CN111435986A - Method for acquiring source image database, training device and electronic equipment - Google Patents

Method for acquiring source image database, training device and electronic equipment Download PDF

Info

Publication number
CN111435986A
CN111435986A CN201911336566.0A CN201911336566A CN111435986A CN 111435986 A CN111435986 A CN 111435986A CN 201911336566 A CN201911336566 A CN 201911336566A CN 111435986 A CN111435986 A CN 111435986A
Authority
CN
China
Prior art keywords
image
target
library
light source
original
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201911336566.0A
Other languages
Chinese (zh)
Other versions
CN111435986B (en
Inventor
杨远飞
徐会
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhuhai Jieli Technology Co Ltd
Original Assignee
Zhuhai Jieli Technology Co Ltd
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 Zhuhai Jieli Technology Co Ltd filed Critical Zhuhai Jieli Technology Co Ltd
Priority to CN201911336566.0A priority Critical patent/CN111435986B/en
Publication of CN111435986A publication Critical patent/CN111435986A/en
Application granted granted Critical
Publication of CN111435986B publication Critical patent/CN111435986B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N9/00—Details of colour television systems
    • H04N9/64—Circuits for processing colour signals
    • H04N9/73—Colour balance circuits, e.g. white balance circuits or colour temperature control
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N3/00—Scanning details of television systems; Combination thereof with generation of supply voltages
    • H04N3/02—Scanning details of television systems; Combination thereof with generation of supply voltages by optical-mechanical means only
    • H04N3/08—Scanning details of television systems; Combination thereof with generation of supply voltages by optical-mechanical means only having a moving reflector

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Image Processing (AREA)
  • Color Television Image Signal Generators (AREA)

Abstract

The invention relates to a method and a system for acquiring a source image database, a training method and a training device of a neural network, a computer readable storage medium, an image processing chip, an image processing system and an electronic device, wherein the acquisition method comprises the following steps: firstly, respectively acquiring a known image library and a corrected image library, and respectively calculating light source information corresponding to respective images; then, establishing a color temperature model according to a plurality of groups of original images with the same color temperature points and light source information of the collected images, and converting the light source information of each original image in the known image library according to the color temperature model to obtain the light source information of each target image in the target image library; and then converting the pixel information of the original image of the known image library into the pixel information of the target image library according to the light source information of the known image library and the light source information of the target image library to obtain a source image database corresponding to the target acquisition unit. The invention greatly reduces the number of image acquisition before neural network training.

Description

Method for acquiring source image database, training device and electronic equipment
Technical Field
The invention relates to the field of image processing, in particular to a method and a system for acquiring a source image database, a method for training a neural network, a training device, a computer-readable storage medium, an image processing chip, an image processing system and electronic equipment.
Background
White balance refers to a method by which a white object can be reduced to white under any light source. The human eye is adaptive so we cannot sometimes find a change in color temperature, but the electronic device does not have this feature, so a white balance algorithm is needed to restore the color. In the prior art, many white balance processing methods are traditional algorithms, such as a gray world method, a perfect reflection method and the like, and the core ideas of the algorithms are that white blocks in an image are directly found for color restoration, but real-world scenes are complex and changeable, and once the white blocks in the scenes are few or no white blocks, the algorithms fail, and the robustness is low. A more advanced white balance processing algorithm is carried out by adopting a neural network, specifically, a target acquisition unit acquires a large number of raw images, then the neural network is trained to obtain a trained model, and the trained model is adopted to carry out white balance processing on the images acquired by the target acquisition unit.
Although the performance of the neural network method is good, the neural network method is supported by a large amount of raw image data, and the color temperature curves corresponding to different lenses or sensors are different, which means that different electronic devices need to re-acquire a large amount of raw images to retrain the neural network as long as one of the lenses and the sensors is changed, for device manufacturers, multiple devices are manufactured, the lenses and the sensors of the devices are often replaced, and even the devices of the same model are often changed, so that each device needs to re-acquire a large amount of raw images, the acquisition work before the neural network is increased sharply, and the production time and the cost of the devices are increased.
Disclosure of Invention
Based on the above situation, the present invention is directed to a method and a system for acquiring a source image database, a method for training a neural network, a training device, a computer-readable storage medium, an image processing chip, an image processing system, and an electronic device, so as to solve the problem of a large increase in workload due to the replacement of a lens and a sensor.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a first aspect of the present invention provides an acquisition method of a source image database, where the source image database is used to train a neural network, and the neural network is used to perform white balance processing on an image acquired by an acquisition unit, and the acquisition method includes the steps of:
s100: acquiring a known image library, and acquiring light source information corresponding to each original image in the known image library; the original images in the known image library cover a plurality of color temperature points, and each original image is provided with a standard color card;
s200: acquiring a correction image library, and calculating light source information corresponding to each acquired image in the correction image library; wherein the correction image library is formed by collecting collected images with standard color cards at different color temperatures by using a target collecting unit;
s300: establishing a color temperature model converted from the original image to the collected image according to the corresponding light source information in the original image and the collected image with the same color temperature points; the color temperature model comprises a conversion matrix c and a conversion coefficient z0Wherein, the product of the light source information of the original image and the conversion matrix c is equal to the product of the light source information of the collected image and the conversion coefficient;
s400: converting the light source information of each original image in the known image library according to the color temperature model to obtain the light source information of each target image in the target image library;
s500: converting the pixel information of each original image of the known image library into the pixel information of each target image of the target image library according to the light source information of the known image library and the light source information of the target image library to obtain a source image database corresponding to the target acquisition unit;
the number of the collected images of the correction image library is smaller than that of the original images of the known image library, and the image sources of the two images are different.
Preferably, the acquiring a known image library in step S100 specifically includes:
acquiring the original image with a standard color card under different scenes by using a non-target acquisition unit to form the known image library; or
The library of known images is acquired by a white balanced common dataset.
Preferably, the step S200 of forming the corrected image library by using the target collecting unit to collect the collected images with the standard color cards at different color temperatures includes: the correction image library is formed by collecting collected images of standard color cards with different standard light sources under a light box by using a target collecting unit.
Preferably, the light source information of an image comprises a luminance value/of a red channel of said imagerBrightness value l of green channelgAnd the luminance value l of the blue channelbThe specific method for obtaining the light source information in step S100 and step S200 includes the steps of:
s001: converting the image to the RGB domain;
s002: respectively calculating the pixel mean value of a red channel, the pixel mean value of a green channel and the pixel mean value of a blue channel for the white blocks which are not over-exposed on the standard color card in the image, wherein the three mean values are the brightness value l of the red channel of the imagerBrightness value l of green channelgAnd the luminance value l of the blue channelb，
Wherein the image refers to the original image or the collected image.
Preferably, the light source information further includes a label; the conversion matrix c is a three-dimensional conversion matrix; the step S300 specifically includes the steps of:
s310: respectively calculating the label x of each original image according to formula (1)1、y1And label x of the captured image2、y2；
Figure BDA0002331087190000031
Wherein lr、lg、lbThe brightness value of the red channel and the brightness value of the green channel of the original image or the collected image are respectivelyAnd a luminance value of the blue channel;
s320: selecting a plurality of groups of original images and collected images with the same color temperature points;
s330: establishing a three-dimensional transformation matrix c and transformation coefficients z0Wherein, in the step (A),
Figure BDA0002331087190000032
Figure BDA0002331087190000033
Figure BDA0002331087190000034
according to formulas (2) and (3) and the labels x of the original images corresponding to the groups1、y1And label x of the captured image2、y2Calculating the transformation matrix c and transformation coefficients z0；
The step S400 specifically includes:
according to the formulas (2) and (3), the transformation matrix c and the transformation coefficient z0Label x of each original image of the known image library1、y1Converting to obtain the label x of each target image in the target image library2、y2Wherein x in the formula (3)2、y2Is replaced by x3、y3，x′2、y′2Is replaced by x'3、y′3。
Preferably, the step S400 further includes: label x of each target image3、y3Substituting x and y in the formula (4) to obtain the brightness value l of the normalized red channel of each target imager0Brightness value l of green channelg0And the luminance value l of the blue channelb0。
Figure BDA0002331087190000041
Preferably, in the step S500, converting the pixel information of each original image in the known image library to the pixel information of each target image in the target image library according to the light source information in the known image library and the light source information in the target image library, specifically including the steps of:
s510: respectively calculating a first red channel gain and a first blue channel gain of each original image, and calculating a third red channel gain and a third blue channel gain of each target image in the target image library;
s520: and converting the pixel information of each original image into the pixel information of each target image in the target image library according to the pixel information of the original image and the first red channel gain, the first blue channel gain, the third red channel gain and the third blue channel gain.
Preferably, the step S510 specifically includes: calculating the third red channel gain r _ gain according to equation (5)3And the third blue channel gain b _ gain3Calculating the first red channel gain r _ gain according to equation (6)1And the first blue channel gain b _ gain1；
Figure BDA0002331087190000042
Figure BDA0002331087190000043
The pixel information of the original image comprises a first red pixel value r1A first blue pixel value b1(ii) a The pixel information of the target image comprises a third red pixel value r3The third blue pixel value b3(ii) a The step S520 specifically includes: calculating pixel information of each target image according to formula (7);
Figure BDA0002331087190000051
preferably, in the step S300, according to multiple sets of light source information corresponding to the original image and the collected image having the same color temperature point, the method specifically includes:
and selecting the original image with the same color temperature point from the known image library for each acquired image so as to form a plurality of groups of corresponding original images and acquired images.
Preferably, the number of the original images is greater than or equal to 100.
Preferably, the color temperature points of the light source corresponding to the original image in the known image library are distributed in a range of 1000k to 10000 k.
Preferably, the step S500 specifically includes:
converting pixel information of each original image of the known image library into pixel information of each target image of the target image library according to the light source information of the known image library and the target image library, merging a plurality of optimized images with the target image library to form an optimized target image library, and obtaining a source image database of the target acquisition unit;
wherein the optimized image is an image acquired using the target acquisition unit.
The method for acquiring the source image database provided by the invention uses a small amount of acquired images acquired by the target acquisition unit, realizes conversion of a large amount of original images in the known image library according to color temperature modeling, further forms the target image library at least equal to the number of images in the known image library, and can directly train the target image library obtained by conversion as the source image database when being applied to neural network training. Obviously, after the acquisition unit is replaced, the method can form a source image database corresponding to the target acquisition unit only by acquiring a small number of acquired images by the acquisition unit, thereby saving the acquisition time of the images before neural network training to a great extent, providing data guarantee for the training of the neural network, improving the working efficiency and reducing the production cost.
A second aspect of the present invention provides a training method for a neural network, where the neural network is configured to perform white balance processing on an image acquired by an acquisition unit, and the training method includes:
and training the initial neural network by using the source image database obtained by the acquisition method, and correcting each parameter of the initial neural network to obtain an optimized neural network corresponding to the target acquisition unit.
A third aspect of the present invention provides an acquisition system of a source image database, where the source image database is used to train a neural network, and the neural network is used to perform white balance processing on an image acquired by an acquisition unit, and the acquisition system includes:
the system comprises an acquisition module, a correction image database and a display module, wherein the acquisition module is used for acquiring a correction image database, and the correction image database acquires acquired images with standard color cards at different color temperatures through a target acquisition unit to form;
the processing module is used for acquiring a known image library and obtaining light source information corresponding to each original image in the known image library; the processing module is further used for calculating light source information corresponding to each collected image, and establishing a color temperature model converted from the original image to the collected image according to a plurality of groups of light source information corresponding to the original image and the collected image with the same color temperature point; converting the light source information of each original image in the known image library according to the color temperature model to obtain the light source information of each target image in the target image library; converting the pixel information of each original image of the known image library into the pixel information of each target image of the target image library according to the light source information of the known image library and the light source information of the target image library to obtain a source image database corresponding to the target acquisition unit;
and the number of the collected images is less than that of the original images, and the two images have different sources.
A fourth aspect of the present invention provides a training apparatus for training a neural network for white balance processing of an image acquired by an acquisition unit, the training apparatus including a training system having an initial neural network and an acquisition system of a source image database,
the acquisition system is used for acquiring a source image database by using the acquisition method as described in any one of the above items;
and the training system is connected with the acquisition system and used for training the initial neural network by using the source image database and correcting all parameters of the initial neural network to obtain an optimized neural network.
Preferably, the device further comprises an output port for outputting the optimized neural network.
A fifth aspect of the invention provides a computer readable storage medium having stored thereon a computer program which, when executed, implements the acquisition method as defined in any one of the above or implements the training method as defined above.
A sixth aspect of the present invention provides an image processing chip having a storage unit that stores the optimized neural network output by the training apparatus as described in any one of the above.
A seventh aspect of the present invention provides an image processing system, including the training apparatus described above and the image processing chip described above, wherein the storage unit stores an optimized neural network output by the training apparatus.
An eighth aspect of the present invention provides an electronic device, which includes an acquisition unit, and further includes the image processing chip as described above, where the image processing chip is connected to the acquisition unit, and is configured to perform white balance processing on an image acquired by the acquisition unit using the optimized neural network. ,
preferably, the electronic device includes at least one of a mobile terminal, a camera, a scanner, a security monitoring device, and a driving recorder.
Other advantages of the present invention will be described in the detailed description, and those skilled in the art will understand the technical features and technical solutions presented in the description.
Drawings
Preferred embodiments of the present invention will be described below with reference to the accompanying drawings. In the figure:
FIG. 1 is a flow chart of a preferred embodiment of the acquisition method provided by the present invention;
FIG. 2 is a color temperature distribution diagram of an image collected by a known image library, a target image library and a target collection unit according to the present invention;
FIG. 3 is a system diagram of a preferred embodiment of an acquisition system provided by the present invention;
fig. 4 is a system diagram of a preferred embodiment of the white balance correction apparatus provided in the present invention.
Detailed Description
The present invention will be described below based on examples, but the present invention is not limited to only these examples. In the following detailed description of the present invention, certain specific details are set forth in order to avoid obscuring the nature of the present invention, well-known methods, procedures, and components have not been described in detail.
Further, those of ordinary skill in the art will appreciate that the drawings provided herein are for illustrative purposes and are not necessarily drawn to scale.
Unless the context clearly requires otherwise, throughout the description and the claims, the words "comprise", "comprising", and the like are to be construed in an inclusive sense as opposed to an exclusive or exhaustive sense; that is, what is meant is "including, but not limited to".
In the description of the present invention, it is to be understood that the terms "first," "second," and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance. In addition, in the description of the present invention, "a plurality" means two or more unless otherwise specified.
The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The invention provides electronic equipment, such as mobile terminals (including mobile phones, pads and the like), cameras, scanners, security monitoring equipment, automobile data recorders and the like. The electronic equipment comprises an acquisition unit and an image processing chip, wherein the acquisition unit comprises a lens and a sensor which are matched with each other, the sensor can be a CCD (charge coupled device), a CMOS (complementary metal oxide semiconductor) device and the like, and the acquisition unit is used for acquiring images, namely the sensor acquires the images through the lens; the image processing chip is connected with the acquisition unit and is used for processing the image acquired by the acquisition unit.
The collecting unit is affected by different light sources when collecting images, so that the collected images have color cast compared with actual images, therefore, the collected images are often subjected to white balance processing, for example, a neural network is established, the neural network is implanted into each electronic device, and the images collected by the electronic device are subjected to white balance processing through the neural network. However, since the lens and the sensor of each device are changed, even for devices of the same model, parameters of the lens and the sensor are not necessarily the same, in order to make the neural network implanted in the electronic device process images shot by the device more perfectly, a large number of images, often more than five hundred images, need to be shot by each electronic device to form a rich source image database, then the source image database is used to train the neural network, so as to obtain a neural network suitable for the acquisition unit of the electronic device, and then the trained neural network is implanted in the electronic device. Obviously, this way of acquiring a large number of images for each electronic device increases the acquisition time before the neural network is modified.
Based on the above problems, the present invention provides a method for acquiring a source image database, where the source image database acquired by the method can directly correspond to a source image database of a target acquisition unit, and can be used to train a neural network, so as to obtain an optimized neural network corresponding to the target acquisition unit, and perform white balance processing on an image acquired by the target acquisition unit by using the optimized neural network. As shown in fig. 1, the acquisition method includes the steps of:
s100: acquiring a known image library, and acquiring light source information corresponding to each original image in the known image library; the known image library is provided with a large number of original images, the original images in the known image library cover a plurality of color temperature points so as to obtain the original images in various scenes as far as possible, and each original image is provided with a standard color card;
s200: acquiring a correction image library, and calculating light source information corresponding to each acquired image in the correction image library; the correction image library is formed by acquiring acquired images with standard color cards at different color temperatures by using a target acquisition unit, namely, the target acquisition unit is used for respectively shooting images with the standard color cards at different color temperatures, the images are called as acquired images, and the acquired images corresponding to a plurality of light sources form the correction image library; it can be understood that the color temperature points corresponding to the same light source are the same, and the collected images in the corrected image library use different light sources, and each collected image corresponds to different color temperature points;
s300: establishing a color temperature model converted from an original image to a collected image according to corresponding light source information in a plurality of groups of original images and collected images with the same color temperature points, specifically, selecting a plurality of groups of original images and collected images with the same color temperature points to form a color temperature set, wherein the color temperature set comprises a plurality of color temperature groups, the color temperature points corresponding to the original images and the collected images in each color temperature group are the same, and correspond to a light source, and then, the color temperature model is converted from the original images to the collected imagesEstablishing a color temperature model according to the light source information of the original image and the light source information of the collected image in each color temperature group; specifically, the color temperature model includes a conversion matrix c and a conversion coefficient z0The product of the light source information of the original image and the conversion matrix c is equal to the product of the light source information of the collected image and the conversion coefficient;
s400: converting the light source information of each original image in the known image library according to the color temperature model to obtain the light source information of each target image in the target image library, namely converting the light source information of all the original images in the known image library, wherein the converted light source information of each original image is the light source information of the corresponding target image in the target image library;
s500: converting the pixel information of each original image of the known image library into the pixel information of each target image of the target image library according to the light source information of the known image library and the target image library, namely converting the pixels of each known image library according to the corresponding light source information of the original image and the target image, namely correspondingly forming the pixel information of the target image, and obtaining a source image database corresponding to the target acquisition unit;
it should be noted that the number of the target images in the target image library is equal to the number of the original images in the known image library, and each target image corresponds to each original image one to one, that is, in a set of one to one corresponding target image and original image, the light source information and the pixel information of the target image are obtained by converting the original images.
The standard color card may be a 24 color card, or may be another color card used by those skilled in the art. The number of the collected images of the corrected image library is far less than that of the original images of the known image library, and the image sources of the two images are different, namely the original images of the known image library are not collected by the target collection unit. It should be noted that, the light sources corresponding to part of the original images in the known image library at least partially coincide with the light sources corresponding to the captured images, that is, the color temperature points corresponding to the captured images are at least partially equal to the color temperature points corresponding to the original images in the known image library. The target acquisition unit is not limited to the current acquisition unit, that is, is not limited to a specific acquisition unit, and in practical application, the target acquisition unit may refer to all acquisition units of the same type, wherein the acquisition units of the same type refer to acquisition units in which the sensor and the lens are the same.
It should be noted that the execution sequence of the steps S100 and S200 is not limited, and both steps are executed before the step S300, that is, the step S100 may be executed first and then the step S200 is executed, the step S200 may be executed first and then the step S100 is executed, or the step S100 and the step S200 are executed simultaneously.
According to the method for acquiring the image database, a small number of acquired images acquired by the target acquisition unit are used, a large number of original images in the known image library are converted according to color temperature modeling, and then the source image database at least equal to the number of the original images in the known image library is formed. Obviously, after the acquisition unit is replaced, the method does not need to use the target acquisition unit to acquire a large number of images, and only needs the acquisition unit to acquire a small number of acquired images, if the number of the acquired images can be three or four, the method can be formed in a source image database corresponding to the target acquisition unit, thereby reducing the acquisition time of the images before neural network training to a great extent, and providing data guarantee for the training of the neural network, thereby improving the working efficiency and reducing the production cost.
The acquiring of the known image library in step S100 specifically includes: the known image library is formed by capturing raw images with a color chart under different scenes by using a non-target capturing unit, or is obtained by a white-balanced common data set. That is, the original image can be acquired by the acquisition unit, but the acquisition unit is different from the target acquisition unit (at least one of the lens and the sensor is different), and this method needs to shoot the standard color card together when acquiring, shoot a large number of original images to form a known image library, for example, the known image library is directly derived from the source image database used by the previous acquisition unit; in another embodiment, the library of known images may be obtained directly from the white balanced common data set by a network or the like, which is more straightforward to facilitate.
Regardless of which method is used to obtain the known image library, in order to improve the effect of the source image database on the neural network training, the number of the original images of the known image library is as large as possible, for example, the number of the original images is greater than or equal to 100, such as 100, 150, 200, 300, 500, 800, etc., and preferably, the number of the original images is greater than 500, such as 500, 600, etc. Of course, the number of original images may also be less than 100, such as 95, 90, 40, etc.
Further, it is known that the original image in the image library covers as wide a color temperature range as possible and contains more scenes, and preferably, the color temperature points of the light source corresponding to the original image are distributed in 1000k to 10000k, such as 1000k, 5000k, 8000k, and the like, and cover as many color temperature points as possible, such as covering not only the color temperature points of various standard light sources, but also other color temperature points, so that the generated source image database is richer, thereby improving robustness to neural network training.
It should be noted that, in the above step S200, for the light source at the time of collecting the image, light sources corresponding to different color temperature points are also selected as much as possible, in order to facilitate the subsequent selection of the original image and the collected image having the same color temperature point, especially in an embodiment where the known image library is collected by other collecting units, in order to enable both of them to have the same light source environment, it is preferable that the step S200 collects the collected image under the light box by using different standard light sources, such as D65, T L84, U30, a light source, and the like, and similarly, when the known image library is obtained by collection, some original images are also collected under the light box by using different standard light sources, so as to better select the collected image and the original image having the same color temperature point in step S300.
The light source information of the image comprises a luminance value l of a red channel of the imagerBrightness value l of green channelgAnd the luminance value l of the blue channelbWhether the light source information corresponding to each original image in the known image library is calculated in step S100 or the light source information corresponding to each collected image in the corrected image library is calculated in step S200, the calculation can be implemented in such a manner that, for convenience of description, the original image and the collected image are both referred to as images, each image has three channels of red, green and blue, and the light source information of the image is obtained, specifically, the luminance value l of the red channel corresponding to the image is calculatedrBrightness value l of green channelgAnd the luminance value l of the blue channelbThe calculation method specifically comprises the following steps:
s001: converting the image into the RGB domain, it being understood that the image, whether obtained from the common data set or acquired by the acquisition unit, is typically in raw format, so that the raw format image is first demosaiced and converted from the bayer domain into the RGB domain for subsequent processing;
s002: respectively calculating the pixel mean value of a red channel, the pixel mean value of a green channel and the pixel mean value of a blue channel for white blocks which are not over-exposed on a standard color card in the image, wherein the three mean values are the brightness value l of the red channel of the imagerBrightness value l of green channelgAnd the luminance value l of the blue channelb。
Understandably, by the above calculation method, the brightness value l of the red channel of each original image can be obtainedrBrightness value l of green channelgAnd the luminance value l of the blue channelbAnd the brightness value l of the red channel of each acquired imagerBrightness value l of green channelgAnd the luminance value l of the blue channelb。
In a preferred embodiment, the step S300 specifically includes, according to the multiple sets of light source information corresponding to the original image and the collected image having the same color temperature point: the method comprises the steps of selecting an original image with the same color temperature point from a known image library for each collected image, and forming multiple groups of corresponding original images and collected images, namely, the number of color temperature groups in a color temperature set is equal to that of the collected images, the color temperature set comprises all the collected images, namely, each collected image selects an original image with the same color temperature point, so that all the collected images can be used when a color temperature model is formed, the accuracy of the color temperature model is improved, and the precision of a trained neural network is improved. It is to be understood that, when the color temperature set is formed in step S300, only the original image corresponding to a part of the collected images in the corrected image library may be selected.
In order to facilitate the labeling of the light sources corresponding to each image, the light source information of the image further includes labels x and y, which correspond to the luminance value l of the red channel of the imagerBrightness value l of green channelgAnd the luminance value l of the blue channelbThe specific calculation formula of the labels x and y can refer to the following formula (1), and in practical application, the source image database can store the brightness value l of the red channel corresponding to each imagerBrightness value l of green channelgAnd the luminance value l of the blue channelbThe corresponding labels x and y are also stored; of course, only one of them may be stored, and the corresponding conversion may be performed when necessary.
The step S300 specifically includes:
s310: respectively calculating labels x of all original images according to formula (1)1、y1And label x of the captured image2、y2；
Figure BDA0002331087190000131
Wherein lr、lg、lbThe luminance value of the red channel, the luminance value of the green channel and the luminance value of the blue channel, respectively, of the original image or of the captured image, that is, when x is calculated1、y1The luminance value l of the red channel of the original image is then taken in the above formularBrightness value l of green channelgAnd the luminance value l of the blue channelb(ii) a When calculating x2、y2When, the above formula(1) Substituted in is the brightness value l of the red channel of the acquired imagerBrightness value l of green channelgAnd the luminance value l of the blue channelb。
S320: selecting a plurality of groups of original images and collected images with the same color temperature points, namely forming a plurality of color temperature groups;
s330: establishing a three-dimensional transformation matrix c and transformation coefficients z0Wherein, in the step (A),
Figure BDA0002331087190000132
s340: according to the formulas (2) and (3) and x of the corresponding original images of each group1、y1And x of the acquired image2、y2Calculating a transformation matrix c and transformation coefficients z0；
Figure BDA0002331087190000133
Figure BDA0002331087190000134
In this embodiment, step S400 specifically includes:
according to the formulas (2) and (3), the transformation matrix c and the transformation coefficient z0Label x of each original image of known image library1、y1Converting to obtain the label x of each target image in the target image library3、y3Wherein x in the formula (3)2、y2Is replaced by x3、y3，x′2、y′2Is replaced by x'3、y′3。
The transformation matrix c is not limited to the three-dimensional matrix, and may be a two-dimensional matrix, a four-dimensional matrix, or a matrix with more dimensions. However, considering that if the dimension is too large, the number of the acquired images to be acquired is large, and the calculation amount is large, in the preferred embodiment of the present invention, the three-dimensional transformation matrix is selected to save the image acquisition time and simplify the calculation amount, as well as the image acquisition time and the calculation amountThe accuracy of the obtained source image database can be ensured, and the number of the specific collected images can be determined according to the transformation matrix c and the transformation coefficient z, for example, at c20、c21When 0 is taken and z is taken as 1, only three collected images can be collected. It should be noted that, in the step of establishing the color temperature model, the order of S310, S320, and S330 is not limited as long as all steps are performed before step S340, for example, the three steps may be performed simultaneously and in parallel; or the two are executed in parallel and then executed with the other in a sequential relationship; or may be performed sequentially in any order.
Considering that in the neural network training, the output of some neural networks is a brightness value (described in detail below) after each image is normalized, in order to facilitate the subsequent training of the neural networks, step S400 further includes: label x of each target image3、y3Substituting x and y in the formula (4) to obtain the brightness value l of the normalized red channel of each target imager0Brightness value l of green channelg0And the luminance value l of the blue channelb0，
Figure BDA0002331087190000141
Note that, for the same image, the luminance value l of the red channel before normalization corresponds theretorBrightness value l of green channelgAnd the luminance value l of the blue channelbAnd the normalized brightness value l of the red channelr0Brightness value l of green channelg0And the luminance value l of the blue channelb0The specific values of (a) are different but the labels x, y are the same for the same image.
In fact, in the embodiment of the known image library obtained by acquiring the image through the acquisition unit, the label x of each original image is obtained through calculation of formula (1)1、y1Then, each luminance value after the normalization of the original image may be calculated by the formula (4). In embodiments where the known image library is obtained by white-balanced common datasets, typically the light source information for each image stored in the common dataset is normalizedThe normalized luminance value, i.e. the luminance value obtained in step S100 at this time, is the luminance value l of the red channel normalized by the original imager0Brightness value l of green channelg0And the luminance value l of the blue channelb0Calculating the label x corresponding to the original image1、y1Then, the normalized brightness value l of the red channel can be directly obtainedr0Brightness value l of green channelg0And the luminance value l of the blue channelb0L in formula (1)r、lg、lb。
Considering that the component of the original image in the bayer domain is relatively large for the green channel, and therefore, the corresponding original image can be considered to be substantially consistent with the green pixel value of the target image, in a preferred embodiment, the converting, in step S500, the pixel information of each original image in the known image library to the pixel information of each target image in the target image library according to the light source information of the known image library and the target image library specifically includes:
s510: respectively calculating first red channel gain r _ gain of each original image1And a first blue channel gain b _ gain1And calculating a third red channel gain r _ gain of each target image in the target image library3And a third blue channel gain b _ gain3；
S520: according to the pixel information of the original image and the first red channel gain r _ gain1And a first blue channel gain b _ gain1A third red channel gain r _ gain3And a third blue channel gain b _ gain3And converting the pixel information of each original image into the pixel information of each target image in the target image library.
Specifically, the image information of each image includes a red pixel value, a green pixel value, and a blue pixel value. For convenience of description, in a corresponding set of the original image and the target image, each pixel of the original image is called a first pixel, and a red pixel value thereof is denoted as a first red pixel value r1And the blue pixel value is recorded as a first blue pixel value b1(ii) a Each pixel of the target image is referred to as a third pixelPixel with its red pixel value denoted as the third red pixel value r3And the blue pixel value is recorded as a third blue pixel value b3. Step S510 specifically includes: calculating a third red channel gain r _ gain according to equation (5)3And a third blue channel gain b _ gain3The first red channel gain r _ gain is calculated according to equation (6)1And a first blue channel gain b _ gain1；
Figure BDA0002331087190000151
Figure BDA0002331087190000152
Step S520 specifically includes: calculating pixel information of each target image according to a formula (7);
Figure BDA0002331087190000153
that is, in a corresponding set of the original image and the target image, the first red channel gain r _ gain of the original image is calculated by equation (6)1And a first blue channel gain b _ gain1Calculating a third red channel gain r _ gain of the target image using equation (5)3And a third blue channel gain b _ gain3. Then due to the first red pixel value r of the first pixel in the original image1And a first blue pixel value b1As is known, therefore, the third red pixel value r of the third pixel corresponding to the first pixel can be calculated using equation (7)2And a third blue pixel value b2And circularly converting the pixel value of each first pixel in the original image to the pixel value of the third pixel of the target image. Similarly, the pixel information of the original images of the corresponding groups is converted respectively to obtain the pixel information of the corresponding target images, so that the pixel information and the light source information of the images in the target image library are obtained. It is noted that, in this embodiment, the green pixel value of the first pixel and the green color of the corresponding third pixelThe pixel values are equal.
In an embodiment, if the converted target image library is directly used as the source image database, there may be power frequency interference at the color temperature point acquired by the target acquisition unit, or there may be some unavoidable deviations between the converted target image library and the color temperature curve of the target acquisition unit due to the existence of other light sources or image noise, standard light source errors, calculation errors, and other factors. Therefore, in a preferred embodiment of the present invention, step S500 specifically includes: converting pixel information of each original image of the known image library into pixel information of each target image of the target image library according to light source information of the known image library and light source information of the target image library, merging the plurality of optimized images and the target image library to form an optimized target image library, and obtaining a source image database of the target acquisition unit; the optimized images are images acquired by using a target acquisition unit, for example, 4 or ten optimized images are acquired by using the target acquisition unit, and then the optimized images are merged with a target image library obtained by conversion to obtain a source image database. That is, the source image database includes the target image and the optimized image in the target image library in this embodiment.
By adopting the method, when the initial neural network is trained, because the real data is added to finely adjust the initial neural network, the training effect of the neural network can be further improved, and the accuracy of optimizing the neural network to process the image acquired by the target acquisition unit is increased.
It is understood that the step of acquiring the optimized image by using the target acquisition unit may be executed together with step S200, or may be executed simultaneously with any one of steps S100, S300, or S400, or may be executed before or after any one of steps S100 to S400, or may be executed together with some steps in step S500, as long as the optimized image is acquired before combining a plurality of optimized images into the target image library to form the optimized target image library and obtaining the source image database of the target acquisition unit.
In actual execution, after acquiring a plurality of optimized images, obtaining the optimized images for conversionThe steps S001 and S002 may be performed on the optimized images, and then the light source information of the optimized images is calculated according to the corresponding formula and stored as the label x in the target image library3、y3Then, the label x of each optimized image is directly calculated according to the formula (1)3、y3(ii) a When the normalized brightness value is stored in the target image library, the normalized brightness value of each optimized image needs to be calculated according to the formulas (1) and (4), that is, the brightness value l of the red channel of each optimized image is firstly calculatedrBrightness value l of green channelgAnd the luminance value l of the blue channelbSubstituting into formula (1), and substituting x and y obtained from formula (1) into formula (4) to obtain brightness value l of normalized red channelr0Brightness value l of green channelg0And the luminance value l of the blue channelb0。
Referring to fig. 2, which depicts color temperature distributions of images captured by the known image library, the target image library and the target capturing unit, wherein a large circle point represents the color temperature distribution of each original image in the known image library, a small circle point represents the color temperature distribution of each target image in the target image library, and × represents the color temperature distribution of an image actually captured by the target capturing unit, it can be seen from fig. 2 that the color temperature distribution of the actually captured image is almost identical to the color temperature distribution of the converted target image library, and therefore, the data of the target image library has validity.
In addition, the invention also provides a training method of a neural network, wherein the neural network is used for carrying out white balance processing on the image acquired by the acquisition unit, and the training method comprises the following steps:
the source image database obtained by the acquisition method in any one of the embodiments is used for training the initial neural network, and parameters of the initial neural network are corrected, so that the optimized neural network corresponding to the target acquisition unit is obtained. For convenience of description, an untrained neural network may be referred to as an initial neural network, and when training the initial neural network, the above steps S100 to S500 are first adopted to obtain a source image database corresponding to a target acquisition unit, and then the source image database is used to train the initial neural networkThe method includes inputting pixel information of an image of a source image database into an initial neural network, outputting light source information corresponding to the image, comparing the light source information with light source information of the image stored in a source image database, and correcting parameters of the initial neural network if a difference exists, so that each parameter of the initial neural network is gradually corrected through each image in the source image database until the difference between the light source information output by the initial neural network and the light source information in the source image database is within a preset range, and considering the corrected initial neural network as an optimized neural network. By adopting the training method, because the number of images in the source image database is large, the color temperature points are wide, a better model (namely, an optimized neural network) can be obtained through neural network training, the output of the optimized neural network can be considered as real light source information, and therefore, the optimized neural network can be used for carrying out white balance processing on other images. In the neural network training, when the light source information output by the initial neural network (or the optimized neural network) is label x3、y3Then, the light source information of the target image and the optimized image stored in the source image database both include the label x3、y3(ii) a When the light source information output by the initial neural network (or the optimized neural network) is the normalized brightness value, the light source information of the target image and the optimized image stored in the source image database are both the normalized brightness value.
The present invention also provides an acquisition system of a source image database, the source image database being used for training a neural network, as shown in fig. 3, the acquisition system including: the device comprises an acquisition module 1 and a processing module 2 which are connected with each other to transmit information. The acquisition module 1 is configured to acquire a correction image library, where the correction image library is formed by acquiring, by a target acquisition unit, acquired images with standard color cards at different color temperatures, where different light sources and standard color cards may refer to the foregoing description, and details are not repeated here. The processing module 2 is used for acquiring a known image library and obtaining light source information corresponding to each original image in the known image library; the processing module 2 is further configured to calculate light source information corresponding to each acquired image, and establish a color temperature model converted from an original image to an acquired image according to a plurality of sets of original images with the same color temperature points and corresponding light source information in the acquired images; converting the light source information of each original image of the known image library according to the color temperature model to obtain the light source information of each target image in the target image library; converting the pixel information of each original image of the known image library into the pixel information of each target image of the target image library according to the light source information of the known image library and the light source information of the target image library to obtain a source image database corresponding to the target acquisition unit; the number of the collected images is smaller than that of the original images, and the two images have different sources.
The above-mentioned acquiring system may be used to execute the acquiring method in the above-mentioned embodiments, but it should be noted that the execution of the above-mentioned acquiring method is not limited to this acquiring system, for example, the acquiring system may further include a known image acquiring module for acquiring a known image library, in which case the processing module 2 may reduce the function of acquiring the corresponding known image library. For another example, the known image obtaining module may be configured to obtain a known image library and calculate light source information corresponding to each original image in the known image library; the acquisition module 1 is used for acquiring a correction image library and calculating light source information corresponding to each acquired image; the processing module 2 can then reduce the corresponding functionality.
In a preferred embodiment of the present invention, a training device is provided for training a neural network, which can be directly used as a device for production, sale, or use, and the like, as shown in fig. 4, the training device 3 includes a training system 3 and an acquisition system 4 of a source image database, the training system 3 has an initial neural network, and the acquisition system 4 acquires the source image database by using the acquisition method described in any one of the above embodiments, and may be the acquisition system 4 described in any one of the above embodiments. The training system 3 is connected to the obtaining system 4, and is configured to train the initial neural network using the source image database, and modify each parameter of the initial neural network to obtain an optimized neural network, and specifically may train using the training method of the neural network described in any of the above embodiments. Obviously, the accuracy of the final neural network can be improved by adopting the training device.
The training device may be directly implanted into the electronic device, but considering that after the neural network is trained, the source image database, the initial neural network and the specific program of the training method are basically not needed any more, in a preferred embodiment of the present invention, the training device further includes an output system, which may include a USB interface, a 232 serial port, etc., or may be a wireless output module, for outputting the optimized neural network, for example, directly outputting the optimized neural network to the electronic device or other storage devices.
Furthermore, the present invention also provides a computer-readable storage medium, such as an optical disc, a usb disk, etc., on which a computer program is stored, which when executed implements the acquisition method as described in any of the above embodiments or implements the training method as described in any of the above embodiments. That is, the above-mentioned acquisition method and training method can be directly stored in a computer-readable storage medium such as an optical disc, and the client can read and execute the computer program through a computer as required to generate the optimized neural network.
In addition, the invention also provides an image processing chip which is provided with a storage unit, wherein the storage unit stores the optimized neural network output by the training device or generated by the computer according to any one of the embodiments. When the white balance processing device is installed on electronic equipment, the white balance processing device can be directly used for carrying out white balance processing on the image acquired by the acquisition unit. Of course, other image processing programs, or other programs such as drivers and applications may be stored in the image processing chip.
The invention further provides an image processing system, which comprises the training device and the image processing chip, wherein the optimized neural network output by the training device is stored in the storage unit, so that the white balance processing is performed on the image acquired by the target acquisition unit.
It should be noted that the optimized neural network output by the training apparatus or generated by the computer may also be directly embedded in the electronic device, for example, directly stored in the central processing unit of the electronic device.
In actual use, corresponding to one electronic device or a plurality of devices (a plurality of devices have the same type of acquisition unit), in one embodiment, a plurality of images can be firstly shot through the acquisition units to form a correction image library; then the training device obtains a source image database corresponding to an acquisition unit (namely a target acquisition unit) of the electronic equipment according to the corrected image database and the known image database, and trains the neural network through the source image database to obtain an optimized neural network; and finally, directly implanting the optimized neural network into the image processing chip, then installing the image processing chip in the electronic equipment, and after the electronic equipment acquires a new image, carrying out white balance processing on the new image by the image processing chip through the optimized neural network.
It will be appreciated by those skilled in the art that the above-described preferred embodiments may be freely combined, superimposed, without conflict.
It will be understood that the embodiments described above are illustrative only and not restrictive, and that various obvious and equivalent modifications and substitutions for details described herein may be made by those skilled in the art without departing from the basic principles of the invention.

Claims (21)

1. An acquisition method of a source image database, the source image database being used for training a neural network, the neural network being used for white balance processing of an image acquired by an acquisition unit, the acquisition method comprising the steps of:
s100: acquiring a known image library, and acquiring light source information corresponding to each original image in the known image library; the original images in the known image library cover a plurality of color temperature points, and each original image is provided with a standard color card;
s200: acquiring a correction image library, and calculating light source information corresponding to each acquired image in the correction image library; wherein the correction image library is formed by collecting collected images with standard color cards at different color temperatures by using a target collecting unit;
s300: establishing a color temperature model converted from the original image to the collected image according to corresponding light source information in the original image and the collected image with the same color temperature points, wherein the color temperature model comprises a conversion matrix c and a conversion coefficient z0Wherein, the product of the light source information of the original image and the conversion matrix c is equal to the product of the light source information of the collected image and the conversion coefficient;
s400: converting the light source information of each original image in the known image library according to the color temperature model to obtain the light source information of each target image in the target image library;
s500: converting the pixel information of each original image of the known image library into the pixel information of each target image of the target image library according to the light source information of the known image library and the light source information of the target image library to obtain a source image database corresponding to the target acquisition unit;
the number of the collected images of the correction image library is smaller than that of the original images of the known image library, and the image sources of the two images are different.
2. The obtaining method according to claim 1, wherein the obtaining of the known image library in step S100 specifically includes:
acquiring the original image with a standard color card under different scenes by using a non-target acquisition unit to form the known image library; or
The library of known images is acquired by a white balanced common dataset.
3. The method according to claim 1, wherein the step S200 of forming the corrected image library by using the target collecting unit to collect the collected images with standard color cards at different color temperatures includes: the correction image library is formed by collecting collected images of standard color cards with different standard light sources under a light box by using a target collecting unit.
4. The method according to claim 1, wherein the light source information of an image includes a luminance value l of a red channel of the imagerBrightness value l of green channelgAnd the luminance value l of the blue channelbThe specific method for obtaining the light source information in step S100 and step S200 includes the steps of:
s001: converting the image to the RGB domain;
s002: respectively calculating the pixel mean value of a red channel, the pixel mean value of a green channel and the pixel mean value of a blue channel for the white blocks which are not over-exposed on the standard color card in the image, wherein the three mean values are the brightness value l of the red channel of the imagerBrightness value l of green channelgAnd the luminance value l of the blue channelb；
Wherein the image refers to the original image or the collected image.
5. The acquisition method according to claim 4, wherein the light source information further includes a label; the conversion matrix c is a three-dimensional conversion matrix;
the step S300 specifically includes the steps of:
s310: respectively calculating the label x of each original image according to formula (1)1、y1And label x of the captured image2、y2；
Figure FDA0002331087180000021
Wherein lr、lg、lbThe brightness values of the red channel, the green channel and the blue channel of the original image or the collected image are respectively;
s320: selecting a plurality of groups of original images and collected images with the same color temperature points;
s330: establishing a three-dimensional transformation matrix c and transformation coefficients z0Wherein, in the step (A),
Figure FDA0002331087180000022
S340：
Figure FDA0002331087180000031
Figure FDA0002331087180000032
according to formulas (2) and (3) and the labels x of the original images corresponding to the groups1、y1And label x of the captured image2、y2Calculating the transformation matrix c and transformation coefficients z0；
The step S400 specifically includes:
according to the formulas (2) and (3), the transformation matrix c and the transformation coefficient z0Label x of each original image of the known image library1、y1Converting to obtain the label x of each target image in the target image library3、y3Wherein x in the formula (3)2、y2Is replaced by x3、y3，x′2、y′2Is replaced by x'3、y′3。
6. The obtaining method according to claim 5, wherein the step S400 further comprises: label x of each target image3、y3Substituting x and y in the formula (4) to obtain the brightness value l of the normalized red channel of each target imager0Brightness value l of green channelg0And the luminance value l of the blue channelb0。
Figure FDA0002331087180000033
7. The method according to claim 5, wherein in the step S500, converting the pixel information of each original image in the known image library to the pixel information of each target image in the target image library according to the light source information in the known image library and the target image library, specifically includes the steps of:
s510: respectively calculating a first red channel gain and a first blue channel gain of each original image, and calculating a third red channel gain and a third blue channel gain of each target image in the target image library;
s520: and converting the pixel information of each original image into the pixel information of each target image in the target image library according to the pixel information of the original image and the first red channel gain, the first blue channel gain, the third red channel gain and the third blue channel gain.
8. The obtaining method according to claim 7, wherein the step S510 specifically includes: calculating the third red channel gain r _ gain according to equation (5)3And the third blue channel gain b _ gain3Calculating the first red channel gain r _ gain according to equation (6)1And the first blue channel gain b _ gain1；
Figure FDA0002331087180000041
Figure FDA0002331087180000042
The pixel information of the original image comprises a first red pixel value r1A first blue pixel value b1(ii) a The pixel information of the target image comprises a third red pixel value r3The third blue pixel value b3；
The step S520 specifically includes: calculating pixel information of each target image according to formula (7);
Figure FDA0002331087180000043
9. the obtaining method according to claim 1, wherein the step S300 specifically includes, according to multiple sets of corresponding light source information in the original image and the acquired image with the same color temperature point:
and selecting the original image with the same color temperature point from the known image library for each acquired image so as to form a plurality of groups of corresponding original images and acquired images.
10. The acquisition method according to claim 1, characterized in that the number of said original images is greater than or equal to 100.
11. The method according to claim 1, wherein the color temperature points of the light source corresponding to the original image in the known image library are distributed in a range of 1000k to 10000 k.
12. The obtaining method according to any one of claims 1 to 11, wherein the step S500 specifically includes:
converting pixel information of each original image of the known image library into pixel information of each target image of the target image library according to the light source information of the known image library and the target image library, merging a plurality of optimized images with the target image library to form an optimized target image library, and obtaining a source image database of the target acquisition unit;
wherein the optimized image is an image acquired using the target acquisition unit.
13. A training method of a neural network, the neural network being used for white balance processing of an image acquired by an acquisition unit, the training method comprising the steps of:
training an initial neural network by using a source image database obtained by the acquisition method according to any one of claims 1 to 12, and correcting parameters of the initial neural network to obtain an optimized neural network corresponding to the target acquisition unit.
14. An acquisition system of a source image database for training a neural network for white balance processing of an image acquired by an acquisition unit, the acquisition system comprising:
the system comprises an acquisition module, a correction image database and a display module, wherein the acquisition module is used for acquiring a correction image database, and the correction image database acquires acquired images with standard color cards at different color temperatures through a target acquisition unit to form;
the processing module is used for acquiring a known image library and obtaining light source information corresponding to each original image in the known image library; the processing module is further used for calculating light source information corresponding to each collected image, and establishing a color temperature model converted from the original image to the collected image according to a plurality of groups of light source information corresponding to the original image and the collected image with the same color temperature point; converting the light source information of each original image in the known image library according to the color temperature model to obtain the light source information of each target image in the target image library; converting the pixel information of each original image of the known image library into the pixel information of each target image of the target image library according to the light source information of the known image library and the light source information of the target image library to obtain a source image database corresponding to the target acquisition unit;
and the number of the collected images is less than that of the original images, and the two images have different sources.
15. A training apparatus for training a neural network for white balance processing of an image acquired by an acquisition unit, the training apparatus comprising a training system having an initial neural network and an acquisition system of a source image database,
the acquisition system is configured to acquire a source image database using the acquisition method of any one of claims 1-12;
and the training system is connected with the acquisition system and used for training the initial neural network by using the source image database and correcting all parameters of the initial neural network to obtain an optimized neural network.
16. The training apparatus of claim 15, further comprising an output port for outputting the optimized neural network.
17. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed, implements the acquisition method according to any one of claims 1-12 or implements the training method according to claim 13.
18. An image processing chip having a storage unit storing the optimized neural network output by the training apparatus of claim 15 or 16.
19. An image processing system comprising the training apparatus according to claim 15 or 16 and the image processing chip according to claim 18, the storage unit storing the optimized neural network output by the training apparatus.
20. An electronic device comprising an acquisition unit, further comprising the image processing chip of claim 18, the image processing chip being connected to the acquisition unit for white balance processing of an image acquired by the acquisition unit using the optimized neural network.
21. The electronic device of claim 20, wherein the electronic device comprises at least one of a mobile terminal, a camera, a scanner, a security monitoring device, and a tachograph.
CN201911336566.0A 2019-12-23 2019-12-23 Method for acquiring source image database, training device and electronic equipment Active CN111435986B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911336566.0A CN111435986B (en) 2019-12-23 2019-12-23 Method for acquiring source image database, training device and electronic equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911336566.0A CN111435986B (en) 2019-12-23 2019-12-23 Method for acquiring source image database, training device and electronic equipment

Publications (2)

Publication Number Publication Date
CN111435986A true CN111435986A (en) 2020-07-21
CN111435986B CN111435986B (en) 2021-11-23

Family

ID=71580935

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911336566.0A Active CN111435986B (en) 2019-12-23 2019-12-23 Method for acquiring source image database, training device and electronic equipment

Country Status (1)

Country Link
CN (1) CN111435986B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115623338A (en) * 2022-08-02 2023-01-17 荣耀终端有限公司 Image processing method and electronic equipment

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB1398147A (en) * 1971-10-05 1975-06-18 Eastman Kodak Co Image reproduction apparatus
JP2001034749A (en) * 1999-07-26 2001-02-09 Toyo Ink Mfg Co Ltd Color reproducing device
US20020149679A1 (en) * 1994-06-28 2002-10-17 Deangelis Douglas J. Line object scene generation apparatus
US20140218540A1 (en) * 2013-02-05 2014-08-07 Google Inc. Noise Models for Image Processing
CN107578390A (en) * 2017-09-14 2018-01-12 长沙全度影像科技有限公司 A kind of method and device that image white balance correction is carried out using neutral net
CN108364267A (en) * 2018-02-13 2018-08-03 北京旷视科技有限公司 Image processing method, device and equipment
CN109427041A (en) * 2017-08-25 2019-03-05 中国科学院上海高等研究院 A kind of image white balance method and system, storage medium and terminal device
CN110267024A (en) * 2019-07-26 2019-09-20 惠州视维新技术有限公司 Adjust the method, apparatus and computer readable storage medium of TV white balance value

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB1398147A (en) * 1971-10-05 1975-06-18 Eastman Kodak Co Image reproduction apparatus
US20020149679A1 (en) * 1994-06-28 2002-10-17 Deangelis Douglas J. Line object scene generation apparatus
JP2001034749A (en) * 1999-07-26 2001-02-09 Toyo Ink Mfg Co Ltd Color reproducing device
US20140218540A1 (en) * 2013-02-05 2014-08-07 Google Inc. Noise Models for Image Processing
CN109427041A (en) * 2017-08-25 2019-03-05 中国科学院上海高等研究院 A kind of image white balance method and system, storage medium and terminal device
CN107578390A (en) * 2017-09-14 2018-01-12 长沙全度影像科技有限公司 A kind of method and device that image white balance correction is carried out using neutral net
CN108364267A (en) * 2018-02-13 2018-08-03 北京旷视科技有限公司 Image processing method, device and equipment
CN110267024A (en) * 2019-07-26 2019-09-20 惠州视维新技术有限公司 Adjust the method, apparatus and computer readable storage medium of TV white balance value

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115623338A (en) * 2022-08-02 2023-01-17 荣耀终端有限公司 Image processing method and electronic equipment
CN115623338B (en) * 2022-08-02 2024-04-12 荣耀终端有限公司 Image processing method and electronic equipment

Also Published As

Publication number Publication date
CN111435986B (en) 2021-11-23

Similar Documents

Publication Publication Date Title
US10791310B2 (en) Method and system of deep learning-based automatic white balancing
Jiang et al. Learning the image processing pipeline
KR102495753B1 (en) Method and electronic device for processing raw image acquired by using camera by using external electronic device
CN105409211B (en) For the automatic white balance positive with skin-color adjustment of image procossing
CN113170028B (en) Method for generating image data of machine learning based imaging algorithm
US10600170B2 (en) Method and device for producing a digital image
JP2002543695A (en) Color device characterization method and apparatus, image color correction method and apparatus, color image data transmission and / or recording method, and recording medium
CN109191403A (en) Image processing method and device, electronic equipment, computer readable storage medium
EP2845164B1 (en) Reference card for scene referred metadata capture
CA3090504A1 (en) Systems and methods for sensor-independent illuminant determination
WO2019015553A1 (en) Image color correction method, device, storage medium and computer equipment
US9723221B2 (en) Imaging device shooting a common subject in synchronization with other imaging devices
CN112508812A (en) Image color cast correction method, model training method, device and equipment
WO2024027287A9 (en) Image processing system and method, and computer-readable medium and electronic device
US20200404235A1 (en) Device for and method of correcting white balance of image
CN111435986B (en) Method for acquiring source image database, training device and electronic equipment
CN110580684A (en) image enhancement method based on black-white-color binocular camera
Lam et al. Automatic white balancing using standard deviation of RGB components
JP5962169B2 (en) Digital camera, color conversion program and recording control program
Lam et al. Automatic white balancing using luminance component and standard deviation of RGB components [image preprocessing]
WO2022052820A1 (en) Data processing method, system, and apparatus
CN115706870B (en) Video processing method, device, electronic equipment and storage medium
CN104954767B (en) A kind of information processing method and electronic equipment
JP2014090410A (en) Method for white balance adjustment
CN114266694A (en) Image processing method, apparatus and computer storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
CB02 Change of applicant information

Address after: 519075 No. 333, Kexing Road, Xiangzhou District, Zhuhai City, Guangdong Province

Applicant after: Zhuhai Jieli Technology Co., Ltd

Address before: 519085 Complex Building 9 (1-4 floors), No. 107, Shihua West Road, Jida, Zhuhai, Guangdong

Applicant before: Zhuhai Jieli Technology Co., Ltd

CB02 Change of applicant information
GR01 Patent grant
GR01 Patent grant