the image quality score of the image to be detected based on the comprehensive weight of each object in the image to be detected.
  • the technical solutions provided in the embodiments of the present application perform image quality detection in units of pixels, combine the obtained position weight, category weight, and state weight of the target pixel to obtain the comprehensive weight of the target pixel, and then traverse all the target pixels.
  • the comprehensive weight of each pixel in the image to be detected can be obtained, and then the image quality score of the image to be detected can be calculated in units of pixels, which further ensures the accuracy of image quality evaluation.
  • FIG. 13 shows a block diagram of an image processing apparatus provided by an embodiment of the present application.
  • the device has the function of realizing the above-mentioned image processing method.
  • the apparatus 1300 may include: a to-be-measured image acquisition module 1310 , a position weight determination module 1320 , a category weight determination module 1330 , a comprehensive weight determination module 1340 , and an image quality detection module 1350 .
  • the to-be-detected image acquisition module 1310 is used to acquire the to-be-detected image.
  • a position weight determination module 1320 configured to obtain the position weight of the target object based on the position information of the target object in the to-be-detected image, where the position weight is used to reflect the position occupied by the target object in the to-be-detected image relative importance of location.
  • a category weight determination module 1330 configured to obtain the category weight of the target object based on the category information of the target object, where the category weight is used to reflect the relative importance of the category to which the target object belongs in the image to be detected .
  • the comprehensive weight determination module 1340 is configured to obtain the comprehensive weight of the target object according to the position weight and category weight of the target object.
  • the image quality detection module 1350 is configured to determine the image quality score of the to-be-detected image based on the comprehensive weight of each object in the to-be-detected image.
  • the location weight determination module 1320 includes: a location information acquisition unit 1321 and a location weight determination unit 1322 .
  • the location information acquiring unit 1321 is configured to acquire location information of the target object in the image to be detected, where the location information includes the location of the target object in the physical space described by the image to be detected.
  • the location weight determination unit 1322 is configured to obtain the location weight of the target object according to the location weight relationship and the location information of the target object, where the location weight relationship reflects the mathematical correspondence between the location and the weight.
  • the target object includes target pixels, and the position information acquisition unit 1321 is used for:
  • image transformation processing is performed on the front view or the side view to obtain a top view corresponding to the front view or the side view;
  • the position coordinates of the target pixel in the top view in the front view or the side view are acquired.
  • the position weight determination unit 1322 is used for:
  • the position weight of the target pixel is obtained, and the position weight curve reflects the pixel position by describing the mathematical relationship between the pixel position and the weight. relative importance.
  • the category weight determination module 1330 includes: a category information identification unit 1331 and a category weight determination unit 1332 .
  • the category information identification unit 1331 is configured to identify category information of the target object in the image to be detected, where the category information includes the category of the entity to which the target object belongs in the image to be detected.
  • the category weight determination unit 1332 is configured to obtain the category weight of the target object according to the category weight relationship and the category information of the target object, where the category weight relationship reflects the mathematical correspondence between categories and weights.
  • the category information identification unit 1331 is used for:
  • Semantic segmentation is performed on the to-be-detected image to obtain category information of each object in the to-be-detected image.
  • the class weight determination unit 1332 is used for:
  • the category weight data structure is searched to obtain the category weight of the target object.
  • the apparatus 1300 further includes: a motion state detection module 1360 , a state weight determination module 1370 and a state category weight determination module 1380 .
  • the motion state detection module 1360 is configured to detect the motion state of the entity to which the target object belongs in the to-be-detected image.
  • the state weight determination module 1370 is configured to determine the state weight of the target object under the condition that the motion state of the entity meets the conditions, and the state weight is used to characterize the different motions of the target object in the to-be-detected image relative importance of the state.
  • the state category weight determination module 1380 is configured to combine the state weight of the target object with the category weight of the target object to obtain the status category weight of the target object, and the state category weight is used to represent different categories of targets The relative importance of objects in different motion states.
  • the comprehensive weight determination module 1340 is configured to obtain the comprehensive weight of the target object according to the position weight and the state category weight of the target object when the target object has a state category weight.
  • the comprehensive weight determination module 1340 is further configured to perform the step of obtaining the comprehensive weight of the target object according to the position weight and the category weight of the target object when the target object does not have a state category weight .
  • the integrated weight determination module 1340 is used to:
  • a weighted summation process is performed on the position weight and the category weight to obtain the comprehensive weight of the target pixel.
  • the target object includes target pixels, and the image to be detected is a compressed image;
  • the image quality detection module 1350 includes: a pixel difference acquisition unit 1351 , a weighted similarity calculation unit 1352 And the quality score evaluation unit 1353.
  • the pixel difference obtaining unit 1351 is configured to obtain a pixel difference value between the image to be detected and a reference image on each pixel, where the reference image refers to an uncompressed original image of the image to be detected.
  • a weighted similarity calculation unit 1352 configured to obtain the weighted similarity between the image to be detected and the reference image based on the summed weight of each pixel and the pixel difference value on each pixel, the weighted similarity
  • the similarity refers to the peak signal-to-noise ratio and/or the structural similarity calculated based on the summation weight, and the peak signal-to-noise ratio and/or the structural similarity are used to reflect the difference between the image to be detected and the reference image degree of difference.
  • the quality score evaluation unit 1353 is configured to obtain the image quality score of the to-be-detected image based on the weighted similarity.
  • the apparatus 1300 further includes: an algorithm adjustment module 1390 .
  • the algorithm adjustment module 1390 is configured to adjust the image processing algorithm based on the image quality score.
  • the target object includes any one of the following: a target pixel, a target image block, and a target entity.
  • the technical solutions provided by the embodiments of the present application obtain the position weight reflecting the importance of the position of the target object by detecting the position of the target object in the image to be detected, and then obtain the target object reflecting the category of the target object by identifying the category of the target object.
  • the category weight of category importance further combines the position weight of the target object with the category weight to obtain a comprehensive weight reflecting the comprehensive importance of the target object. Detecting the image quality of the image makes the image quality evaluation process more reasonable, the obtained image quality score reflecting the image quality is more accurate, and the accuracy of the image quality evaluation is improved.
  • the embodiment of the present application designs a position weight model, which can accurately calculate the position weight of each pixel in the image; and through the semantic segmentation model, the image to be detected is divided into different regions according to categories, and the category of each region is calculated.
  • the weights ensure the accuracy of determining the category weights; finally, the position weights and category weights are integrated, and the final image quality scores are calculated by combining traditional PSNR/SSIM methods, which further ensures the reliability of the obtained image quality scores.
  • FIG. 15 shows a structural block diagram of a computer device 1500 provided by an embodiment of the present application.
  • the computer device 1500 may be an electronic device such as a mobile phone, a tablet computer, a multimedia playback device, a wearable device, a PC (Personal Computer), a language learning terminal, an intelligent teaching machine, and the like.
  • the computer device is used to implement the image processing method provided in the above embodiment.
  • the computer device may be the terminal 10 or the server 20 in the application execution environment shown in FIG. 1 .
  • computer device 1500 includes: processor 1501 and memory 1502 .
  • the processor 1501 may include one or more processing cores, such as a 4-core processor, an 8-core processor, and the like.
  • the processor 1501 can be implemented by at least one hardware form among DSP (Digital Signal Processing, digital signal processing), FPGA (Field Programmable Gate Array, field programmable gate array), and PLA (Programmable Logic Array, programmable logic array).
  • the processor 1501 may also include a main processor and a coprocessor.
  • the main processor is a processor used to process data in the wake-up state, also called CPU (Central Processing Unit, central processing unit); the coprocessor is A low-power processor for processing data in a standby state.
  • the processor 1501 may be integrated with a GPU (Graphics Processing Unit, image processor), and the GPU is used for rendering and drawing the content that needs to be displayed on the display screen.
  • the processor 1501 may further include an AI (Artificial Intelligence, artificial intelligence) processor, where the AI processor is used to process computing operations related to machine learning.
  • AI Artificial Intelligence, artificial intelligence
  • Memory 1502 may include one or more computer-readable storage media, which may be non-transitory. Memory 1502 may also include high-speed random access memory, as well as non-volatile memory, such as one or more disk storage devices, flash storage devices. In some embodiments, a non-transitory computer-readable storage medium in memory 1502 is used to store at least one instruction, at least one program, code set or instruction set, the at least one instruction, at least one program, code set or instruction set and configured to be executed by one or more processors to implement the image processing method described above.
  • the computer device 1500 may also optionally include: a peripheral device interface 1503 and at least one peripheral device.
  • the processor 1501, the memory 1502 and the peripheral device interface 1503 can be connected through a bus or a signal line.
  • Each peripheral device can be connected to the peripheral device interface 1503 through a bus, a signal line or a circuit board.
  • FIG. 15 does not constitute a limitation on the computer device 1500, and may include more or less components than the one shown, or combine some components, or adopt different component arrangements.
  • a computer-readable storage medium is also provided, and a computer program is stored in the storage medium, and the computer program, when executed by a processor, implements the above-mentioned image processing method.
  • the computer-readable storage medium may include: ROM (Read Only Memory, read-only memory), RAM (Random Access Memory, random access memory), SSD (Solid State Drives, solid-state hard disk), or an optical disk.
  • the random access memory may include ReRAM (Resistance Random Access Memory, resistive random access memory) and DRAM (Dynamic Random Access Memory, dynamic random access memory).
  • a computer program product or computer program comprising computer instructions stored in a computer readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the above-mentioned image processing method.
  • references herein to "a plurality” means two or more.
  • "And/or" which describes the association relationship of the associated objects, means that there can be three kinds of relationships, for example, A and/or B, which can mean that A exists alone, A and B exist at the same time, and B exists alone.
  • the character "/" generally indicates that the associated objects are an "or” relationship.
  • the numbering of the steps described in this document only exemplarily shows a possible execution sequence between the steps. In some other embodiments, the above steps may also be executed in different order, such as two different numbers. The steps are performed at the same time, or two steps with different numbers are performed in a reverse order to that shown in the figure, which is not limited in this embodiment of the present application.

Abstract

本申请公开了一种图像处理方法、装置、设备及存储介质，属于人工智能技术领域。所述方法包括：获取待检测图像；基于待检测图像中目标对象的位置信息，获取目标对象的位置权重；基于目标对象的类别信息，获取目标对象的类别权重；根据目标对象的位置权重与类别权重，得到目标对象的综合权重；基于待检测图像中各对象的综合权重，确定待检测图像的图像质量分。本申请实施例提供的技术方案中，通过获取目标对象的位置权重以及类别权重，进而得到反映目标对象的综合重要性的综合权重，最终基于各对象的综合权重来有侧重地评估待检测图像的图像质量，得到合理的反映图像质量的图像质量分，提升图像质量评估的准确性。

Description

图像处理方法、装置、设备及存储介质 技术领域
本申请涉及人工智能技术领域，特别涉及一种图像处理方法、装置、设备及存储介质。
背景技术
近年来，随着对数字图像领域的广泛研究，图像质量评价的研究也越来越受到研究者的关注，提出并完善了许多图像质量评价的指标和方法。
相关技术中，通常通过计算输入图片与基准图片之间的差别或者距离，即全参考的图像质量评估(FR-IQA)。基准图片一般是未压缩的原始图片，距离越大，说明输入图片的质量越差。
相关技术中，计算的内容与任务无关，丢弃了图片中的重要信息，图片质量评估的准确性差。
发明内容
本申请实施例提供了一种图像处理方法、装置、设备及存储介质，能够提高图片质量评估的准确性。所述技术方案如下：
根据本申请实施例的一个方面，提供了一种图像处理方法，所述方法包括：
获取待检测图像；
基于所述待检测图像中目标对象的位置信息，获取所述目标对象的位置权重，所述位置权重用于反映所述目标对象在所述待检测图像中所占据位置的相对重要程度；
基于所述目标对象的类别信息，获取所述目标对象的类别权重，所述类别权重用于反映所述目标对象所属类别在所述待检测图像中的相对重要程度；
根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重；
基于所述待检测图像中各对象的综合权重，确定所述待检测图像的图像质量分。
通过检测目标对象在待检测图像中的位置，从而得到反映目标对象位置重要性的位置权重，再通过识别目标对象的类别来获取反映目标对象类别重要性的类别权重，进一步地将目标对象的位置权重与类别权重进行结合，得到反映目标对象的综合重要性的综合权重，最终在考虑各对象的综合权重的情况下，有侧重地评估待检测图像的图像质量，使得图像质量评估过程更加合理，所得到反映图像质量的图像质量分更加准确，提升图像质量评估的准确性。
在一个可能的设计中，所述基于所述待检测图像中目标对象的位置信息，获取所述目标对象的位置权重，包括：
获取所述待检测图像中所述目标对象的位置信息，所述位置信息包括所述目标对象在所述待检测图像描述的物理空间中的位置；
根据位置权重关系以及所述目标对象的位置信息，得到所述目标对象的位置权重，所述位置权重关系反映位置与权重之间的数学对应关系。
通过上述方式，可以根据目标对象的实际空间位置获得更加精确的位置权重。
在一个可能的设计中，所述目标对象包括目标像素，所述获取所述待检测图像中目标对象的位置信息，包括：
在所述待检测图像是前视图或侧视图的情况下，对所述前视图或所述侧视图进行图像变换处理，得到与所述前视图或所述侧视图对应的俯视图；
基于所述前视图或所述侧视图与所述俯视图之间的像素映射关系，获取所述前视图或所述侧视图中所述目标像素在所述俯视图中的位置坐标。
通过上述方式，可以更加准确地得到反映目标对象的实际空间位置的数据信息。
在一个可能的设计中，所述根据位置权重关系以及所述目标对象的位置信息，得到所述目标对象的位置权重，包括：
根据位置权重曲线以及所述目标像素在所述俯视图中的位置坐标，获取所述目标像素的位置权重，所述位置权重曲线通过描述像素位置与权重之间的数学关系来反映所述像素位置的相对重要程度。
通过上述方式，可以保证目标对象的空间位置与位置权重之间的映射关系的稳定性。
在一个可能的设计中，所述基于所述目标对象的类别信息，获取所述目标对象的类别权重，包括：
识别所述待检测图像中所述目标对象的类别信息，所述类别信息包括所述目标对象在所述待检测图像中所属实体的类别；
根据类别权重关系以及所述目标对象的类别信息，得到所述目标对象的类别权重，所述类别权重关系反映类别与权重之间的数学对应关系。
通过上述方式，根据识别得到的目标对象的类别，进而保证得到的目标对象的位置权重的准确性。
在一个可能的设计中，所述识别所述待检测图像中所述目标对象的类别信息，包括：
对所述待检测图像进行语义分割处理，得到所述待检测图像中各对象的类别信息；
所述根据类别权重关系以及所述目标对象的类别信息，得到所述目标对象的类别权重，包括：
基于所述目标对象的类别信息，在类别权重数据结构中进行查找，得到所述目标对象的类别权重。
通过上述方式，可以保证类别识别的准确性，并提升确定类别权重的效率。
在一个可能的设计中，所述获取待检测图像之后，还包括：
检测所述目标对象在所述待检测图像中所属实体的运动状态；
在所述实体的运动状态符合条件的情况下，确定所述目标对象的状态权重，所述状态权重用于表征所述待检测图像中所述目标对象在不同运动状态下的相对重要程度；
将所述目标对象的状态权重与所述目标对象的类别权重进行结合，得到所述目标对象的状态类别权重，所述状态类别权重用于表征不同类别的目标对象在不同运动状态下的相对重要程度；
在所述目标对象具有状态类别权重的情况下，根据所述目标对象的位置权重与状态类别权重，得到所述目标对象的综合权重；
在所述目标对象不具有状态类别权重的情况下，执行所述根据所述目标对象的位置权重 与类别权重，得到所述目标对象的综合权重的步骤。
通过上述方式，可以更加准确的通过权重反映目标对象的重要程度，权重不仅反映目标对象的位置信息和类别信息，还可反映目标对象的运动状态。
在一个可能的设计中，所述根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重，包括：
对所述位置权重与所述类别权重进行加权求和处理，得到所述目标像素的综合权重。
通过上述方式，使得综合权重的计算方式更加科学，从而能够更加客观的反映目标像素的重要程度。
在一个可能的设计中，所述目标对象包括目标像素，所述待检测图像是压缩图像；
所述基于所述待检测图像中各对象的综和权重，确定所述待检测图像的图像质量分，包括：
获取所述待检测图像与参考图像在各像素上的像素差值，所述参考图像是指所述待检测图像未经压缩的原始图像；
基于所述各像素的综和权重与所述各像素上的像素差值，得到所述待检测图像与所述参考图像之间的加权相似度，所述加权相似度是指基于所述综和权重计算的峰值信噪比和/或结构相似度，所述峰值信噪比和/或结构相似度用于反映所述待检测图像与所述参考图像之间的差异程度；
基于所述加权相似度，得到所述待检测图像的图像质量分。
通过上述方式，能够更加准确地评估图像质量，合理运用权重指标。
在一个可能的设计中，所述基于所述待检测图像中各对象的综合权重，确定所述待检测图像的图像质量分之后，还包括：
基于所述图像质量分，对图像处理算法进行调整。
通过上述方式，能够使用图像质量分提升图像处理算法所希望达到的目标效果，在尽可能不丢失图像中的重要信息的情况下，对图像进行处理。
在一个可能的设计中，所述目标对象包括以下任意一种：目标像素、目标图像块、目标实体。
通过上述方式，能够进行不同量级的图像质量评价。
根据本申请实施例的一个方面，提供了一种图像处理装置，所述装置包括：
待测图像获取模块，用于获取待检测图像；
位置权重确定模块，用于基于所述待检测图像中目标对象的位置信息，获取所述目标对象的位置权重，所述位置权重用于反映所述目标对象在所述待检测图像中所占据位置的相对重要程度；
类别权重确定模块，用于基于所述目标对象的类别信息，获取所述目标对象的类别权重，所述类别权重用于反映所述目标对象所属类别在所述待检测图像中的相对重要程度；
综合权重确定模块，用于根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重；
图像质量检测模块，用于基于所述待检测图像中各对象的综合权重，确定所述待检测图像的图像质量分。
根据本申请实施例的一个方面，提供了一种计算机设备，所述计算机设备包括处理器和 存储器，所述存储器中存储有计算机程序，所述计算机程序由所述处理器加载并执行以实现上述图像处理方法。
根据本申请实施例的一个方面，提供了一种计算机可读存储介质，所述存储介质中存储有计算机程序，所述计算机程序由处理器加载并执行以实现上述图像处理方法。
根据本申请实施例的一个方面，提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述图像处理方法。
本申请实施例提供的技术方案可以带来如下有益效果：
通过检测目标对象在待检测图像中的位置，从而得到反映目标对象位置重要性的位置权重，再通过识别目标对象的类别来获取反映目标对象类别重要性的类别权重，进一步地将目标对象的位置权重与类别权重进行结合，得到反映目标对象的综合重要性的综合权重，最终在考虑各对象的综合权重的情况下，有侧重地评估待检测图像的图像质量，使得图像质量评估过程更加合理，所得到反映图像质量的图像质量分更加准确，提升图像质量评估的准确性。
另外，本申请实施例设计了一个位置权重模型，能够准确计算图像中每个像素点的位置权重；并且通过语义分割模型，将待检测图像按照类别分割为不同区域，并计算每个区域的类别权重，保证确定类别权重的准确性；最终综合位置权重和类别权重，并结合传统的PSNR/SSIM等方法，计算最终的图像质量分数，进一步保证了得到的图像质量分的可靠性。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本申请一个实施例提供的实施环境的示意图；
图2是本申请一个实施例提供的图像处理方法的流程图；
图3、图4、图5、图6示例性示出了几种待检测图像的示意图；
图7示例性示出了一种反映类别-权重关系的示意图；
图8示例性示出了一种行车图像的示意图；
图9是本申请另一个实施例提供的图像处理方法的流程图；
图10示例性示出了本申请实施例提供的图像处理流程的示意图；
图11是本申请一个实施例提供的图像处理方法的流程图；
图12示例性示出了一种自动驾驶任务中前视图与俯视图的示意图；
图13是本申请一个实施例提供的图像处理装置的框图；
图14是本申请另一个实施例提供的图像处理装置的框图；
图15是本申请一个实施例提供的计算机设备的结构框图。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚，下面将结合附图对本申请实施方式作进 一步地详细描述。
首先，对本申请中的部分用语进行解释说明，以便于本领域技术人员理解。
图像质量评价(Image Quality Assessment，IQA)是图像处理中的基本技术之一，主要通过对图像进行特性分析研究，然后评估出图像优劣(图像失真程度)。图像质量评价在图像处理系统中，对于算法分析比较、系统性能评估等方面有着重要的作用。图像质量客观评价可分为全参考(Full-Reference，FR)，部分参考(Reduced-Reference，RR)和无参考(No-Reference，NR)三种类型。
全参考图像质量评估(Full Reference Image Quality Assessment，FR-IQA)，是指在选择理想图像作为参考图像的情况下，比较待评图像与参考图像之间的差异，分析待评图像的失真程度，从而得到待评图像的质量评估。常用的全参考图像质量客观评价主要以像素统计、信息论、结构信息三方面为基础。
1、基于图像像素统计基础
基于图像像素统计基础，峰值信噪比(Peak-Signal to Noise Ratio，PSNR)和均方误差(Mean Square Error，MSE)是比较常见的两种质量评价方法。它们通过计算待评测图像和参考图像对应像素点灰度值之间的差异，从统计角度来衡量待评图像的质量优劣。
2、基于信息论基础
基于信息论中信息熵基础，互信息被广泛用来评价图像质量。近些年，Sheikh和Bovik等人提出来了信息保真度准则(Information Fidelity Criterion，IFC)和视觉信息保真度(Visual Information Fidelity，VIF)两种算法。它们通过计算待评图像与参考图像之间的互信息来衡量待评图像的质量优劣。这两种方法具有一定的理论支撑，在信息保真度上拓展了图像与人眼之间的联系，但是这类方法对于图像的结构信息没有反应。
3、基于结构信息基础
人眼视觉的主要功能是提取背景中的结构信息，而且人眼视觉系统能高度自适应地实现这一目标，因此对图像的结构失真的度量应是图像感知质量的最好近似。在此基础上给出了一种符合人眼视觉系统特性的图像质量客观评判标准-结构相似度(Structure Similarity，SSIM)。
SSIM根据图像像素间的相关性构造出参考图像与待评图像之间的结构相似性，SSIM值越大，图像质量越好。该指标算法实现简单，质量评估性比较可靠，同时很多研究者结合人眼视觉系统对其又进行了许多改进，目前在图像处理各个方面都受到广泛应用。SSIM使用的两张图像中，一张为未经压缩的无失真图像，另一张为失真后的图像。作为结构相似性理论的实现，结构相似度指数从图像组成的角度将结构信息定义为独立于亮度、对比度的，反映场景中物体结构的属性，并将失真建模为亮度、对比度和结构三个不同因素的组合。用均值作为亮度的估计，标准差作为对比度的估计，协方差作为结构相似程度的度量。其计算方式如下图。其中x，y分别表示两张图像，μ和δ分别是均值和方差，c为常数。
Figure PCTCN2020133674-appb-000001
无参考图像质量评估(Non Reference Image Quality Assessment，NR-IQA)，无参考方法也称为首评价方法，因为一般的理想图像很难获得，所以这种完全脱离了对理想参考图像依赖的质量评价方法应用较为广泛。无参考方法一般都是基于图像统计特性。
1、均值
均值是指图像像素的平均值，它反映了图像的平均亮度，平均亮度越大，图像质量越好。
2、标准差
标准差是指图像像素灰度值相对于均值的离散程度。如果标准差越大，表明图像中灰度级分别越分散，图像质量也就越好。
3、平均梯度
平均梯度能反映图像中细节反差和纹理变换，它在一定程度上反映了图像的清晰程度。
1)熵
熵是指图像的平均信息量，它从信息论的角度衡量图像中信息的多少，图像中的信息熵越大，说明图像包含的信息越多。
一般而言，无参考图像质量评价方法首先对理想图像的特征作出某种假设，再为该假设建立相应的数学分析模型，最后通过计算待评图像在该模型下的表现特征，从而得到图像的质量评价结果。
高动态范围图像(High-Dynamic Range，简称HDR)，相比普通的图像，可以提供更多的动态范围和图像细节，根据不同的曝光时间的LDR(Low-Dynamic Range，低动态范围图像)，并利用每个曝光时间相对应最佳细节的LDR图像来合成最终HDR图像。它能够更好的反映出真实环境中的视觉效果。
峰值信噪比(Peak Signal to Noise Ratio，PSNR),是指信号的可能最大值和噪声的比值。由于信号常常有很宽的动态范围，所以常用对数的方式计算PSNR。通常情况下，PSNR可用于计算一张图像与基准图像的差异。其计算公式如下：
Figure PCTCN2020133674-appb-000002
Figure PCTCN2020133674-appb-000003
其中，I表示输入图像，K表示基准图像，MAX表示最大允许的图像值。是一种评价图像的客观标准，它具有局限性，一般是用于最大值信号和背景噪音之间的一个工程项目。Peak的中文意思是顶点。而Ratio的意思是比率或比例的。整个意思就是到达噪音比率的顶点信号，PSNR一般是用于最大值信号和背景噪音之间的一个工程项目。通常在经过影像压缩之后，输出的影像都会在某种程度与原始影像不同。为了衡量经过处理后的影像品质，我们通常会参考PSNR值来衡量某个处理程序能否令人满意。它是原图像与被处理图像之间的均方误差相对于(2 n-1) 2的对数值(信号最大值的平方，n是每个采样值的比特数)，它的单位是dB。
请参考图1，其示出了本申请一个实施例提供的实施环境的示意图。该方案实施环境可以实现成为一个自动驾驶任务系统。该实施环境可以包括：终端10和服务器20。
终端10可以是诸如手机、平板电脑、多媒体播放设备、PC(Personal Computer，个人电 脑)等电子设备，也可是装载于无人驾驶车辆、普通车辆中的智能车载终端，还可以是任何涉及图像处理的终端。终端10可以配置或者连接摄像头，通过该摄像头采集图像。
服务器20可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云计算服务的云服务器。服务器20可以是为上述终端10提供服务的后台服务器。
终端10和服务器20之间可以通过网络进行通信，本申请在此不作限制。
本申请实施例提供的图像处理方法，各步骤的执行主体可以是服务器20，也可以是终端10，还可以是由终端10和服务器20交互配合执行。为了便于说明，在下述方法实施例中，仅以各步骤的执行主体为计算机设备进行介绍说明，但对此不构成限定。在一个可能的设计中，本申请提供的图像质量评估方法主要应用于自动驾驶场景。
请参考图2，其示出了本申请一个实施例提供的图像处理方法的流程图。该方法可以包括以下几个步骤(210～250)：
步骤210，获取待检测图像。
待检测图像是指待检测图像质量的图像。本申请实施例对图像的数据格式不作限制。
在一种可能的实施方式中，待检测图像是压缩图像。在一个示例中，如图3所示，其示例性示出了一种待检测图像的示意图。其中，图3(a)是未经压缩的原始图像，图3(b)、图3(c)分别是压缩图3(a)后得到的压缩图像，图3(b)、图3(c)的压缩程度不同。可通过本申请实施例提供的方法对图3(b)、图3(c)的图像质量进行评估，具体方法参见后文。
在一种可能的实施方式中，待检测图像是车辆行驶过程中采集的车道图像。所谓的自动驾驶场景，即待检测图像是由自动驾驶车辆的前视、后视或者侧视摄像头拍摄而来，图像中通常包括道路、行人、车辆、路标等内容。在一个示例中，如图4所示，其示例性示出了一种待检测图像的示意图。图4是车辆行驶过程中采集的车道图像。
在一种可能的实施方式中，待检测图像是去雾度图像。一个示例中，如图5所示，其示例性示出了一种待检测图像的示意图。其中，图5(a)是未经去雾处理的原始图像，图5(b)、图5(c)、图5(d)分别是对图5(a)进行去雾图像处理后得到的去雾度图像，图5(b)、图5(c)、图5(d)的去雾效果不同。可通过本申请实施例提供的方法对图5(b)、图5(c)、图5(d)的图像质量进行评估，具体方法参见后文。
在一种可能的实施方式中，待检测图像是终端采集的视频监控图像，可通过本申请实施例提供的方法对视频监控图像的图像质量进行评估，具体方法参见后文。
步骤220，基于待检测图像中目标对象的位置信息，获取目标对象的位置权重。
图像中往往具有多个对象作为观察目标，可根据图像中所包含的对象来获取图像中蕴含的信息。目标对象是指用于检测图像质量的任一对象。目标对象包括以下任意一种：目标像素、目标图像块、目标实体。目标像素是指待检测图像中任一像素。目标图像块是指待检测图像中任一图像块，上述图像块可以是规则形状的图像块，也可以是非规则形状的图像块。目标实体是指待检测图像中任一实体，上述实体包括人体、物体，例如行人、车辆、树木、路面等实体。
根据待检测图像中目标对象的属性信息，可以得到目标对象的权重信息。属性信息包括 描述目标对象的各项数据信息。可选地，属性信息包括位置信息。可选地，属性信息包括类别信息。权重信息反映目标对象在待检测图像中的相对重要程度。
位置信息用于反映目标对象的位置，既可以是目标对象在待检测图像的图像坐标系中的位置，也可以是目标对象在待检测图像描述的物理空间中的位置，本申请实施例对此不作限定。
位置权重用于反映目标对象在待检测图像中所占据位置的相对重要程度。可选地，位置权重指的是图像中的每个像素，在对应的实际物理空间中的位置的重要性。在本方案中，距离自车越近的像素，其位置权重越高。例如，自车正前方的位置权重最高，而远程天空的位置权重最低。
在示例性实施例中，上述步骤220包括如下子步骤：
步骤221，获取待检测图像中目标对象的位置信息。
位置信息包括目标对象在待检测图像描述的物理空间中的位置。
步骤222，根据位置权重关系以及目标对象的位置信息，得到目标对象的位置权重。
位置权重关系反映位置与权重之间的数学对应关系。可选地，通过设置位置与权重的函数反应位置权重关系。
步骤230，基于目标对象的类别信息，获取目标对象的类别权重。
类别权重用于反映目标对象所属类别在待检测图像中的相对重要程度。
在示例性实施例中，如图6所示，其示例性示出了一种待检测图像的示意图。其中，图6(a)是压缩前的车辆行驶图像，图6(b)是压缩后的车辆行驶图像，其中图6(a)中的树木1和图6(b)中的树木2是同一颗树木。对于自动驾驶任务，更关注压缩算法是否对原始图片中的车辆等造成模糊，而对树木所在区域进行图片压缩而造成的模糊影响不大，但在传统的图片质量评价方法中图片的所有区域和像素的重要程度等同，不符合具体任务场景，在此种方式下树木造成的影响会体现出来，但是这种影响本申请并不关注，通过为树木设置一个较低的权重，可降低对图片中树木的图像质量的关注程度。因此，对不同目标或区域增加不同权重，可以有区分的关注不同的图像内容。
在示例性实施例中，上述步骤230包括如下子步骤：
步骤231，识别待检测图像中目标对象的类别信息。
类别信息包括目标对象在待检测图像中所属实体的类别；
对待检测图像进行语义分割处理，得到待检测图像中各对象的类别信息。
基于神经网络的语义分割算法，会对图像中的每个对象进行分类，输出每个像素对应的物体类别。例如车辆、行人等。
步骤232，根据类别权重关系以及目标对象的类别信息，得到目标对象的类别权重。
类别权重关系反映类别与权重之间的数学对应关系。
基于目标对象的类别信息，在类别权重数据结构中进行查找，得到目标对象的类别权重。
自定义每种类别的权重。具体来说，根据先验知识，预设每种类别的权重系数。
在一个示例中，如图7所示，其示例性示出了一种反映类别-权重关系的示意图。图中横轴表示类别，纵轴表示权重值，通过该折线可知，不同的目标类别设不同的权重。
结合语义分割算法的类别输出，赋予不同对象不同的类别权重。在语义分割图上对应类别区域，赋予此权重值。例如，语义分割图中包括车辆区域与树木区域，那么车辆区域的类 别权重要高于树木区域。
步骤240，根据目标对象的位置权重与类别权重，得到目标对象的综合权重。
综合权重是指综合各方面因素来反映目标像素的相对重要程度的数据指标。
在示例性实施例中，对位置权重与类别权重进行加权求和处理，得到目标像素的综合权重。
结合位置权重和类别权重，利用加权和的方式计算像素的综合权重。在一个示例中，可通过如下公式计算像素的综合权重：
W ij＝m*W d+n*W object
其中，m，n为加权系数，根据任务灵活调整。W ij是目标像素的综合权重，W d是目标像素的位置权重，W object是目标像素的类别权重。
在一个示例中，如图8所示，其示例性示出了一种行车图像的示意图。本申请提供的技术方案对与自车中心不同距离的物体有不同的关注程度，例如对近处车辆的关注程度大于远处车辆，例如图中的前方车辆50的位置权重大于旁边车辆30的位置权重。此外，本申请提供的技术方案中不同目标或区域的重要程度不同，不同类别的实体的重要程度不同，如车辆、行人的重要程度大于树木、灯柱，例如图中前方车辆50的类别权重大于树木40的类别权重；如路上>路边，地面>天空，例如图中地面区域10的权重大于高处区域20的权重。
步骤250，基于待检测图像中各对象的综合权重，确定待检测图像的图像质量分。
上述图像质量分是指衡量图像质量的量化数据指标。利用计算机视觉的方法，计算图像质量，可以输出一个相对值或者绝对值，作为图像质量的度量，有利于对图像压缩算法、图像生成类算法和图像增强类算法进行定量评估。
基于待检测图像中各显示元素的权重信息，对待检测图像进行图像质量检测处理，得到待检测图像的图像质量分。
可选地，基于待检测图像中各像素的类别权重，对待检测图像进行图像质量检测处理，得到待检测图像的图像质量分。
可选地，基于待检测图像中各像素的位置权重，对待检测图像进行图像质量检测处理，得到待检测图像的图像质量分。
在图像质量分用于评价图像压缩算法的情况下，图像质量分用于评价图像的关键信息在图像压缩前后的保存程度。可选地，图像质量分越高，压缩效果越好，压缩后的图片中丢失的关键信息越少。
在示例性实施例中，计算PSNR/SSIM等数据指标时，可对不同的像素乘以不同的综合权重，获得最终的质量分数。下述步骤251-253为步骤250的一种实施方式，能够根据权重计算PSNR/SSIM：
步骤251，获取待检测图像与参考图像在各像素上的像素差值。
参考图像是指待检测图像未经压缩的原始图像。
步骤252，基于各像素的综和权重与各像素上的像素差值，得到待检测图像与参考图像之间的加权相似度。
加权相似度是指基于综和权重计算的峰值信噪比和/或结构相似度，峰值信噪比和/或结构相似度用于反映待检测图像与参考图像之间的差异程度。
将各像素的综和权重代入与待检测图像与参考图像之间在各像素上的均方误差计算过程 中，得到待检测图像与参考图像之间的加权均方误差。
基于待检测图像与参考图像之间的加权均方误差，得到待检测图像与参考图像之间的峰值信噪比和/或结构相似度。
在一个示例中，上述计算过程可由如下公式体现：
Figure PCTCN2020133674-appb-000004
Figure PCTCN2020133674-appb-000005
其中，W ij为第i行第j列像素的综合权重。
这里仅以PSNR的计算作为示例，综合权重也可以用于SSIM等多种方法的度量。
步骤253，基于加权相似度，得到待检测图像的图像质量分。
可选地，将加权相似度作为待检测图像的图像质量分。
可选地，图像质量分是由加权相似度进行标准化处理得到的数值，例如归一化处理，图像质量分的取值区间在[0，1]内，便于比较。
综上所述，本申请实施例提供的技术方案，通过检测目标对象在待检测图像中的位置，从而得到反映目标对象位置重要性的位置权重，再通过识别目标对象的类别来获取反映目标对象类别重要性的类别权重，进一步地将目标对象的位置权重与类别权重进行结合，得到反映目标对象的综合重要性的综合权重，最终在考虑各对象的综合权重的情况下，有侧重地评估待检测图像的图像质量，使得图像质量评估过程更加合理，所得到反映图像质量的图像质量分更加准确，提升图像质量评估的准确性。
另外，本申请实施例设计了一个位置权重模型，能够准确计算图像中每个像素点的位置权重；并且通过语义分割模型，将待检测图像按照类别分割为不同区域，并计算每个区域的类别权重，保证确定类别权重的准确性；最终综合位置权重和类别权重，并结合传统的PSNR/SSIM等方法，计算最终的图像质量分数，进一步保证了得到的图像质量分的可靠性。
在示例性实施例中，图像质量分具有多种用途，下面对此进行介绍。
不同的图像处理过程，可以获得不同的质量的图像。这些图像处理过程包括摄像系统、图像信号处理、存储、传输、压缩等。
图像质量评估(IQA)是对图像质量进行经验性地衡量的方式，包括主观和客观两大类方法。主观的方法基于人类视觉对图像中的物体或属性的感知，得出图像好坏程度的判断；客观的方法则基于预设的计算模型得出定量的数值。本申请提供的技术方案主要涉及客观的图像质量评估方法，即设计一种计算模型，用于计算图像质量分数。
图像质量评估可以应用于很多不同的领域，且各领域使用图像质量评估的方法和目的各有不同。例如，评测机构用于评估不同摄像机拍摄图片的优劣；机器学习模型训练过程中，用于指导算法收敛方向等。
本申请实施例提供的图像质量评估方法所输出的图像质量分可用于衡量该图像是否适用于自动驾驶任务。在一种可能的实施方式中，基于图像质量分，对图像处理算法进行调整。 例如，对图像压缩算法进行调整。
下面结合相关技术，对本申请提供的技术方案所产生的有益效果进行进一步的说明。
在一种相关技术中，通过峰值信噪比或者结构相似性度量去衡量两张图片的相似程度，无论是用峰值信噪比或者结构相似性度量，其方案都需要计算输入图片与基准图片之间的差别或者距离，即全参考的图像质量评估(FR-IQA)。基准图片一般是未压缩的原始图片，距离越大，说明输入图片的质量越差。此类技术的缺点大致反应在如下两个方面：
一是计算的内容与任务无关。这两种方法都属于直接利用图像进行计算，其计算过程与计算内容都与具体任务无关，因此无法反映出图像质量对某一种实际任务产生的影响，例如自动驾驶任务。
二是计算的内容不能代表真实情况。一方面，公式中包含了过多与场景或者任务无关的内容，例如PSNR计算了所有像素的MSE，但事实上，并非所有的像素都会对任务有同等价值的影响；另一方面，公式的计算过程舍弃了部分图像的信息，导致结果偏离实际情况，例如PSNR计算了单个像素点的MSE，但事实上对于图像而言，像素点之间有紧密联系，且物体的大小和形状都体现在这些像素点的关系之中，因此PSNR的计算丢弃了这些重要的信息。
以自动驾驶场景为例，本申请提供的IQA评估方法适用于自动驾驶任务，该方法可以评估在自动驾驶领域中输入图像的质量情况，并且算法简单易实现。IQA通常是作为图像处理算法(如图像压缩)的评价或者神经网络的损失函数来使用，需要避免算法太复杂或者难以实现。因此，本申请设计地是一种简单、高效、通用的IQA算法，适用于多种应用场景，权重可以根据具体的应用场景进行实现。
在示例性实施例中，还可以根据如下信息增加权重值，例如车辆速度信息，行人运动速度和方向，行人到车道的距离等反映某一对象的运动状态的数据指标。请参考图9，上述步骤210之后还包括如下步骤：
步骤260，检测目标对象在待检测图像中所属实体的运动状态。
可选地，通过选取待检测图像的前几帧图像或者前一帧图像，与待检测图像进行对比，得到目标对象在待检测图像中所属实体的运动状态。例如，行人若是处于步行的运动状态，那么行人的位置信息就会发生变化，可通过与之前的行人位置进行对比，确定当前行人处于运动的状态。
步骤270，判断目标对象在待检测图像中所属实体的运动状态是否符合条件。若是，则执行步骤280；否则，执行步骤240。
步骤280，确定目标对象的状态权重。
在实体的运动状态符合条件的情况下，确定目标对象的状态权重，状态权重用于表征待检测图像中目标对象在不同运动状态下的相对重要程度。
步骤290，将目标对象的状态权重与目标对象的类别权重进行结合，得到目标对象的状态类别权重，状态类别权重用于表征不同类别的目标对象在不同运动状态下的相对重要程度。
在一个示例中，状态类别权重可以表示为W object，可通过如下公式进行计算：
W object＝W o*W enhance
其中，W o表示为目标对象所属类别预设的类别权重，W enhance表示根据目标对象的运动 状态增加的状态权重，例如根据车辆或行人状态增加的权重系数。
相应地，上述步骤240可以替换为如下步骤291：
步骤291，根据目标对象的位置权重与状态类别权重，得到目标对象的综合权重。
在目标对象具有状态类别权重的情况下，根据目标对象的位置权重与状态类别权重，得到目标对象的综合权重。
此种情况下，位置权重与状态类别权重的结合办法可以参考上文中位置权重与类别权重之间的结合方法，这里不再赘述。
在目标对象不具有状态类别权重的情况下，执行根据目标对象的位置权重与类别权重，得到目标对象的综合权重的步骤。
在一个示例中，如图10所示，其示例性示出了本申请实施例提供的图像处理流程的示意图，用于评价压缩算法是否对自动驾驶任务产生影响。图中包括图像处理流程中的图像输入环节、位置权重计算环节、语义分割环节、类别权重计算环节、全局图像质量分数计算环节以及加权图像质量分数计算环节。其中，图像输入环节，用于获取待评估图像和参考图像。位置权重计算环节，用于计算图像的每个像素点的位置权重。语义分割环节，利用语义分割模型，获得图像中不同像素的类别。类别权重计算环节，根据语义分割模型输出的类别结果和预设的不同类别的权重，获得每个像素点对应的类别权重值。全局图像质量分数计算环节，利用输入图像和参考图像，计算全局图像质量分数。可用方法包括PSNR和SSIM。加权图像质量分数计算环节，综合位置权重和类别权重，获得每个像素的综合权重，并结合全局图像质量分数计算最终的图像质量分。
综上所述，本申请实施例提供的技术方案，除了获取目标对象反映目标对象位置重要性的位置权重以及反映目标对象类别重要性的类别权重之外，还根据目标对象的运动状态为目标对象增加状态权重，进而将三种权重合理地进行结合，得到从多方面反映目标对象的综合重要性的综合权重，更加科学可靠地计算目标对象的权重，进一步保证图像质量评估的准确性。
请参考图11，其示出了本申请一个实施例提供的图像处理方法的流程图。该方法可以包括以下几个步骤(1～14)：
步骤1，获取待检测图像。
步骤2，检测待检测图像是否为俯视图。若是，则执行步骤3；否则，执行步骤5。
步骤3，对前视图或侧视图进行图像变换处理，得到与前视图或侧视图对应的俯视图。
在待检测图像是前视图或侧视图的情况下，对前视图或侧视图进行图像变换处理，得到与前视图或侧视图对应的俯视图。
步骤4，基于前视图或侧视图与俯视图之间的像素映射关系，获取前视图或侧视图中目标像素在俯视图中的位置坐标。
步骤5，根据位置权重曲线以及目标像素在俯视图中的位置坐标，获取目标像素的位置权重。
位置权重曲线通过描述像素位置与权重之间的数学关系来反映像素位置的相对重要程度。
上述过程可以简单理解为，如果输入图像是前视图或者侧视图，将其转换为俯视图；在一种可能的设计中，利用如下公式，计算俯视图中不同像素点对应的位置权重。
Figure PCTCN2020133674-appb-000006
其中，a、b为曲线形状相关参数。P 0(x0，y0)为自车中心点。△x＝x-x0,△y＝y-y0表示P(x，y)点与P 0的距离。可选地，上述公式可以应用于上述俯视图中。
最后将位置权重映射回前视图或俯视图，获得原始输入图像的每个像素的位置权重值w d。
待检测图像中各像素的位置权重，与待检测图像中各像素在俯视图中对应的像素的位置权重相等。
在一个示例中，如图12所示，其示例性示出了一种自动驾驶任务中前视图与俯视图的示意图。其中，图12(a)是当前行驶车辆前方的道路前视图，图12(b)是由图12(a)所示的前视图进行图像变换得到的俯视图，通过上述公式可以得到俯视图中每个像素的位置权重值，基于前视图与俯视图之间的变换关系，得到前视图中每个像素的位置权重值。在俯视图中距离自车中心n越近的像素点权重越高，其在前视图中对应的位置越在自车的正前方，例如图中他车中心n-1的权重值就需要根据自车中心n与他车中心n-1之间的距离进行计算。
出于精准检测图片质量的考虑，车辆的正前方无论是否有物体，都需要保证此区域的图片质量，远处图片质量可以较低，设置权重值的目的也在于有侧重地关注图像质量，更加关注与车辆附近区域的图像质量，使得图像质量评价更加合理。
步骤6，对待检测图像进行语义分割处理，得到待检测图像中各像素的类别信息。
步骤7，基于目标像素的类别信息，在类别权重数据结构中进行查找，得到目标像素的类别权重。
步骤8，检测目标对象在待检测图像中所属实体的运动状态。
步骤9，判断目标对象在待检测图像中所属实体的运动状态是否符合条件。若是，则执行步骤10；否则，执行步骤13。
步骤10，确定目标对象的状态权重。
步骤11，将目标对象的状态权重与目标对象的类别权重进行结合，得到目标对象的状态类别权重。
步骤12，根据目标对象的位置权重与状态类别权重，得到目标对象的综合权重。
步骤13，根据目标对象的位置权重与类别权重，得到目标对象的综合权重。
步骤14，基于待检测图像中各对象的综合权重，确定待检测图像的图像质量分。
综上所述，本申请实施例提供的技术方案，以像素为单位进行图像质量检测，将获取到的目标像素的位置权重、类别权重以及状态权重进行结合，得到目标像素的综合权重之后遍历所有像素，便可得到待检测图像中各像素的综合权重，进而以像素为单位有侧重地计算待检测图像的图像质量分，进一步保证了图像质量评估的准确性。
下述为本申请装置实施例，可用于执行本申请方法实施例。对于本申请装置实施例中未披露的细节，请参照本申请方法实施例。
请参考图13，其示出了本申请一个实施例提供的图像处理装置的框图。该装置具有实现上述图像处理方法的功能。该装置1300可以包括：待测图像获取模块1310、位置权重确定模块1320、类别权重确定模块1330、综合权重确定模块1340以及图像质量检测模块1350。
待测图像获取模块1310，用于获取待检测图像。
位置权重确定模块1320，用于基于所述待检测图像中目标对象的位置信息，获取所述目标对象的位置权重，所述位置权重用于反映所述目标对象在所述待检测图像中所占据位置的相对重要程度。
类别权重确定模块1330，用于基于所述目标对象的类别信息，获取所述目标对象的类别权重，所述类别权重用于反映所述目标对象所属类别在所述待检测图像中的相对重要程度。
综合权重确定模块1340，用于根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重。
图像质量检测模块1350，用于基于所述待检测图像中各对象的综合权重，确定所述待检测图像的图像质量分。
在示例性实施例中，请参考图14，所述位置权重确定模块1320包括：位置信息获取单元1321以及位置权重确定单元1322。
位置信息获取单元1321，用于获取所述待检测图像中所述目标对象的位置信息，所述位置信息包括所述目标对象在所述待检测图像描述的物理空间中的位置。
位置权重确定单元1322，用于根据位置权重关系以及所述目标对象的位置信息，得到所述目标对象的位置权重，所述位置权重关系反映位置与权重之间的数学对应关系。
在示例性实施例中，请参考图14，所述目标对象包括目标像素，所述位置信息获取单元1321用于：
在所述待检测图像是前视图或侧视图的情况下，对所述前视图或所述侧视图进行图像变换处理，得到与所述前视图或所述侧视图对应的俯视图；
基于所述前视图或所述侧视图与所述俯视图之间的像素映射关系，获取所述前视图或所述侧视图中所述目标像素在所述俯视图中的位置坐标。
在示例性实施例中，请参考图14，所述位置权重确定单元1322用于：
根据位置权重曲线以及所述目标像素在所述俯视图中的位置坐标，获取所述目标像素的位置权重，所述位置权重曲线通过描述像素位置与权重之间的数学关系来反映所述像素位置的相对重要程度。
在示例性实施例中，请参考图14，所述类别权重确定模块1330包括：类别信息识别单元1331以及类别权重确定单元1332。
类别信息识别单元1331，用于识别所述待检测图像中所述目标对象的类别信息，所述类别信息包括所述目标对象在所述待检测图像中所属实体的类别。
类别权重确定单元1332，用于根据类别权重关系以及所述目标对象的类别信息，得到所述目标对象的类别权重，所述类别权重关系反映类别与权重之间的数学对应关系。
在示例性实施例中，请参考图14，所述类别信息识别单元1331用于：
对所述待检测图像进行语义分割处理，得到所述待检测图像中各对象的类别信息。
所述类别权重确定单元1332用于：
基于所述目标对象的类别信息，在类别权重数据结构中进行查找，得到所述目标对象的类别权重。
在示例性实施例中，请参考图14，所述装置1300还包括：运动状态检测模块1360、状态权重确定模块1370以及状态类别权重确定模块1380。
运动状态检测模块1360，用于检测所述目标对象在所述待检测图像中所属实体的运动状 态。
状态权重确定模块1370，用于在所述实体的运动状态符合条件的情况下，确定所述目标对象的状态权重，所述状态权重用于表征所述待检测图像中所述目标对象在不同运动状态下的相对重要程度。
状态类别权重确定模块1380，用于将所述目标对象的状态权重与所述目标对象的类别权重进行结合，得到所述目标对象的状态类别权重，所述状态类别权重用于表征不同类别的目标对象在不同运动状态下的相对重要程度。
所述综合权重确定模块1340，用于在所述目标对象具有状态类别权重的情况下，根据所述目标对象的位置权重与状态类别权重，得到所述目标对象的综合权重。
所述综合权重确定模块1340，还用于在所述目标对象不具有状态类别权重的情况下，执行所述根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重的步骤。
在示例性实施例中，所述综合权重确定模块1340用于：
对所述位置权重与所述类别权重进行加权求和处理，得到所述目标像素的综合权重。
在示例性实施例中，请参考图14，所述目标对象包括目标像素，所述待检测图像是压缩图像；所述图像质量检测模块1350包括：像素差获取单元1351、加权相似度计算单元1352以及质量分评估单元1353。
像素差获取单元1351，用于获取所述待检测图像与参考图像在各像素上的像素差值，所述参考图像是指所述待检测图像未经压缩的原始图像。
加权相似度计算单元1352，用于基于所述各像素的综和权重与所述各像素上的像素差值，得到所述待检测图像与所述参考图像之间的加权相似度，所述加权相似度是指基于所述综和权重计算的峰值信噪比和/或结构相似度，所述峰值信噪比和/或结构相似度用于反映所述待检测图像与所述参考图像之间的差异程度。
质量分评估单元1353，用于基于所述加权相似度，得到所述待检测图像的图像质量分。
在示例性实施例中，请参考图14，所述装置1300还包括：算法调整模块1390。
算法调整模块1390，用于基于所述图像质量分，对图像处理算法进行调整。
在示例性实施例中，所述目标对象包括以下任意一种：目标像素、目标图像块、目标实体。
综上所述，本申请实施例提供的技术方案，通过检测目标对象在待检测图像中的位置，从而得到反映目标对象位置重要性的位置权重，再通过识别目标对象的类别来获取反映目标对象类别重要性的类别权重，进一步地将目标对象的位置权重与类别权重进行结合，得到反映目标对象的综合重要性的综合权重，最终在考虑各对象的综合权重的情况下，有侧重地评估待检测图像的图像质量，使得图像质量评估过程更加合理，所得到反映图像质量的图像质量分更加准确，提升图像质量评估的准确性。
另外，本申请实施例设计了一个位置权重模型，能够准确计算图像中每个像素点的位置权重；并且通过语义分割模型，将待检测图像按照类别分割为不同区域，并计算每个区域的类别权重，保证确定类别权重的准确性；最终综合位置权重和类别权重，并结合传统的PSNR/SSIM等方法，计算最终的图像质量分数，进一步保证了得到的图像质量分的可靠性。
需要说明的是，上述实施例提供的装置，在实现其功能时，仅以上述各功能模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能模块完成，即将 设备的内部结构划分成不同的功能模块，以完成以上描述的全部或者部分功能。另外，上述实施例提供的装置与方法实施例属于同一构思，其具体实现过程详见方法实施例，这里不再赘述。
请参考图15，其示出了本申请一个实施例提供的计算机设备1500的结构框图。该计算机设备1500可以是诸如手机、平板电脑、多媒体播放设备、可穿戴设备、PC(Personal Computer)、语言学习终端、智能教学机等电子设备。该计算机设备用于实施上述实施例中提供的图像处理方法。该计算机设备可以是图1所示应用程序运行环境中的终端10或者服务器20。
通常，计算机设备1500包括有：处理器1501和存储器1502。
处理器1501可以包括一个或多个处理核心，比如4核心处理器、8核心处理器等。处理器1501可以采用DSP(Digital Signal Processing，数字信号处理)、FPGA(Field Programmable Gate Array，现场可编程门阵列)、PLA(Programmable Logic Array，可编程逻辑阵列)中的至少一种硬件形式来实现。处理器1501也可以包括主处理器和协处理器，主处理器是用于对在唤醒状态下的数据进行处理的处理器，也称CPU(Central Processing Unit，中央处理器)；协处理器是用于对在待机状态下的数据进行处理的低功耗处理器。在一些实施例中，处理器1501可以在集成有GPU(Graphics Processing Unit，图像处理器)，GPU用于负责显示屏所需要显示的内容的渲染和绘制。一些实施例中，处理器1501还可以包括AI(Artificial Intelligence，人工智能)处理器，该AI处理器用于处理有关机器学习的计算操作。
存储器1502可以包括一个或多个计算机可读存储介质，该计算机可读存储介质可以是非暂态的。存储器1502还可包括高速随机存取存储器，以及非易失性存储器，比如一个或多个磁盘存储设备、闪存存储设备。在一些实施例中，存储器1502中的非暂态的计算机可读存储介质用于存储至少一个指令，至少一段程序、代码集或指令集，所述至少一条指令、至少一段程序、代码集或指令集，且经配置以由一个或者一个以上处理器执行，以实现上述图像处理方法。
在一些实施例中，计算机设备1500还可选包括有：外围设备接口1503和至少一个外围设备。处理器1501、存储器1502和外围设备接口1503之间可以通过总线或信号线相连。各个外围设备可以通过总线、信号线或电路板与外围设备接口1503相连。
本领域技术人员可以理解，图15中示出的结构并不构成对计算机设备1500的限定，可以包括比图示更多或更少的组件，或者组合某些组件，或者采用不同的组件布置。
在示例性实施例中，还提供了一种计算机可读存储介质，所述存储介质中存储有计算机程序，所述计算机程序在被处理器执行时以实现上述图像处理方法。
可选地，该计算机可读存储介质可以包括：ROM(Read Only Memory，只读存储器)、RAM(Random Access Memory，随机存取记忆体)、SSD(Solid State Drives，固态硬盘)或光盘等。其中，随机存取记忆体可以包括ReRAM(Resistance Random Access Memory,电阻式随机存取记忆体)和DRAM(Dynamic Random Access Memory，动态随机存取存储器)。
在示例性实施例中，还提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处 理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述图像处理方法。
应当理解的是，在本文中提及的“多个”是指两个或两个以上。“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。字符“/”一般表示前后关联对象是一种“或”的关系。另外，本文中描述的步骤编号，仅示例性示出了步骤间的一种可能的执行先后顺序，在一些其它实施例中，上述步骤也可以不按照编号顺序来执行，如两个不同编号的步骤同时执行，或者两个不同编号的步骤按照与图示相反的顺序执行，本申请实施例对此不作限定。
以上所述仅为本申请的示例性实施例，并不用以限制本申请，凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的保护范围之内。

Claims (24)

  1. 一种图像处理方法，其特征在于，所述方法包括：
    获取待检测图像；
    基于所述待检测图像中目标对象的位置信息，获取所述目标对象的位置权重，所述位置权重用于反映所述目标对象在所述待检测图像中所占据位置的相对重要程度；
    基于所述目标对象的类别信息，获取所述目标对象的类别权重，所述类别权重用于反映所述目标对象所属类别在所述待检测图像中的相对重要程度；
    根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重；
    基于所述待检测图像中各对象的综合权重，确定所述待检测图像的图像质量分。
  2. 根据权利要求1所述的方法，其特征在于，所述基于所述待检测图像中目标对象的位置信息，获取所述目标对象的位置权重，包括：
    获取所述待检测图像中所述目标对象的位置信息，所述位置信息包括所述目标对象在所述待检测图像描述的物理空间中的位置；
    根据位置权重关系以及所述目标对象的位置信息，得到所述目标对象的位置权重，所述位置权重关系反映位置与权重之间的数学对应关系。
  3. 根据权利要求2所述的方法，其特征在于，所述目标对象包括目标像素，所述获取所述待检测图像中目标对象的位置信息，包括：
    在所述待检测图像是前视图或侧视图的情况下，对所述前视图或所述侧视图进行图像变换处理，得到与所述前视图或所述侧视图对应的俯视图；
    基于所述前视图或所述侧视图与所述俯视图之间的像素映射关系，获取所述前视图或所述侧视图中所述目标像素在所述俯视图中的位置坐标。
  4. 根据权利要求3所述的方法，其特征在于，所述根据位置权重关系以及所述目标对象的位置信息，得到所述目标对象的位置权重，包括：
    根据位置权重曲线以及所述目标像素在所述俯视图中的位置坐标，获取所述目标像素的位置权重，所述位置权重曲线通过描述像素位置与权重之间的数学关系来反映所述像素位置的相对重要程度。
  5. 根据权利要求1所述的方法，其特征在于，所述基于所述目标对象的类别信息，获取所述目标对象的类别权重，包括：
    识别所述待检测图像中所述目标对象的类别信息，所述类别信息包括所述目标对象在所述待检测图像中所属实体的类别；
    根据类别权重关系以及所述目标对象的类别信息，得到所述目标对象的类别权重，所述类别权重关系反映类别与权重之间的数学对应关系。
  6. 根据权利要求5所述的方法，其特征在于，所述识别所述待检测图像中所述目标对象的类别信息，包括：
    对所述待检测图像进行语义分割处理，得到所述待检测图像中各对象的类别信息；
    所述根据类别权重关系以及所述目标对象的类别信息，得到所述目标对象的类别权重，包括：
    基于所述目标对象的类别信息，在类别权重数据结构中进行查找，得到所述目标对象的类别权重。
  7. 根据权利要求1所述的方法，其特征在于，所述获取待检测图像之后，还包括：
    检测所述目标对象在所述待检测图像中所属实体的运动状态；
    在所述实体的运动状态符合条件的情况下，确定所述目标对象的状态权重，所述状态权重用于表征所述待检测图像中所述目标对象在不同运动状态下的相对重要程度；
    将所述目标对象的状态权重与所述目标对象的类别权重进行结合，得到所述目标对象的状态类别权重，所述状态类别权重用于表征不同类别的目标对象在不同运动状态下的相对重要程度；
    在所述目标对象具有状态类别权重的情况下，根据所述目标对象的位置权重与状态类别权重，得到所述目标对象的综合权重；
    在所述目标对象不具有状态类别权重的情况下，执行所述根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重的步骤。
  8. 根据权利要求1所述的方法，其特征在于，所述根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重，包括：
    对所述位置权重与所述类别权重进行加权求和处理，得到所述目标像素的综合权重。
  9. 根据权利要求1至8任一项所述的方法，其特征在于，所述目标对象包括目标像素，所述待检测图像是压缩图像；
    所述基于所述待检测图像中各对象的综和权重，确定所述待检测图像的图像质量分，包括：
    获取所述待检测图像与参考图像在各像素上的像素差值，所述参考图像是指所述待检测图像未经压缩的原始图像；
    基于所述各像素的综和权重与所述各像素上的像素差值，得到所述待检测图像与所述参考图像之间的加权相似度，所述加权相似度是指基于所述综和权重计算的峰值信噪比和/或结构相似度，所述峰值信噪比和/或结构相似度用于反映所述待检测图像与所述参考图像之间的差异程度；
    基于所述加权相似度，得到所述待检测图像的图像质量分。
  10. 根据权利要求1至8任一项所述的方法，其特征在于，所述基于所述待检测图像中各对象的综合权重，确定所述待检测图像的图像质量分之后，还包括：
    基于所述图像质量分，对图像处理算法进行调整。
  11. 根据权利要求1至8任一项所述的方法，其特征在于，所述目标对象包括以下任意一种：目标像素、目标图像块、目标实体。
  12. 一种图像处理装置，其特征在于，所述装置包括：
    待测图像获取模块，用于获取待检测图像；
    位置权重确定模块，用于基于所述待检测图像中目标对象的位置信息，获取所述目标对象的位置权重，所述位置权重用于反映所述目标对象在所述待检测图像中所占据位置的相对重要程度；
    类别权重确定模块，用于基于所述目标对象的类别信息，获取所述目标对象的类别权重，所述类别权重用于反映所述目标对象所属类别在所述待检测图像中的相对重要程度；
    综合权重确定模块，用于根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重；
    图像质量检测模块，用于基于所述待检测图像中各对象的综合权重，确定所述待检测图像的图像质量分。
  13. 根据权利要求12所述的装置，其特征在于，所述位置权重确定模块包括：
    位置信息获取单元，用于获取所述待检测图像中所述目标对象的位置信息，所述位置信息包括所述目标对象在所述待检测图像描述的物理空间中的位置；
    位置权重确定单元，用于根据位置权重关系以及所述目标对象的位置信息，得到所述目标对象的位置权重，所述位置权重关系反映位置与权重之间的数学对应关系。
  14. 根据权利要求13所述的装置，其特征在于，所述目标对象包括目标像素，所述位置信息获取单元用于：
    在所述待检测图像是前视图或侧视图的情况下，对所述前视图或所述侧视图进行图像变换处理，得到与所述前视图或所述侧视图对应的俯视图；
    基于所述前视图或所述侧视图与所述俯视图之间的像素映射关系，获取所述前视图或所述侧视图中所述目标像素在所述俯视图中的位置坐标。
  15. 根据权利要求14所述的装置，其特征在于，所述位置权重确定单元用于：
    根据位置权重曲线以及所述目标像素在所述俯视图中的位置坐标，获取所述目标像素的位置权重，所述位置权重曲线通过描述像素位置与权重之间的数学关系来反映所述像素位置的相对重要程度。
  16. 根据权利要求12所述的装置，其特征在于，所述类别权重确定模块包括：
    类别信息识别单元，用于识别所述待检测图像中所述目标对象的类别信息，所述类别信息包括所述目标对象在所述待检测图像中所属实体的类别；
    类别权重确定单元，用于根据类别权重关系以及所述目标对象的类别信息，得到所述目标对象的类别权重，所述类别权重关系反映类别与权重之间的数学对应关系。
  17. 根据权利要求16所述的装置，其特征在于，所述类别信息识别单元用于：
    对所述待检测图像进行语义分割处理，得到所述待检测图像中各对象的类别信息；
    所述类别权重确定单元用于：
    基于所述目标对象的类别信息，在类别权重数据结构中进行查找，得到所述目标对象的类别权重。
  18. 根据权利要求12所述的装置，其特征在于，所述装置还包括：
    运动状态检测模块，用于检测所述目标对象在所述待检测图像中所属实体的运动状态；
    状态权重确定模块，用于在所述实体的运动状态符合条件的情况下，确定所述目标对象的状态权重，所述状态权重用于表征所述待检测图像中所述目标对象在不同运动状态下的相对重要程度；
    状态类别权重确定模块，用于将所述目标对象的状态权重与所述目标对象的类别权重进行结合，得到所述目标对象的状态类别权重，所述状态类别权重用于表征不同类别的目标对象在不同运动状态下的相对重要程度；
    所述综合权重确定模块，用于在所述目标对象具有状态类别权重的情况下，根据所述目标对象的位置权重与状态类别权重，得到所述目标对象的综合权重；
    所述综合权重确定模块，还用于在所述目标对象不具有状态类别权重的情况下，执行所述根据所述目标对象的位置权重与类别权重，得到所述目标对象的综合权重的步骤。
  19. 根据权利要求12所述的装置，其特征在于，所述综合权重确定模块用于：
    对所述位置权重与所述类别权重进行加权求和处理，得到所述目标像素的综合权重。
  20. 根据权利要求12至19任一项所述的装置，其特征在于，所述目标对象包括目标像素，所述待检测图像是压缩图像；
    所述图像质量检测模块包括：
    像素差获取单元，用于获取所述待检测图像与参考图像在各像素上的像素差值，所述参考图像是指所述待检测图像未经压缩的原始图像；
    加权相似度计算单元，用于基于所述各像素的综和权重与所述各像素上的像素差值，得到所述待检测图像与所述参考图像之间的加权相似度，所述加权相似度是指基于所述综和权重计算的峰值信噪比和/或结构相似度，所述峰值信噪比和/或结构相似度用于反映所述待检测图像与所述参考图像之间的差异程度；
    质量分评估单元，用于基于所述加权相似度，得到所述待检测图像的图像质量分。
  21. 根据权利要求12至19任一项所述的装置，其特征在于，所述装置还包括：
    算法调整模块，用于基于所述图像质量分，对图像处理算法进行调整。
  22. 根据权利要求12至19任一项所述的装置，其特征在于，所述目标对象包括以下任意一种：目标像素、目标图像块、目标实体。
  23. 一种计算机设备，其特征在于，所述计算机设备包括处理器和存储器，所述存储器中存储有计算机程序，所述计算机程序由所述处理器加载并执行以实现如权利要求1至11任一项所述的方法。
  24. 一种计算机可读存储介质，其特征在于，所述存储介质中存储有计算机程序，所述计算机程序由处理器加载并执行以实现如权利要求1至11任一项所述的方法。
PCT/CN2020/133674 2020-12-03 2020-12-03 图像处理方法、装置、设备及存储介质 WO2022116104A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202080004518.6A CN112703532B (zh) 2020-12-03 2020-12-03 图像处理方法、装置、设备及存储介质
PCT/CN2020/133674 WO2022116104A1 (zh) 2020-12-03 2020-12-03 图像处理方法、装置、设备及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/133674 WO2022116104A1 (zh) 2020-12-03 2020-12-03 图像处理方法、装置、设备及存储介质

Publications (1)

Publication Number Publication Date
WO2022116104A1 true WO2022116104A1 (zh) 2022-06-09

Family

ID=75514813

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/133674 WO2022116104A1 (zh) 2020-12-03 2020-12-03 图像处理方法、装置、设备及存储介质

Country Status (2)

Country Link
CN (1) CN112703532B (zh)
WO (1) WO2022116104A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115620030A (zh) * 2022-12-06 2023-01-17 浙江正泰智维能源服务有限公司 一种图像匹配方法、装置、设备、介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113191266B (zh) * 2021-04-30 2021-10-22 江苏航运职业技术学院 船舶动力装置远程监控管理方法及系统
CN113364963B (zh) * 2021-08-09 2021-11-16 浙江大华技术股份有限公司 成像控制方法及相关设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108010060A (zh) * 2017-12-06 2018-05-08 北京小米移动软件有限公司 目标检测方法及装置
CN108550259A (zh) * 2018-04-19 2018-09-18 何澜 道路拥堵判断方法、终端设备及计算机可读存储介质
CN108829826A (zh) * 2018-06-14 2018-11-16 清华大学深圳研究生院 一种基于深度学习和语义分割的图像检索方法
US20200117952A1 (en) * 2018-10-11 2020-04-16 James Carroll Target object position prediction and motion tracking
CN111476806A (zh) * 2020-06-23 2020-07-31 腾讯科技（深圳）有限公司 图像处理方法、装置、计算机设备和存储介质
CN111696112A (zh) * 2020-06-15 2020-09-22 携程计算机技术（上海）有限公司 图像自动裁剪方法、系统、电子设备及存储介质

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101853504B (zh) * 2010-05-07 2012-04-25 厦门大学 基于视觉特性与结构相似度的图像质量评测方法
CN101976444B (zh) * 2010-11-11 2012-02-08 浙江大学 一种基于像素类型的结构类似性图像质量客观评价方法
CN102722888A (zh) * 2012-05-22 2012-10-10 天津大学 基于生理与心理立体视觉的立体图像客观质量评价方法
JP6100064B2 (ja) * 2013-04-10 2017-03-22 株式会社東芝 電子機器および画像処理方法
CN104834898B (zh) * 2015-04-09 2018-05-15 华南理工大学 一种人物摄影图像的质量分类方法
CN106202089B (zh) * 2015-05-04 2020-03-27 阿里巴巴集团控股有限公司 一种确定图片质量和网页展示的方法及设备
CN109345552A (zh) * 2018-09-20 2019-02-15 天津大学 基于区域权重的立体图像质量评价方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108010060A (zh) * 2017-12-06 2018-05-08 北京小米移动软件有限公司 目标检测方法及装置
CN108550259A (zh) * 2018-04-19 2018-09-18 何澜 道路拥堵判断方法、终端设备及计算机可读存储介质
CN108829826A (zh) * 2018-06-14 2018-11-16 清华大学深圳研究生院 一种基于深度学习和语义分割的图像检索方法
US20200117952A1 (en) * 2018-10-11 2020-04-16 James Carroll Target object position prediction and motion tracking
CN111696112A (zh) * 2020-06-15 2020-09-22 携程计算机技术（上海）有限公司 图像自动裁剪方法、系统、电子设备及存储介质
CN111476806A (zh) * 2020-06-23 2020-07-31 腾讯科技（深圳）有限公司 图像处理方法、装置、计算机设备和存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115620030A (zh) * 2022-12-06 2023-01-17 浙江正泰智维能源服务有限公司 一种图像匹配方法、装置、设备、介质
CN115620030B (zh) * 2022-12-06 2023-04-18 浙江正泰智维能源服务有限公司 一种图像匹配方法、装置、设备、介质

Also Published As

Publication number Publication date
CN112703532B (zh) 2022-05-31
CN112703532A (zh) 2021-04-23

Similar Documents

Publication Publication Date Title
WO2022116104A1 (zh) 图像处理方法、装置、设备及存储介质
CN108895981B (zh) 一种三维测量方法、装置、服务器和存储介质
US11830230B2 (en) Living body detection method based on facial recognition, and electronic device and storage medium
CN110222787B (zh) 多尺度目标检测方法、装置、计算机设备及存储介质
US20230214976A1 (en) Image fusion method and apparatus and training method and apparatus for image fusion model
CN113065558A (zh) 一种结合注意力机制的轻量级小目标检测方法
WO2022156640A1 (zh) 一种图像的视线矫正方法、装置、电子设备、计算机可读存储介质及计算机程序产品
CN111476827B (zh) 目标跟踪方法、系统、电子装置及存储介质
CN110532970B (zh) 人脸2d图像的年龄性别属性分析方法、系统、设备和介质
CN112308095A (zh) 图片预处理及模型训练方法、装置、服务器及存储介质
CN113052835B (zh) 一种基于三维点云与图像数据融合的药盒检测方法及其检测系统
CN110189294B (zh) 基于深度可信度分析的rgb-d图像显著性检测方法
CN108470178B (zh) 一种结合深度可信度评价因子的深度图显著性检测方法
CN110825900A (zh) 特征重构层的训练方法、图像特征的重构方法及相关装置
CN112036339B (zh) 人脸检测的方法、装置和电子设备
CN111784658B (zh) 一种用于人脸图像的质量分析方法和系统
CN112149476A (zh) 目标检测方法、装置、设备和存储介质
CN111914938A (zh) 一种基于全卷积二分支网络的图像属性分类识别方法
CN111753671A (zh) 一种现实场景的人群计数方法
WO2022247126A1 (zh) 视觉定位方法、装置、设备、介质及程序
CN114581318A (zh) 一种低照明度图像增强方法及系统
WO2020087434A1 (zh) 一种人脸图像清晰度评价方法及装置
Babu et al. An efficient image dahazing using Googlenet based convolution neural networks
CN113762009B (zh) 一种基于多尺度特征融合及双注意力机制的人群计数方法
Zheng et al. Overwater image dehazing via cycle-consistent generative adversarial network

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20963950

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20963950

Country of ref document: EP

Kind code of ref document: A1