WO2021051382A1 - 白平衡处理方法和设备、可移动平台、相机 - Google Patents

白平衡处理方法和设备、可移动平台、相机 Download PDF

Info

Publication number
WO2021051382A1
WO2021051382A1 PCT/CN2019/106950 CN2019106950W WO2021051382A1 WO 2021051382 A1 WO2021051382 A1 WO 2021051382A1 CN 2019106950 W CN2019106950 W CN 2019106950W WO 2021051382 A1 WO2021051382 A1 WO 2021051382A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
value
gray
white balance
pixel
Prior art date
Application number
PCT/CN2019/106950
Other languages
English (en)
French (fr)
Inventor
肖瑾
曹子晟
胡攀
Original Assignee
深圳市大疆创新科技有限公司
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 深圳市大疆创新科技有限公司 filed Critical 深圳市大疆创新科技有限公司
Priority to CN201980033268.6A priority Critical patent/CN112204957A/zh
Priority to PCT/CN2019/106950 priority patent/WO2021051382A1/zh
Publication of WO2021051382A1 publication Critical patent/WO2021051382A1/zh

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N23/00Cameras or camera modules comprising electronic image sensors; Control thereof
    • H04N23/80Camera processing pipelines; Components thereof
    • H04N23/84Camera processing pipelines; Components thereof for processing colour signals
    • H04N23/88Camera processing pipelines; Components thereof for processing colour signals for colour balance, e.g. white-balance circuits or colour temperature control

Definitions

  • the present disclosure relates to the field of image processing technology, in particular, to white balance processing methods and equipment, movable platforms, and cameras.
  • the human visual system has the characteristic of color constancy, that is, the human eye can adapt to different lighting and restore the color of the scene under different lighting to the color of the scene illuminated by white light.
  • the image acquisition device for example, a camera
  • Traditional white balance processing methods can be roughly divided into two categories: based on statistical priors and based on learning algorithms.
  • the first method is mostly based on making observations on images with no color cast and assuming a priori, and then use the statistical information of the image to infer the illumination of the image. But this kind of algorithm basically assumes that the normal image conforms to a certain prior knowledge. In some extreme cases, the prior knowledge may not be satisfied. In this way, algorithms based on statistical priors will produce large errors.
  • the embodiments of the present disclosure propose a white balance processing method and device, a movable platform, and a camera to solve the technical problem of low accuracy in related technologies.
  • a white balance processing method includes:
  • a white balance processing device includes a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor executes The program implements the following methods:
  • a movable platform including:
  • a power system installed in the machine body and used to provide power to the movable platform;
  • the white balance processing device described in any embodiment.
  • a camera including:
  • the lens assembly is arranged inside the housing;
  • a sensor assembly which is arranged inside the housing, and is used to sense light passing through the lens assembly and generate an electrical signal
  • the white balance processing device described in any embodiment.
  • a computer-readable storage medium is provided, and a number of computer instructions are stored on the readable storage medium, and the computer instructions implement the steps of the method described in any embodiment when the computer instructions are executed.
  • Fig. 1 is a flowchart of a white balance processing method according to an embodiment of the present disclosure.
  • Fig. 2 is an effect comparison diagram of a histogram constructed according to the prior art and the present invention according to an embodiment of the present disclosure.
  • Fig. 3 shows a network framework of a gray detection network according to an embodiment of the present disclosure.
  • Fig. 4 is a schematic structural diagram of a computer device for implementing the method of the embodiment of the present disclosure according to the embodiment of the present disclosure.
  • Fig. 5 is a block diagram showing a movable platform according to an embodiment of the present disclosure.
  • Fig. 6 is a block diagram showing a camera according to an embodiment of the present disclosure.
  • first, second, third, etc. may be used in this specification to describe various information, the information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other.
  • first information may also be referred to as second information, and similarly, the second information may also be referred to as first information.
  • word “if” as used herein can be interpreted as "when” or “when” or "in response to determination”.
  • FIG. 1 it is a flowchart of a white balance processing method according to an embodiment of the present disclosure.
  • the method may include:
  • Step S101 detecting gray points on the image to be processed
  • Step S102 Construct a histogram according to the gray points
  • Step S103 Estimate the illumination of the image to be processed according to the histogram, and perform white balance processing on the image to be processed according to the illumination.
  • the pixel when a certain pixel on an image satisfies: the R channel color component, G channel color component, and B channel color component of the pixel on the image are equal, then the pixel is the Gray dots on the image.
  • an embodiment of the present specification uses a machine learning model to detect gray points.
  • the image to be processed can be input into a pre-trained machine learning model to obtain the first probability value of each pixel on the image to be processed as a gray point; the pixel with the first probability value greater than the preset value is taken as Gray point.
  • the first probability value can be set based on experience. The larger the first probability value, the higher the confidence that the corresponding pixel point is a gray point. Detecting gray points through a machine learning model improves the accuracy of gray point detection, thereby improving the white balance processing effect.
  • the machine learning model used to detect gray points is a convolutional neural network model (Convolutional Neural Networks, CNN).
  • CNN convolutional Neural Networks
  • Other neural network models can also be used here, or other machine learning models can be used. Since CNN has a relatively strong expressive ability and can effectively extract key features, the use of CNN to detect gray points can further improve the detection accuracy.
  • the above-mentioned convolutional neural network model can be obtained by training according to the training image and the training label of each pixel on the training image, and according to the preset loss function.
  • the light of the image is given instead of the label of the gray point. Labeling gray dots is not only time-consuming, but also expensive. Because the definition of gray under white light is clear, that is, the gray point satisfies:
  • L c represents the color value of the light L on the R, G, and B color channels.
  • the training label of each pixel on the training image is determined according to the second probability value, namely:
  • p(h,w) represents the second probability value corresponding to the pixel at the coordinates (h,w) on the training image
  • ⁇ ,> represents the inner product operation
  • I c (h,w) represents the The RGB value of the pixel at the coordinates (h, w) on the training image
  • the full image of the training image can be used as input, and the CNN can be trained according to the loss function.
  • the embodiment of this specification adopts the Unet neural network structure to maintain the spatial size of the output.
  • the input of the Unet neural network is a color image of H*W*3, and the output is the third probability value of H*W, denoted as Among them, H and W are the pixel height and pixel width of the training image, respectively.
  • the loss function is a cross entropy function.
  • the loss function of an embodiment can be obtained according to the following formula, namely:
  • Loss is the loss function
  • the third probability value corresponding to the pixel at the coordinates (h, w) on the training image output by the convolutional neural network model, p(h, w) indicates that the coordinates on the training image are (h, The second probability value corresponding to the pixel at w).
  • the test image can be input into the network to obtain the predicted second probability value, and then set the threshold ⁇ , if the pixels in the test image The predicted second probability value of the point is greater than the threshold ⁇ , then it is determined that the pixel point is a gray point; otherwise, it is determined that the pixel point is a color point. According to the accuracy of the predicted second probability value, the performance of the trained machine learning model can be verified. If the accuracy is greater than the preset accuracy threshold, the gray points on the image to be processed can be detected according to the trained model.
  • I and W represent the observed image and the image illuminated by natural white light (ie ideal image)
  • L represents the illumination
  • c represents the three color channels of R, G, and B
  • n is the length and width subscript of the image (image) .
  • the color values in the RGB space may be mapped to a preset target space to obtain the mapping value of the RGB value in the target space, and then a histogram is constructed according to the mapping value.
  • the multiplicative relationship between the image and the illumination is converted into an additive relationship.
  • the step of mapping the RGB value of the gray point on the image to be processed to a preset target space, and obtaining the mapping value of the RGB value on the target space includes: according to the gray point
  • the first mapping value of the target space is generated from the R channel color component and the G channel color component of the gray point
  • the second mapping value of the target space is generated according to the G channel color component and the B channel color component of the gray point. Assuming that the first mapping value and the second mapping value are represented by u and v, respectively, then:
  • r, g, and b are the R channel color component, G channel color component, and B channel color component of the gray point on the image to be processed, respectively.
  • I u , Wu and Lu are the observation image I, the first mapping value of the ideal image W and the illumination L in the target space
  • I v , W v and L v are the observation image I, the ideal image W and the illumination respectively The second mapping value of L in the target space.
  • the image W and the light L conform to the multiplicative relationship in the RGB space. Due to the nature of the log function, after turning to the uv space, the multiplicative relationship is converted into an additive relationship. Calculate the two-dimensional statistical histogram for the image ⁇ W u ,W v ⁇ in uv space. Due to the additive relationship, the deviation of the histogram pattern corresponds to ⁇ L u ,L v ⁇ , and then convert it back to RGB space. Get light ⁇ L r ,L g ,L b ⁇ .
  • each dimension of the target space can be evenly divided into several equal parts to obtain multiple intervals; according to the mapping value, the number of gray points falling in each interval can be counted; The number of gray points corresponding to each interval constructs a histogram.
  • the u-axis and v-axis of the uv space can be divided into N equal parts to obtain N*N intervals (N is a positive integer), where the gray point in the i-th space satisfies: U i1 ⁇ u i ⁇ U i2, V i1 ⁇ v i ⁇ V i2, 1 ⁇ i ⁇ N, u i and V i represent the first mapping and the second mapping value within the i th interval value gray points, U i1 and U i2 respectively Represents the lower limit and upper limit of the first mapping value of the i-th interval, and Vi1 and Vi2 represent the lower limit and upper limit of the second mapping value of the i-th interval, respectively.
  • the first mapping value of a gray point falls within the value range of the first mapping value in the corresponding interval
  • the second mapping value of the gray point falls within the value range of the second mapping value in the corresponding interval
  • the The gray points are the gray points in the interval. Therefore, the number of pixel points in each interval can be calculated according to the first mapping value and the second mapping value of each gray point and the upper and lower limits of the first mapping value of each interval.
  • FIG. 2 shows the effect comparison diagram of the histogram constructed in the prior art and the present invention, in which the left side is the histogram constructed in the prior art, and the right figure is the histogram constructed in the embodiment of the present invention. It can be seen that, compared with the histogram of the prior art, the histogram based on gray points proposed by the present invention removes the discrete noise pattern in the histogram, thereby making the subsequent illumination estimation simpler and more accurate.
  • the embodiment of this specification may use a preselected filter template and the histogram to do sliding window convolution to obtain the filtered response, and then use the illumination of the pixel with the largest response as the waiting Process the lighting of the image.
  • H I represents the histogram
  • F represents the filter template
  • the illumination ⁇ L r ,L g ,L b ⁇ can be obtained.
  • white balance processing can be performed according to the illumination ⁇ L r , L g , L b ⁇ .
  • it can also be based on with Perform white balance processing to obtain W u and W v corresponding to the uv space respectively, and then inversely map W u and W v to the RGB space to generate an ideal image.
  • the traditional white balance processing method generally uses each pixel in the image to be processed for white balance processing. However, some color points that deviate greatly from gray may have no effect on the final estimated illumination, or even have a negative effect. . Using all pixels, including color points, for white balance processing may cause the histogram to be too scattered, including large deviated noise points, which will affect the final lighting calculation, resulting in poor white balance processing.
  • the network framework of a gray detection network shown in the embodiment of the present disclosure is shown in FIG. 3.
  • the image to be processed is input into a machine learning model (for example, CNN) to detect gray points in the image to be processed, and then the RGB values of the gray points are spatially converted to obtain the mapping value corresponding to the RGB value, according to the mapping value
  • a machine learning model for example, CNN
  • the RGB values of the gray points are spatially converted to obtain the mapping value corresponding to the RGB value, according to the mapping value
  • To construct a histogram, and estimate the illumination based on the histogram because only the gray points useful for illumination estimation are used when constructing the histogram, and the color points that may interfere with the illumination estimation are removed, so the accuracy of the illumination estimation is improved;
  • the generalization ability is improved.
  • the gray points only occupy a small part of the pixels on the image to be processed, the amount of data is reduced, the computational complexity is reduced, and the white balance processing efficiency is improved.
  • the embodiment of this specification achieves the numerical effect of ranking first.
  • the white balance method based on gray detection proposed in the embodiments of this specification shows better cross-database generalization ability than other learning algorithms.
  • the white balance processing device in the embodiment of this specification may be, for example, a server or a terminal device.
  • the method embodiments can be implemented by software, or can be implemented by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory by the processor that processes the file where it is located. From a hardware perspective, as shown in FIG. 4, it is a hardware structure diagram of the white balance processing device 400 that implements the method of this specification, except for the processor 401, the memory 402, the network interface 403, and the hardware structure shown in FIG. In addition to the lossy memory 404, the white balance processing device used to implement the method of the present specification in the embodiment may also include other hardware according to the actual function of the white balance processing device, which will not be repeated here.
  • the embodiment of this specification provides a white balance processing device, the white balance processing device includes a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor executes the following when the program is executed method:
  • the traditional white balance processing method generally uses each pixel in the image to be processed for white balance processing. However, some color points that deviate greatly from gray may have no effect on the final estimated illumination, or even have a negative effect. . Using all pixels, including color points, for white balance processing may cause the histogram to be too scattered, including large deviated noise points, which will affect the final lighting calculation, resulting in poor white balance processing.
  • the gray points useful for illumination estimation are used when constructing the histogram, and the color points that may interfere with the illumination estimation are removed. Therefore, the accuracy of the illumination estimation is improved; at the same time, because different image acquisition devices collect The gray of the image has commonality, so the generalization ability is improved.
  • the gray points only occupy a small part of the pixels on the image to be processed, the amount of data is reduced, the computational complexity is reduced, and the white balance processing efficiency is improved.
  • the step of the processor constructing a histogram according to the gray point includes: mapping the RGB value of the gray point on the image to be processed to a preset target space to obtain the RGB value in the The mapping value on the target space; a histogram is constructed according to the mapping value.
  • the multiplicative relationship between the image and the illumination is converted into an additive relationship. Through space conversion, the multiplicative relationship is converted into an additive relationship, which is convenient for processing and reduces the computational complexity.
  • the RGB value includes an R channel color component, a G channel color component, and a B channel color component
  • the mapping value includes a first mapping value and a second mapping value
  • the processor places the gray point at the The RGB value on the image to be processed is mapped to a preset target space
  • the step of obtaining the mapping value of the RGB value on the target space includes: generating the RGB value based on the R channel color component and the G channel color component of the gray point.
  • the first mapping value of the target space generates the second mapping value of the target space according to the G channel color component and the B channel color component of the gray point. Assuming that the first mapping value and the second mapping value are represented by u and v, respectively, then:
  • r, g, and b are the R channel color component, G channel color component, and B channel color component of the gray point on the image to be processed, respectively.
  • the step of the processor detecting gray points on the image to be processed includes: inputting the image to be processed into a pre-trained machine learning model to obtain the first probability value of each pixel on the image to be processed as a gray point ;
  • the pixel points with the first probability value greater than the preset value are regarded as gray points.
  • the machine learning model is a convolutional neural network model. Since CNN has a relatively strong expressive ability and can effectively extract key features, the use of CNN to detect gray points can further improve the detection accuracy.
  • the processor executes the program, the following method is also implemented: training the convolutional neural network model according to the training image and the training label of each pixel on the training image, and according to a preset loss function.
  • the processor executes the program, the following method is also implemented: calculating the corresponding pixel as the first gray point according to the RGB value of each pixel on the training image and the RGB value of the gray point on the training image. Two probability values; the training label of each pixel on the training image is determined according to the second probability value.
  • the light of the image is given instead of the label of the gray point. Labeling gray dots is not only time-consuming, but also expensive. By detecting the gray points in the above manner, marking the gray points one by one is avoided, and the efficiency of gray point detection is improved.
  • the step of calculating the second probability value of the corresponding pixel as a gray point by the processor according to the RGB value of each pixel on the training image and the RGB value of the gray point on the training image respectively includes: The inner product of the RGB value of the gray point on the training image and the RGB value of each pixel on the training image; and the second probability value of the corresponding pixel point being a gray point is calculated according to the inner product.
  • the loss function is a cross entropy function.
  • the step of the processor constructing a histogram according to the mapping value includes: evenly dividing each dimension of the target space into a number of equal parts to obtain a plurality of intervals; according to the mapping value, statistics fall within The number of gray points in each interval; construct a histogram according to the number of gray points corresponding to each interval.
  • the step of the processor estimating the illumination of the image to be processed according to the histogram includes: using a filter template and histogram to perform sliding window convolution to obtain a filtered response; The illumination is used as the illumination of the image to be processed.
  • the embodiment of this specification also provides a movable platform 500, including: a body 501; a power system 502 installed in the body 501 to provide power for the movable platform; and any The white balance processing device 400 described in the embodiment.
  • the movable platform 500 is a vehicle, a drone or a movable robot.
  • an embodiment of the present specification also provides a camera 600, which includes: a housing 601; a lens assembly 602, which is arranged inside the housing 601; a sensor assembly 603, which is arranged inside the housing 601 for sensing passing The light of the lens assembly 602 generates an electrical signal; and the white balance processing device 400 according to any one of the embodiments.
  • the embodiments of this specification also provide a computer-readable storage medium on which a number of computer instructions are stored, and when the computer instructions are executed, the steps of the method described in any of the embodiments are implemented.
  • the embodiments of this specification may adopt the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program codes.
  • Computer usable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be realized by any method or technology.
  • the information can be computer-readable instructions, data structures, program modules, or other data.
  • Examples of computer storage media include, but are not limited to: phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disc (DVD) or other optical storage, Magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media can be used to store information that can be accessed by computing devices.
  • PRAM phase change memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • RAM random access memory
  • ROM read-only memory
  • EEPROM electrically erasable programmable read-only memory
  • flash memory or other memory technology
  • CD-ROM compact disc
  • DVD digital versatile disc
  • Magnetic cassettes magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media can be used to store information that can be accessed by computing devices.

Landscapes

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

Abstract

一种白平衡处理方法和设备、可移动平台、相机,先检测图像中的灰色点,再利用所有灰色点构建统计直方图,最后利用直方图估计光照并进行白平衡处理。由于构建直方图时只采用了对光照估计有用的灰色点,而去掉了可能会干扰光照估计的彩色点,因此,提高了光照估计的准确性;同时,由于不同图像采集装置采集图像的灰色存在共通性,因此,提高了泛化能力。

Description

白平衡处理方法和设备、可移动平台、相机 技术领域
本公开涉及图像处理技术领域,具体而言,涉及白平衡处理方法和设备、可移动平台、相机。
背景技术
人眼视觉***具有颜色恒常特性,即人眼可以适应不同的光照,将不同光照下的场景颜色恢复成白光照射下的场景颜色。然而图像采集装置(例如,相机)在拍照时却往往不具备颜色恒常能力。因此,需要对图像采集装置采集到的图像进行白平衡处理,从而检测出拍摄场景的光照,并消除光照的影响,使图片回复到白色光照射的颜色。
传统的白平衡处理方法大致可以分为两大类:基于统计先验的和基于学习算法的。第一种方式大多根据对正常无色偏的图像做出观测并假设先验,再运用图像的统计信息推测出图像的光照。但是这类算法基本都假设正常图像符合某种先验知识,在某些极端情况下,先验知识可能会不满足,这样基于统计先验的算法会产生很大的误差。
发明内容
有鉴于此,本公开的实施例提出了白平衡处理方法和设备、可移动平台、相机,以解决相关技术中准确度低的技术问题。
根据本公开实施例的第一方面,提出一种白平衡处理方法,所述方法包括:
检测待处理图像上的灰色点;
根据所述灰色点构建直方图;
根据所述直方图估计所述待处理图像的光照,并根据所述光照对所述待处理图像进行白平衡处理。
根据本公开实施例的第二方面,提出一种白平衡处理设备,所述白平衡处理设备包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序,所述处理器执行所述程序时实现以下方法:
检测待处理图像上的灰色点;
根据所述灰色点构建直方图;
根据所述直方图估计所述待处理图像的光照,并根据所述光照对所述待处理图像进行白平衡处理。
根据本公开实施例的第三方面,提出一种可移动平台,包括:
机体;
动力***,安装在所述机体内,用于为所述可移动平台提供动力;以及,
任一实施例所述的白平衡处理设备。
根据本公开实施例的第四方面,提出一种相机,包括:
外壳;
镜头组件,设于所述外壳内部;
传感器组件,设于所述外壳内部,用于感知通过所述镜头组件的光并生成电信号;以及,
任一实施例所述的白平衡处理设备。
根据本公开实施例的第五方面,提出一种计算机可读存储介质,所述可读存储介质上存储有若干计算机指令,所述计算机指令被执行时实现任一实施例所述方法的步骤。
应用本说明书实施例方案,先检测图像中的灰色点,再利用所有灰色点构建统计直方图,最后利用直方图并估计光照并进行白平衡处理。由于构建直方图时只采用了对光照估计有用的灰色点,而去掉了可能会干扰光照估计的彩色点,因此,提高了光照估计的准确性;同时,由于不同图像采集装置采集图像的灰色存在共通性,因此,提高了泛化能力。
附图说明
为了更清楚地说明本公开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公开的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动性的前提下,还可以根据这些附图获得其他的附图。
图1是根据本公开的实施例示出的白平衡处理方法流程图。
图2是根据本公开的实施例示出的现有技术与本发明构建的直方图的效果对比图。
图3是根据本公开的实施例示出的一种灰色检测网络的网络框架。
图4是根据本公开的实施例示出的用于实施本本公开的实施例的方法的计算机设备的结构示意图。
图5是根据本公开的实施例示出的可移动平台的框图。
图6是根据本公开的实施例示出的相机的框图。
具体实施方式
这里将详细地对示例性实施例进行说明,其示例表示在附图中。下面的描述涉及附图时,除非另有表示,不同附图中的相同数字表示相同或相似的要素。以下示例性实施例中所描述的实施方式并不代表与本说明书相一致的所有实施方式。相反,它们仅是与如所附权利要求书中所详述的、本说明书的一些方面相一致的装置和方法的例子。
在本说明书使用的术语是仅仅出于描述特定实施例的目的,而非旨在限制本说明书。在本说明书和所附权利要求书中所使用的单数形式的“一种”、“所述”和“该”也旨在包括多数形式,除非上下文清楚地表示其他含义。还应当理解,本文中使用的术语“和/或”是指并包含一个或多个相关联的列出项目的任何或所有可能组合。
应当理解,尽管在本说明书可能采用术语第一、第二、第三等来描述各种信息,但这些信息不应限于这些术语。这些术语仅用来将同一类型的信息彼此区分开。例如,在不脱离本说明书范围的情况下,第一信息也可以被称为第二信息,类似地,第二信 息也可以被称为第一信息。取决于语境,如在此所使用的词语“如果”可以被解释成为“在……时”或“当……时”或“响应于确定”。
如图1所示,是根据本公开的实施例示出的白平衡处理方法流程图。所述方法可包括:
步骤S101:检测待处理图像上的灰色点;
步骤S102:根据所述灰色点构建直方图;
步骤S103:根据所述直方图估计所述待处理图像的光照,并根据所述光照对所述待处理图像进行白平衡处理。
在本实施例中,当一幅图像上的某个像素点满足:该像素点在所述图像上的R通道颜色分量、G通道颜色分量和B通道颜色分量相等,则该像素点为所述图像上的灰色点。
由于光照的影响,图像中的灰色会显示成其他颜色。现有的灰色检测方法基本基于一些简单的先验假设,例如,假设RGB三通道局部对比度是相同的像素点为灰色。这种简单的假设在现实场景中复杂的情况可能会不满足,从而影响灰色点的估计。为了解决上述问题,本说明书一个实施例通过机器学习模型来检测灰色点。
具体来说,可以将所述待处理图像输入预先训练的机器学***衡处理效果。
在一个实施例中,用来检测灰色点的机器学习模型为卷积神经网络模型(Convolutional Neural Networks,CNN)。这里也可以采用其他的神经网络模型,或者采用其他的机器学习模型。由于CNN的表达能力比较强,能够有效的提取关键特征,因此,采用CNN来检测灰色点能够进一步提高检测准确性。
上述卷积神经网络模型可以根据训练图像以及所述训练图像上各个像素点的训练标签,并根据预设的损失函数训练得到。现存已有的数据库中,给定的是图像的光照而不是灰色点的标签。对灰色点打标签不仅费时长,而且代价高。由于白光照射下的灰色定义是明确的,即灰色点满足:
Figure PCTCN2019106950-appb-000001
其中,
Figure PCTCN2019106950-appb-000002
Figure PCTCN2019106950-appb-000003
分别是训练图像上灰色点在R、G、B三个颜色通道上的颜色值,y是常量。
考虑到光照L的影响,灰色点的观测颜色
Figure PCTCN2019106950-appb-000004
会发生以下变化:
Figure PCTCN2019106950-appb-000005
L c表示光照L在R、G、B颜色通道上的颜色值。
因此,可以对观测图像I生成是否为灰色点的标签,即,分别根据所述训练图像上各个像素点的RGB值与所述训练图像上灰色点的RGB值计算对应像素点为灰色点的第二概率值;根据所述第二概率值确定所述训练图像上各个像素点的训练标签,即:
Figure PCTCN2019106950-appb-000006
其中,p(h,w)表示所述训练图像上坐标为(h,w)处的像素点对应的第二概率值,<,>表示内积操作,I c(h,w)表示所述训练图像上坐标为(h,w)处的像素点的RGB值,
Figure PCTCN2019106950-appb-000007
表示所述训练图像上灰色点的RGB值,c∈{r,g,b}。
若训练图像上一个像素点与灰色点
Figure PCTCN2019106950-appb-000008
的内积值越大,表示这两个点的余弦距离越小,这个像素点是灰色的概率(即,第二概率值)就越大,反之亦然。通过上述方式检测灰色点,避免了对灰色点逐个进行打标,提高了灰色点检测效率。
得到灰色点的标签之后,可以将训练图像全图作为输入,并根据损失函数来训练出CNN。为了输出每个点的灰色概率,本说明书实施例采用Unet神经网络结构,保持输出的空间尺寸。Unet神经网络的输入为H*W*3的彩色图像,输出为H*W的第三概率值,记为
Figure PCTCN2019106950-appb-000009
其中,H和W分别为训练图像的像素高度和像素宽度。
在一个实施例中,损失函数为交叉熵函数。具体来说,一个实施例的损失函数可以根据如下公式得到,即:
Figure PCTCN2019106950-appb-000010
其中,Loss为所述损失函数,
Figure PCTCN2019106950-appb-000011
为所述卷积神经网络模型输出的所述训练图像上坐标为(h,w)处的像素点对应的第三概率值,p(h,w)表示所述训练图像上坐标为 (h,w)处的像素点对应的第二概率值。
根据上述损失函数以及训练图像及其灰色点标签来训练出模型之后,在测试阶段,可以将测试图输入到网络中,得到预测的第二概率值,再设定阈值λ,若测试图像中像素点的预测的第二概率值大于阈值λ,则判定该像素点为灰色点;否则,判定该像素点为彩色点。根据预测的第二概率值的准确率可以验证训练出的机器学习模型的性能,如果准确率大于预设的准确率阈值,就可以根据训练出的模型来检测待处理图像上的灰色点。
白平衡问题的数学模型可以简化成以下形式:
Figure PCTCN2019106950-appb-000012
其中,I和W分别表示观测图像和自然白光照射下的图像(即,理想图像),L表示光照,c表示R、G、B三个颜色通道,n是图像(image)的长宽下标。白平衡过程即给定观测图像I,估计光照L,并通过W=I/L去除光照影响,将图像恢复成白光照射的图像。
为了便于处理,可以将RGB空间上的颜色值映射到预设的目标空间,得到所述RGB值在所述目标空间上的映射值,再根据该映射值构建直方图。在一个实施例中,在该目标空间上,图像与光照之间的乘性关系转换为加性关系。具体来说,将所述灰色点在所述待处理图像上的RGB值映射到预设的目标空间,得到所述RGB值在所述目标空间上的映射值的步骤包括:根据所述灰色点的R通道颜色分量和G通道颜色分量生成所述目标空间的第一映射值,根据所述灰色点的G通道颜色分量和B通道颜色分量生成所述目标空间的第二映射值。假设第一映射值和第二映射值分别用u和v表示,则:
Figure PCTCN2019106950-appb-000013
其中,r,g和b分别为所述灰色点在所述待处理图像上的R通道颜色分量、G通道颜色分量和B通道颜色分量。
将观测图像I,理想图像W和光照L都由原先的RGB空间转换到uv空间,如下:
I u=W u+L u,I v=W v+L v
其中,I u,W u和L u分别为观测图像I,理想图像W和光照L在目标空间的第一映射值,I v,W v和L v分别为观测图像I,理想图像W和光照L在目标空间的第二映射值。
图像W与光照L在RGB空间符合乘性关系,由于log函数的性质,转到uv空间后,乘性关系转换成了加性关系。对uv空间的图像{W u,W v}计算二维统计直方图,由于加性的关系,直方图的图样的偏离对应了{L u,L v},再转换回到RGB空间,就可以得到光照{L r,L g,L b}。
在根据映射值构建直方图时,可以将目标空间的每个维度均匀地划分成若干等分,得到多个区间;根据所述映射值分别统计落在各个区间内的灰色点的个数;根据各个区间对应的灰色点的个数构建直方图。
例如,可以将uv空间的u轴和v轴均划分为N等分,以得到N*N个区间(N为正整数),其中,第i个空间内的灰色点满足:U i1≤u i<U i2,V i1≤v i<V i2,1≤i≤N,u i和v i分别表示第i个区间内灰色点的第一映射值和第二映射值,U i1和U i2分别表示第i个区间的第一映射值的取值下限和取值上限,V i1和V i2分别表示第i个区间的第二映射值的取值下限和取值上限。当一个灰色点的第一映射值落入对应区间内第一映射值的取值范围内,且该灰色点的第二映射值落入对应区间内第二映射值的取值范围内,则该灰色点为所述区间内的灰色点。从而可以根据各个灰色点的第一映射值和第二映射值以及各个区间的第一映射值的取值上下限计算各个区间内像素点的个数。
由于灰色点只占据了颜色空间的很少一部分,因此可以采用较少的像素点来构建直方图,而不是采用全图中所有像素点来构建直方图。例如,在本说明书实施例中,可以采用64*64的直方图代替256*256的直方图。图2给出了现有技术与本发明构建的直方图的效果对比图,其中左边是现有技术中构建的直方图,右图是本发明实施例构建的直方图。可以看出,本发明提出的基于灰色点的直方图,相对于现有技术的直方图,去除了直方图中离散的噪声图样,从而使得后续的光照估计更加简单且准确。
对于上述计算出的直方图,本说明书实施例可以采用一个预先选定的滤波器模板与直方图做滑窗卷积,得到滤波后的响应,然后将响应最大的像素点的光照作为所述待处理图像的光照。即:
Figure PCTCN2019106950-appb-000014
其中,
Figure PCTCN2019106950-appb-000015
Figure PCTCN2019106950-appb-000016
分别表示第一映射值的估计值和第二映射值的估计值,H I表示直方图,F表示滤波器模板。
然后,再将
Figure PCTCN2019106950-appb-000017
Figure PCTCN2019106950-appb-000018
逆映射到RGB空间,即可得到光照{L r,L g,L b}。最后,可以根据光照{L r,L g,L b}进行白平衡处理。除此之外,在得到
Figure PCTCN2019106950-appb-000019
Figure PCTCN2019106950-appb-000020
之后,也可以先根据
Figure PCTCN2019106950-appb-000021
Figure PCTCN2019106950-appb-000022
进行白平衡处理,分别得到uv空间对应的W u和W v,然后,再将W u和W v逆映射到RGB空间,从而生成理想图像。
传统的白平衡处理方式一般都是采用待处理图像中的每一个像素点来进行白平衡处理,然而,一些大幅度偏离灰色的彩色点对最终的估计光照可能是没有作用,甚至起负面作用的。将包括彩色点在内的所有像素点一起用于白平衡处理可能会造成直方图中图样过于分散,包含大偏离的噪声点,影响最后的光照计算,从而导致白平衡处理效果较差。
本公开的实施例示出的一种灰色检测网络的网络框架如图3所示。本实施例通过将待处理图像输入机器学***衡处理效率。
为了验证本说明书实施例提出的算法的效果,在两个常用的白平衡数据库上做实验:重新处理的CCD数据库和NUS数据库。对比实验结果分别见表1和表2。其中CCD数据库包含568张图像,由同一个相机拍摄采集,NUS数据库包含1736张图像,分别由8个不同相机拍摄。训练和测试时,用于计算真实光照的标准色板都被扣除。本发明采用3重交叉验证的方法,并采用角度误差作为数值评价标准。
目前基于学***衡方法都对不同的相机传感器训练不同的白平衡模型,因此模型的泛化能力也成为评价白平衡算法的重要指标,即在一个数据集上训练的模型,能否很好的对其他数据集进行处理。本说明书实施例采用CCD数据集上训练的不同白平衡模型,在NUS数据集上进行测试,反之亦然。本说明书实施例对比了现存的白平衡效果最好的学习算法FFCC(Fast Fourier Color Constancy,快速傅里叶色彩恒常性) 和FC 4,实验结果见表3所示。
Figure PCTCN2019106950-appb-000023
表1不同白平衡算法在CCD数据集上的效果比较
Figure PCTCN2019106950-appb-000024
表2不同白平衡算法在NUS数据集上效果比较
Figure PCTCN2019106950-appb-000025
表3白平衡跨库实验结果
根据上述表格可以看出,得益于有效的灰色检测网络,本说明书实施例实现了排列首位的数值效果。并且,由于不同相机拍摄图像的灰色存在共通性,因此本说明书实施例提出的基于灰色检测的白平衡方法,比其他学习算法表现出更好的跨库泛化能力。
本说明书实施例的白平衡处理设备例如可以是服务器或终端设备。方法实施例可以通过软件实现,也可以通过硬件或者软硬件结合的方式实现。以软件实现为例,作为一个逻辑意义上的装置,是通过其所在文件处理的处理器将非易失性存储器中对 应的计算机程序指令读取到内存中运行形成的。从硬件层面而言,如图4所示,为实施本说明书方法的白平衡处理设备400的一种硬件结构图,除了图4所示的处理器401、内存402、网络接口403、以及非易失性存储器404之外,实施例中用于实施本说明书方法的白平衡处理设备,通常根据该白平衡处理设备的实际功能,还可以包括其他硬件,对此不再赘述。
本说明书实施例提供一种白平衡处理设备,所述白平衡处理设备包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序,所述处理器执行所述程序时实现以下方法:
检测待处理图像上的灰色点;
根据所述灰色点构建直方图;
根据所述直方图估计所述待处理图像的光照,并根据所述光照对所述待处理图像进行白平衡处理。
传统的白平衡处理方式一般都是采用待处理图像中的每一个像素点来进行白平衡处理,然而,一些大幅度偏离灰色的彩色点对最终的估计光照可能是没有作用,甚至起负面作用的。将包括彩色点在内的所有像素点一起用于白平衡处理可能会造成直方图中图样过于分散,包含大偏离的噪声点,影响最后的光照计算,从而导致白平衡处理效果较差。
本说明书实施例由于构建直方图时只采用了对光照估计有用的灰色点,而去掉了可能会干扰光照估计的彩色点,因此,提高了光照估计的准确性;同时,由于不同图像采集装置采集图像的灰色存在共通性,因此,提高了泛化能力。另外,由于灰色点只占据了待处理图像上像素点的很少一部分,因此减小了数据量,降低了计算复杂度,提高了白平衡处理效率。
进一步地,所述处理器根据所述灰色点构建直方图的步骤包括:将所述灰色点在所述待处理图像上的RGB值映射到预设的目标空间,得到所述RGB值在所述目标空间上的映射值;根据所述映射值构建直方图。在一个实施例中,在该目标空间上,图像与光照之间的乘性关系转换为加性关系。通过空间转换,将乘性关系转换为加性关系,从而便于处理,减小了计算复杂度。
进一步地,所述RGB值包括R通道颜色分量、G通道颜色分量和B通道颜色分量,所述映射值包括第一映射值和第二映射值;所述处理器将所述灰色点在所述待 处理图像上的RGB值映射到预设的目标空间,得到所述RGB值在所述目标空间上的映射值的步骤包括:根据所述灰色点的R通道颜色分量和G通道颜色分量生成所述目标空间的第一映射值,根据所述灰色点的G通道颜色分量和B通道颜色分量生成所述目标空间的第二映射值。假设第一映射值和第二映射值分别用u和v表示,则:
Figure PCTCN2019106950-appb-000026
其中,r,g和b分别为所述灰色点在所述待处理图像上的R通道颜色分量、G通道颜色分量和B通道颜色分量。
进一步地,所述处理器检测待处理图像上的灰色点的步骤包括:将所述待处理图像输入预先训练的机器学习模型,以获取待处理图像上各个像素点为灰色点的第一概率值;将第一概率值大于预设值的像素点作为灰色点。
进一步地,所述机器学习模型为卷积神经网络模型。由于CNN的表达能力比较强,能够有效的提取关键特征,因此,采用CNN来检测灰色点能够进一步提高检测准确性。
进一步地,所述处理器执行所述程序时还实现以下方法:根据训练图像以及所述训练图像上各个像素点的训练标签,并根据预设的损失函数训练所述卷积神经网络模型。
进一步地,所述处理器执行所述程序时还实现以下方法:分别根据所述训练图像上各个像素点的RGB值与所述训练图像上灰色点的RGB值计算对应像素点为灰色点的第二概率值;根据所述第二概率值确定所述训练图像上各个像素点的训练标签。
现存已有的数据库中,给定的是图像的光照而不是灰色点的标签。对灰色点打标签不仅费时长,而且代价高。通过上述方式检测灰色点,避免了对灰色点逐个进行打标,提高了灰色点检测效率。
进一步地,所述处理器分别根据所述训练图像上各个像素点的RGB值与所述训练图像上灰色点的RGB值计算对应像素点为灰色点的第二概率值的步骤包括:分别计算所述训练图像上灰色点的RGB值与所述训练图像上各个像素点的RGB值的内积;根据所述内积计算对应像素点为灰色点的第二概率值。
进一步地,所述损失函数为交叉熵函数。
进一步地,所述处理器根据所述映射值构建直方图的步骤包括:将所述目标空 间的每个维度均匀地划分成若干等分,得到多个区间;根据所述映射值分别统计落在各个区间内的灰色点的个数;根据各个区间对应的灰色点的个数构建直方图。
进一步地,所述处理器根据所述直方图估计所述待处理图像的光照的步骤包括:采用滤波器模板与直方图做滑窗卷积,得到滤波后的响应;将响应最大的像素点的光照作为所述待处理图像的光照。
本说明书实施例的白平衡处理设备中的处理器所执行的方法的其他实施例与上述白平衡处理方法的实施例相同,此处不再赘述。
如图5所示,本说明书实施例还提供一种可移动平台500,包括:机体501;动力***502,安装在所述机体501内,用于为所述可移动平台提供动力;以及任一实施例所述的白平衡处理设备400。
可选地,所述可移动平台500为车辆、无人机或者可移动机器人。
如图6所示,本说明书实施例还提供一种相机600,包括:外壳601;镜头组件602,设于所述外壳601内部;传感器组件603,设于所述外壳601内部,用于感知通过所述镜头组件602的光并生成电信号;以及任一实施例所述的白平衡处理设备400。
本说明书实施例还提供一种计算机可读存储介质,所述可读存储介质上存储有若干计算机指令,所述计算机指令被执行时实任一实施例所述方法的步骤。
以上实施例中的各种技术特征可以任意进行组合,只要特征之间的组合不存在冲突或矛盾,但是限于篇幅,未进行一一描述,因此上述实施方式中的各种技术特征的任意进行组合也属于本说明书公开的范围。
本说明书实施例可采用在一个或多个其中包含有程序代码的存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。计算机可用存储介质包括永久性和非永久性、可移动和非可移动媒体,可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括但不限于:相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带,磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质,可用于存储可以被计算 设备访问的信息。
本领域技术人员在考虑说明书及实践这里公开的说明书后,将容易想到本公开的其它实施方案。本公开旨在涵盖本公开的任何变型、用途或者适应性变化,这些变型、用途或者适应性变化遵循本公开的一般性原理并包括本公开未公开的本技术领域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的,本公开的真正范围和精神由下面的权利要求指出。
应当理解的是,本公开并不局限于上面已经描述并在附图中示出的精确结构,并且可以在不脱离其范围进行各种修改和改变。本公开的范围仅由所附的权利要求来限制。
以上所述仅为本公开的较佳实施例而已,并不用以限制本公开,凡在本公开的精神和原则之内,所做的任何修改、等同替换、改进等,均应包含在本公开保护的范围之内。

Claims (26)

  1. 一种白平衡处理方法,其特征在于,所述方法包括:
    检测待处理图像上的灰色点;
    根据所述灰色点构建直方图;
    根据所述直方图估计所述待处理图像的光照,并根据所述光照对所述待处理图像进行白平衡处理。
  2. 根据权利要求1所述的方法,其特征在于,根据所述灰色点构建直方图的步骤包括:
    将所述灰色点在所述待处理图像上的RGB值映射到预设的目标空间,得到所述RGB值在所述目标空间上的映射值;
    根据所述映射值构建直方图。
  3. 根据权利要求2所述的方法,其特征在于,所述RGB值包括R通道颜色分量、G通道颜色分量和B通道颜色分量,所述映射值包括第一映射值和第二映射值;
    将所述灰色点在所述待处理图像上的RGB值映射到预设的目标空间,得到所述RGB值在所述目标空间上的映射值的步骤包括:
    根据所述灰色点的R通道颜色分量和G通道颜色分量生成所述目标空间的第一映射值,根据所述灰色点的G通道颜色分量和B通道颜色分量生成所述目标空间的第二映射值。
  4. 根据权利要求1所述的方法,其特征在于,检测待处理图像上的灰色点的步骤包括:
    将所述待处理图像输入预先训练的机器学习模型,以获取待处理图像上各个像素点为灰色点的第一概率值;
    将第一概率值大于预设值的像素点作为灰色点。
  5. 根据权利要求4所述的方法,其特征在于,所述机器学习模型为卷积神经网络模型。
  6. 根据权利要求5所述的方法,其特征在于,所述方法还包括:
    根据训练图像以及所述训练图像上各个像素点的训练标签,并根据预设的损失函数训练所述卷积神经网络模型。
  7. 根据权利要求6所述的方法,其特征在于,所述方法还包括:
    分别根据所述训练图像上各个像素点的RGB值与所述训练图像上灰色点的RGB值计算对应像素点为灰色点的第二概率值;
    根据所述第二概率值确定所述训练图像上各个像素点的训练标签。
  8. 根据权利要求7所述的方法,其特征在于,分别根据所述训练图像上各个像素点的RGB值与所述训练图像上灰色点的RGB值计算对应像素点为灰色点的第二概率值的步骤包括:
    分别计算所述训练图像上灰色点的RGB值与所述训练图像上各个像素点的RGB值的内积;
    根据所述内积计算对应像素点为灰色点的第二概率值。
  9. 根据权利要求8所述的方法,其特征在于,所述损失函数为交叉熵函数。
  10. 根据权利要求1所述的方法,其特征在于,根据所述映射值构建直方图的步骤包括:
    将所述目标空间的每个维度均匀地划分成若干等分,得到多个区间;
    根据所述映射值分别统计落在各个区间内的灰色点的个数;
    根据各个区间对应的灰色点的个数构建直方图。
  11. 根据权利要求1所述的方法,其特征在于,根据所述直方图估计所述待处理图像的光照的步骤包括:
    采用滤波器模板与直方图做滑窗卷积,得到滤波后的响应;
    将响应最大的像素点的光照作为所述待处理图像的光照。
  12. 一种白平衡处理设备,其特征在于,所述白平衡处理设备包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序,所述处理器执行所述程序时实现以下方法:
    检测待处理图像上的灰色点;
    根据所述灰色点构建直方图;
    根据所述直方图估计所述待处理图像的光照,并根据所述光照对所述待处理图像进行白平衡处理。
  13. 根据权利要求12所述的白平衡处理设备,其特征在于,所述处理器根据所述灰色点构建直方图的步骤包括:
    将所述灰色点在所述待处理图像上的RGB值映射到预设的目标空间,得到所述RGB值在所述目标空间上的映射值;
    根据所述映射值构建直方图。
  14. 根据权利要求13所述的白平衡处理设备,其特征在于,所述RGB值包括R通道颜色分量、G通道颜色分量和B通道颜色分量,所述映射值包括第一映射值和第 二映射值;
    所述处理器将所述灰色点在所述待处理图像上的RGB值映射到预设的目标空间,得到所述RGB值在所述目标空间上的映射值的步骤包括:
    根据所述灰色点的R通道颜色分量和G通道颜色分量生成所述目标空间的第一映射值,根据所述灰色点的G通道颜色分量和B通道颜色分量生成所述目标空间的第二映射值。
  15. 根据权利要求12所述的白平衡处理设备,其特征在于,所述处理器检测待处理图像上的灰色点的步骤包括:
    将所述待处理图像输入预先训练的机器学习模型,以获取待处理图像上各个像素点为灰色点的第一概率值;
    将第一概率值大于预设值的像素点作为灰色点。
  16. 根据权利要求15所述的白平衡处理设备,其特征在于,所述机器学习模型为卷积神经网络模型。
  17. 根据权利要求16所述的白平衡处理设备,其特征在于,所述处理器执行所述程序时还实现以下方法:
    根据训练图像以及所述训练图像上各个像素点的训练标签,并根据预设的损失函数训练所述卷积神经网络模型。
  18. 根据权利要求17所述的白平衡处理设备,其特征在于,所述处理器执行所述程序时还实现以下方法:
    分别根据所述训练图像上各个像素点的RGB值与所述训练图像上灰色点的RGB值计算对应像素点为灰色点的第二概率值;
    根据所述第二概率值确定所述训练图像上各个像素点的训练标签。
  19. 根据权利要求18所述的白平衡处理设备,其特征在于,所述处理器分别根据所述训练图像上各个像素点的RGB值与所述训练图像上灰色点的RGB值计算对应像素点为灰色点的第二概率值的步骤包括:
    分别计算所述训练图像上灰色点的RGB值与所述训练图像上各个像素点的RGB值的内积;
    根据所述内积计算对应像素点为灰色点的第二概率值。
  20. 根据权利要求19所述的白平衡处理设备,其特征在于,所述损失函数为交叉熵函数。
  21. 根据权利要求12所述的白平衡处理设备,其特征在于,所述处理器根据所述 映射值构建直方图的步骤包括:
    将所述目标空间的每个维度均匀地划分成若干等分,得到多个区间;
    根据所述映射值分别统计落在各个区间内的灰色点的个数;
    根据各个区间对应的灰色点的个数构建直方图。
  22. 根据权利要求12所述的白平衡处理设备,其特征在于,所述处理器根据所述直方图估计所述待处理图像的光照的步骤包括:
    采用滤波器模板与直方图做滑窗卷积,得到滤波后的响应;
    将响应最大的像素点的光照作为所述待处理图像的光照。
  23. 一种可移动平台,其特征在于,包括:
    机体;
    动力***,安装在所述机体内,用于为所述可移动平台提供动力;以及,
    如权利要求12至22任意一项所述的白平衡处理设备。
  24. 根据权利要求23所述的可移动平台,其特征在于,所述可移动平台为车辆、无人机或者可移动机器人。
  25. 一种相机,其特征在于,包括:
    外壳;
    镜头组件,设于所述外壳内部;
    传感器组件,设于所述外壳内部,用于感知通过所述镜头组件的光并生成电信号;以及,
    如权利要求12至22任意一项所述的白平衡处理设备。
  26. 一种计算机可读存储介质,其特征在于,所述可读存储介质上存储有若干计算机指令,所述计算机指令被执行时实现权利要求1至11任一项所述方法的步骤。
PCT/CN2019/106950 2019-09-20 2019-09-20 白平衡处理方法和设备、可移动平台、相机 WO2021051382A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN201980033268.6A CN112204957A (zh) 2019-09-20 2019-09-20 白平衡处理方法和设备、可移动平台、相机
PCT/CN2019/106950 WO2021051382A1 (zh) 2019-09-20 2019-09-20 白平衡处理方法和设备、可移动平台、相机

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/106950 WO2021051382A1 (zh) 2019-09-20 2019-09-20 白平衡处理方法和设备、可移动平台、相机

Publications (1)

Publication Number Publication Date
WO2021051382A1 true WO2021051382A1 (zh) 2021-03-25

Family

ID=74004735

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/106950 WO2021051382A1 (zh) 2019-09-20 2019-09-20 白平衡处理方法和设备、可移动平台、相机

Country Status (2)

Country Link
CN (1) CN112204957A (zh)
WO (1) WO2021051382A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114630095A (zh) * 2022-03-15 2022-06-14 锐迪科创微电子(北京)有限公司 目标场景图像的自动白平衡方法及装置、终端
WO2022257713A1 (zh) * 2021-06-07 2022-12-15 荣耀终端有限公司 一种ai自动白平衡的算法和电子设备

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114757856B (zh) * 2022-06-16 2022-09-20 深圳深知未来智能有限公司 一种基于无监督深度学***衡算法及***

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070091185A1 (en) * 2005-10-20 2007-04-26 Samsung Electronics Co., Ltd. Method and apparatus for color temperature correction in a built-in camera of a portable terminal
CN101957988A (zh) * 2009-07-20 2011-01-26 华为技术有限公司 获得图像灰度点概率分布的方法、装置及白平衡方法、装置
CN103491357A (zh) * 2013-10-14 2014-01-01 旗瀚科技有限公司 一种图像传感器白平衡处理方法
CN103974053A (zh) * 2014-05-12 2014-08-06 华中科技大学 一种基于灰点提取的自动白平衡矫正方法
CN104954772A (zh) * 2015-06-26 2015-09-30 济南中维世纪科技有限公司 一种应用于自动白平衡算法的图像近灰色像素选取算法
CN107578390A (zh) * 2017-09-14 2018-01-12 长沙全度影像科技有限公司 一种使用神经网络进行图像白平衡校正的方法及装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9336582B1 (en) * 2015-04-17 2016-05-10 Google Inc. Convolutional color correction
US9794540B2 (en) * 2015-04-17 2017-10-17 Google Inc. Hardware-based convolutional color correction in digital images
US10949958B2 (en) * 2016-11-15 2021-03-16 Google Llc Fast fourier color constancy

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070091185A1 (en) * 2005-10-20 2007-04-26 Samsung Electronics Co., Ltd. Method and apparatus for color temperature correction in a built-in camera of a portable terminal
CN101957988A (zh) * 2009-07-20 2011-01-26 华为技术有限公司 获得图像灰度点概率分布的方法、装置及白平衡方法、装置
CN103491357A (zh) * 2013-10-14 2014-01-01 旗瀚科技有限公司 一种图像传感器白平衡处理方法
CN103974053A (zh) * 2014-05-12 2014-08-06 华中科技大学 一种基于灰点提取的自动白平衡矫正方法
CN104954772A (zh) * 2015-06-26 2015-09-30 济南中维世纪科技有限公司 一种应用于自动白平衡算法的图像近灰色像素选取算法
CN107578390A (zh) * 2017-09-14 2018-01-12 长沙全度影像科技有限公司 一种使用神经网络进行图像白平衡校正的方法及装置

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022257713A1 (zh) * 2021-06-07 2022-12-15 荣耀终端有限公司 一种ai自动白平衡的算法和电子设备
CN114630095A (zh) * 2022-03-15 2022-06-14 锐迪科创微电子(北京)有限公司 目标场景图像的自动白平衡方法及装置、终端
CN114630095B (zh) * 2022-03-15 2024-02-09 锐迪科创微电子(北京)有限公司 目标场景图像的自动白平衡方法及装置、终端

Also Published As

Publication number Publication date
CN112204957A (zh) 2021-01-08

Similar Documents

Publication Publication Date Title
CN106960446B (zh) 一种面向无人艇应用的水面目标检测跟踪一体化方法
CN107230225B (zh) 三维重建的方法和装置
TWI676963B (zh) 目標獲取方法及設備
CN109903331B (zh) 一种基于rgb-d相机的卷积神经网络目标检测方法
WO2021051382A1 (zh) 白平衡处理方法和设备、可移动平台、相机
JP2018022360A (ja) 画像解析装置、画像解析方法およびプログラム
CN111144349B (zh) 一种室内视觉重定位方法及***
CN109063549B (zh) 基于深度神经网络的高分辨率航拍视频运动目标检测方法
WO2019007253A1 (zh) 图像识别方法、装置及设备、可读介质
US11651581B2 (en) System and method for correspondence map determination
CN108182421A (zh) 视频分割方法和装置
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
CN110276831B (zh) 三维模型的建构方法和装置、设备、计算机可读存储介质
US11741615B2 (en) Map segmentation method and device, motion estimation method, and device terminal
TW201421424A (zh) 影像追蹤裝置及其影像追蹤方法
CN110992393B (zh) 一种基于视觉的目标运动跟踪方法
CN110443228B (zh) 一种行人匹配方法、装置、电子设备及存储介质
CN109740527B (zh) 一种视频帧中图像处理方法
CN112396016B (zh) 一种基于大数据技术的人脸识别***
DE102020207974B4 (de) Systeme und verfahren zum nachweis von bewegung während 3d-datenrekonstruktion
CN112329662B (zh) 基于无监督学习的多视角显著性估计方法
CN117726747A (zh) 补全弱纹理场景的三维重建方法、装置、存储介质和设备
CN117036442A (zh) 一种鲁棒单目深度补全方法、***及储存介质
WO2019230965A1 (ja) 物体らしさ推定装置、方法、およびプログラム
CN112101303B (zh) 一种图像数据处理方法、装置以及计算机可读存储介质

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: 19945959

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: 19945959

Country of ref document: EP

Kind code of ref document: A1