WO2023125503A1 - 一种暗光图像降噪方法及装置 - Google Patents

一种暗光图像降噪方法及装置 Download PDF

Info

Publication number
WO2023125503A1
WO2023125503A1 PCT/CN2022/142230 CN2022142230W WO2023125503A1 WO 2023125503 A1 WO2023125503 A1 WO 2023125503A1 CN 2022142230 W CN2022142230 W CN 2022142230W WO 2023125503 A1 WO2023125503 A1 WO 2023125503A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
network model
noise
dark light
noise reduction
Prior art date
Application number
PCT/CN2022/142230
Other languages
English (en)
French (fr)
Inventor
朱才志
王林
周晓
汝佩哲
Original Assignee
英特灵达信息技术(深圳)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 英特灵达信息技术(深圳)有限公司 filed Critical 英特灵达信息技术(深圳)有限公司
Publication of WO2023125503A1 publication Critical patent/WO2023125503A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20172—Image enhancement details

Definitions

  • the present application relates to the technical field of image processing, and in particular to a method and device for image noise reduction in dark light.
  • the imaging element of an imaging device is usually a CCD (Charge-coupled Device, charge-coupled device) or CMOS (Complementary Metal-Oxide-Semiconductor, Complementary Metal Oxide Semiconductor), which contains millions of photosensitive units. If a photosensitive unit is damaged, Then it becomes a bad pixel, and the pixel value of the corresponding pixel position in the image will be significantly different from the surrounding pixels. For low-light images, dead pixels are usually high-brightness dead pixels.
  • FIG. 1 is a schematic diagram of the influence of dead pixels in the dark light image in the RAW domain after noise reduction processing using the existing technology. There is a highlighted dead pixel in the original dark light image before the noise reduction processing, as shown in Figure 1 As shown in the figure, after the noise reduction process, the dead pixel has affected multiple surrounding pixels.
  • the purpose of the embodiments of the present application is to provide a method and device for denoising a low-light image, so as to significantly reduce the influence of bad pixels of an image on the noise-reduction process of a low-light image, and improve the noise-reduction quality of a low-light image.
  • the specific technical scheme is as follows:
  • the present application provides a method for image noise reduction in dark light, the method comprising:
  • the noise reduction network model is trained according to sample images, and the Sample images include: simulated bad pixel dark light images and noise-free images;
  • the step of performing preset image enhancement transformation on the RAW domain dark-light image includes:
  • the following steps are used to train the noise reduction network model:
  • the converged initial neural network model is determined as the noise reduction network model.
  • the following steps are used to obtain the simulated bad pixel dark light image:
  • the embodiment of the present application also provides a low-light image noise reduction device, the device comprising:
  • An acquisition module configured to acquire dark light images in the RAW domain
  • the noise reduction module is used to perform preset image enhancement transformation on the RAW domain dark light image, and input the transformed image into a pre-trained noise reduction network model to obtain an output image; wherein, the noise reduction network model is based on the sample
  • the sample images include: simulated bad pixel dark light images and noise-free images;
  • An inverse transform module configured to perform an inverse transform of the preset image enhancement transform on the output image to obtain a noise-reduced image.
  • the noise reduction module includes an enhanced transformation submodule, and the enhanced transformation submodule is specifically used for:
  • a training module is also included, and the training module is specifically used for:
  • the converged initial neural network model is determined as the noise reduction network model.
  • generating module is specifically used for:
  • the embodiment of the present application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory complete mutual communication through the communication bus;
  • the processor is used for implementing the steps of any one of the methods for reducing noise in low-light images when executing the program stored in the memory.
  • the embodiment of the present application also provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the steps of any one of the methods for reducing noise in dark light images above are implemented.
  • the noise reduction network model can perform noise reduction processing on the input RAW domain image, and automatically Perform bad point suppression, that is, greatly reduce the impact of bad points in the image on image noise reduction, and improve the quality of image noise reduction.
  • the noise reduction network model can automatically suppress dead pixels during the image noise reduction process, there is no need to specifically perform dead point correction to avoid loss of image edge details during the bad point correction process.
  • the image enhancement transformation is performed first to effectively suppress the maximum bad point and enhance the details of the dark part of the image, which is more conducive to noise reduction in the subsequent network model. Improve image noise reduction quality.
  • Fig. 1 is a schematic diagram of the influence of dead pixels in dark light images in the RAW domain after noise reduction processing using the prior art
  • FIG. 2 is a schematic flow chart of a method for reducing noise in dark-light images provided by an embodiment of the present application
  • Figure 3(a) is a schematic diagram of the results of noise reduction processing using the prior art
  • Figure 3(b) is a schematic diagram of the results of noise reduction processing using the dark light image noise reduction method provided by the embodiment of the present application;
  • FIG. 4 is a schematic structural diagram of a noise reduction network model provided in an embodiment of the present application.
  • FIG. 5 is a schematic diagram of the influence of bad pixels in a dark light image in the RAW domain after being processed by the dark light image noise reduction method provided by the embodiment of the present application;
  • FIG. 6 is a schematic structural diagram of a dark-light image noise reduction device provided in an embodiment of the present application.
  • FIG. 7 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
  • inventions of the present application provide a method and device for noise reduction of dark-light images, which can be applied to electronic equipment, such as
  • the electronic device may be a device capable of image processing such as a desktop computer, a server, a tablet computer, or a mobile phone.
  • FIG. 2 is a schematic flow chart of a dark light image noise reduction method provided in the embodiment of the present application. As shown in FIG. 2, the method includes the following steps:
  • the low-light image noise reduction method provided in the embodiments of the present application is aimed at low-light images in the RAW domain. In this step, the dark light image in the RAW domain that needs noise reduction processing is obtained.
  • the image collected by the imaging device under dark light conditions is a dark light image. Due to the serious lack of light, the collected image contains a lot of noise, making the image unclear and poor in quality. Therefore, noise reduction processing is required.
  • the image output by the sensor is a RAW domain image (also called an original image file), and after the ISP processing is performed on the RAW domain image, the noise properties of the image will become more complicated and difficult to process. Therefore, noise reduction is usually performed before ISP processing, that is, the noise reduction process is directly performed on the RAW domain image.
  • the image directly output by the sensor of the imaging device is the RAW domain dark light image.
  • the dark light condition refers to insufficient light in the shooting environment, that is, insufficient light input of the imaging device.
  • S202 Perform preset image enhancement transformation on the dark-light image in the RAW domain, and input the transformed image into a pre-trained noise reduction network model to obtain an output image; wherein, the noise reduction network model is trained according to sample images, and the sample images include: Simulate bad-pixel dark-light images and noise-free images.
  • the idea of deep learning is adopted to train the neural network model, so as to realize the automatic suppression of dead pixels during the noise reduction process.
  • Image enhancement transformations may include image normalization and gamma transformations. Among them, image normalization can reduce the influence of the highlight part in the RAW domain dark light image, and facilitate the subsequent gamma transformation. Gamma transformation can compress the part with high gray level in the image and stretch the part with low gray level, thereby enhancing the details of dark parts.
  • the image is input to the noise reduction network model. Since the noise reduction network model is pre-trained based on the sample image, the image can be denoised and the output image can be obtained.
  • the sample images used to train the denoising network model include a large number of simulated dark-light images and noise-free images.
  • the simulated dead pixel dark light image and the noise-free image can be in one-to-one correspondence.
  • the sample image contains 1000 simulated dead pixel dark light images and 1000 noise-free images, each simulated dead pixel dark A noise-free image, the image content of the corresponding simulated dead pixel low-light image and the noise-free image are the same, but the image quality is different.
  • the noise-free image in the sample image can be understood as an image in an ideal state, which does not contain noise or dead pixels, while the simulated dead pixel dark light image contains noise and dead pixels.
  • the following steps are used to obtain the dark light image of the simulated dead point:
  • a dark-light noise image that is, a dark-light image containing noise
  • the number of dead pixels in a 1080p resolution image usually ranges from hundreds to thousands, and the number of dead pixels is related to the sensor process. Then, in the embodiment of the present application, some pixel coordinates can be randomly determined on the RAW domain dark light image containing noise, and the pixel values at these positions are set as abnormal values, or pixel values that obviously do not satisfy the Poisson distribution and Gaussian distribution, so as to Simulate the distribution of dead pixels, and generate simulated dark-light images of dead pixels.
  • the noise-free image can be obtained first, and the noise-free image can be processed to obtain the dark light noise image, and then The dead pixels are generated on the dark light noise image, so as to obtain the simulated dead pixel dark light image corresponding to each noise-free image.
  • the ISO value (sensitivity) of the imaging device is adjusted to the lowest value, and a noise-free RAW image is captured. Then divide the pixel value of the pixel in the noise-free image by different multiples, such as 10, 100, and 200 times, to obtain different levels of low-light and noise-free images, that is, the noise-free images in the sample image.
  • the noise-reduction network model is trained in advance by using simulated dark-light images with bad pixels and noise-free images, and the trained noise-reduction network model can perform noise reduction processing on input RAW images.
  • the transformed image is input into the noise reduction network model to obtain an output image.
  • S203 Perform an inverse transformation of a preset image enhancement transformation on the output image to obtain a noise-reduced image.
  • the inverse transformation of the preset image enhancement transformation is performed to obtain the RAW domain image after noise reduction.
  • the inverse transformation may include inverse gamma transformation and inverse normalization transformation in turn.
  • the noise reduction network model can perform noise reduction processing on the input RAW domain image, and automatically Perform bad point suppression, that is, greatly reduce the impact of bad points in the image on image noise reduction, and improve the quality of image noise reduction.
  • the noise reduction network model can automatically suppress dead pixels during the image noise reduction process, there is no need to specifically perform dead point correction to avoid loss of image edge details during the bad point correction process.
  • the image enhancement transformation is performed first to effectively suppress the maximum bad point and enhance the details of the dark part of the image, which is more conducive to noise reduction in the subsequent network model. Improve image noise reduction quality.
  • FIG. 3( a ) is a schematic diagram of the result of noise reduction processing using the prior art
  • FIG. 3( b ) is a schematic diagram of the result of noise reduction processing using the dark light image noise reduction method provided by the embodiment of the present application.
  • Figure 3(a) cannot eliminate the influence of bad pixels, resulting in the image containing many pixels with abnormal pixel values, which looks blurry, that is, the image noise reduction quality is poor.
  • Figure 3(b) adopts the dark light image noise reduction method provided by the embodiment of the present application, which can basically eliminate the influence of bad pixels on image noise reduction, so it basically does not contain pixels with abnormal pixel values, and the image as a whole is cleaner and smoother. clear.
  • the training process of the noise reduction network model is introduced below, and the specific training steps may include:
  • Step 11 Get the initial neural network model and sample images.
  • the initial neural network model may be a convolutional neural network model
  • the structure of the network model may be an encoder-decoder structure.
  • Fig. 4 is a schematic structural diagram of a denoising network model provided by an embodiment of the present application.
  • input represents the input image
  • output represents the input image
  • conv1 represents the first convolution structure
  • kernel size (convolution kernel) 3 ⁇ 3, stride (step size) is 1
  • conv2 represents the first convolution structure
  • Convolution kernel 2 ⁇ 2, step size is 2
  • deconv means deconvolution (transposed convolution) structure
  • convolution kernel 2 ⁇ 2, step size is 2
  • relu means linear rectification function.
  • the sample images include a large number of simulated dead pixel dark-light images and noise-free images corresponding to each other. For details, refer to the relevant introduction in step S202.
  • Step 12 Input the simulated bad pixel dark light image after the preset image enhancement transformation into the initial neural network model; calculate the loss value based on the output result of the initial neural network model and the noise-free image after the preset image enhancement transformation.
  • the preset image enhancement transformation may be performed on the simulated bad pixel dark light image and the noise-free image respectively.
  • Preset image enhancement transformations include normalization and gamma transformations.
  • the simulated bad point dark light image be I, (i, j) be the pixel coordinates, and the pixel value of the pixel point of the coordinates (i, j) be x ij .
  • Normalization processing can significantly reduce the influence of the highlight part in the RAW domain, and the image after normalization processing is represented as I', and then gamma transformation is performed on it.
  • the gamma transformation with a coefficient less than 1 can compress the part with high gray level in the image and stretch the part with low gray level, so as to enhance the details of the dark part of the dark light image.
  • x is the simulated bad pixel dark light image after image enhancement transformation
  • y is the noise-free image after image enhancement transformation
  • f is the functional relationship of network model fitting
  • y is the output result of the neural network model
  • y and The absolute value of the difference can be used as the loss value.
  • computing y and The absolute value of the difference between the pixel values of the corresponding pixel points in the center is averaged to obtain the loss value.
  • Step 13 Adjust the model parameters of the initial neural network based on the loss value, and return to the step of inputting the simulated bad pixel dark light image after the preset image enhancement transformation into the initial neural network model until the initial neural network model converges.
  • the method of backpropagation is used to adjust the model parameters and complete a round of training.
  • step 12 After multiple iterations of training, stable model parameters can be obtained. At this time, the initial neural network model can be considered to be convergent.
  • Step 14 Determine the converged initial neural network model as the noise reduction network model.
  • the dark light image of simulated dead pixels is generated by simulating the distribution of dead pixels, and the noise reduction network model is trained in combination with the noise-free images.
  • the parameters of the network model are constantly adjusted, so that the network model can process the input image, and the processing result is constantly approaching the noise-free image to achieve image noise reduction.
  • the trained denoising network model can be well suited for noise reduction of dark-light images with dead pixels, that is, automatically suppresses the effect of bad pixels on The effect of image noise reduction, improve image noise reduction quality.
  • FIG. 5 is a schematic diagram of the influence of bad pixels in a dark-light image in the RAW domain after being processed by the method for reducing noise in a dark-light image provided by an embodiment of the present application.
  • Figure 5 corresponds to the original low-light image before the noise reduction process, which has a highlighted dead pixel.
  • the highlighted dead pixel is no longer visible. It can be seen that after this After the low-light image noise reduction method provided by the application is processed, it can significantly suppress the highlighted dead pixels in the dark-light image, and the highlighted dead pixels no longer spread to several or even dozens of pixels as shown in Figure 1.
  • Fig. 6 is a schematic structural diagram of a dark light image noise reduction device provided in the embodiment of the present application. As shown in Fig. 6, the device may include:
  • the noise reduction module 602 is used to perform preset image enhancement transformation on the RAW domain dark light image, input the transformed image into the pre-trained noise reduction network model, and obtain the output image; wherein, the noise reduction network model is trained according to the sample image , the sample images include: simulated bad pixel dark light image and noise-free image;
  • the inverse transformation module 603 is configured to perform inverse transformation of the preset image enhancement transformation on the output image to obtain a noise-reduced image.
  • the noise reduction module 602 includes an enhanced transformation submodule, and the enhanced transformation submodule is specifically used for:
  • a training module may also be included, and the training module is specifically used for:
  • the converged initial neural network model is determined as the noise reduction network model.
  • a generation module may also be included, and the generation module is specifically used for:
  • the noise reduction network model can perform noise reduction processing on the input RAW domain image, and automatically Perform bad point suppression, that is, greatly reduce the impact of bad points in the image on image noise reduction, and improve the quality of image noise reduction.
  • the noise reduction network model can automatically suppress dead pixels during the image noise reduction process, there is no need to specifically perform dead point correction to avoid loss of image edge details during the bad point correction process.
  • the image enhancement transformation is performed first to effectively suppress the maximum bad point and enhance the details of the dark part of the image, which is more conducive to noise reduction in the subsequent network model. Improve image noise reduction quality.
  • the embodiment of the present application also provides an electronic device, as shown in FIG. complete the mutual communication,
  • Memory 703 used to store computer programs
  • the noise reduction network model is trained according to sample images, and the Sample images include: simulated bad pixel dark light images and noise-free images;
  • the communication bus mentioned in the above-mentioned electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus or the like.
  • PCI Peripheral Component Interconnect
  • EISA Extended Industry Standard Architecture
  • the communication bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.
  • the communication interface is used for communication between the electronic device and other devices.
  • the memory may include a random access memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-Volatile Memory, NVM), such as at least one magnetic disk memory.
  • RAM Random Access Memory
  • NVM non-Volatile Memory
  • the memory may also be at least one storage device located far away from the aforementioned processor.
  • the above-mentioned processor can be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; it can also be a digital signal processor (Digital Signal Processor, DSP), a dedicated integrated Circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
  • CPU Central Processing Unit
  • NP Network Processor
  • DSP Digital Signal Processor
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • the denoising network model is trained based on sample images, which include: simulated bad pixel dark light images and noise-free images; the inverse transformation of the preset image enhancement transformation is performed on the output image to obtain a denoising image.
  • the noise reduction network model can perform noise reduction processing on the input RAW domain image, and automatically Perform bad point suppression, that is, greatly reduce the impact of bad points in the image on image noise reduction, and improve the quality of image noise reduction.
  • the noise reduction network model can automatically suppress dead pixels during the image noise reduction process, there is no need to specifically perform dead point correction to avoid loss of image edge details during the bad point correction process.
  • the image enhancement transformation is performed first to effectively suppress the maximum bad point and enhance the details of the dark part of the image, which is more conducive to noise reduction in the subsequent network model. Improve image noise reduction quality.
  • a computer-readable storage medium is also provided, and a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, any of the above dark-light images can be realized.
  • the steps of the noise reduction method are also provided, and a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, any of the above dark-light images can be realized. The steps of the noise reduction method.
  • a computer program product including instructions is also provided, which, when run on a computer, causes the computer to execute any method for noise reduction of low-light images in the above embodiments.
  • the computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices.
  • the computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website, computer, server or data center Transmission to another website site, computer, server, or data center by wired (eg, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.).
  • the computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrated with one or more available media.
  • the available medium may be a magnetic medium (such as a floppy disk, a hard disk, or a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a Solid State Disk (SSD)).
  • each embodiment in this specification is described in a related manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments.
  • the description is relatively simple. Please refer to the part description of the embodiment of the image noise reduction method in dark light.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Processing (AREA)

Abstract

本申请实施例提供了一种暗光图像降噪方法及装置，方法包括：获取RAW域暗光图像；对RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，降噪网络模型是根据样本图像训练的，样本图像包括：模拟坏点暗光图像和无噪图像；对输出图像进行预设图像增强变换的逆变换，得到降噪图像。能够显著降低图像坏点对暗光图像降噪过程产生的影响，提高暗光图像的降噪质量。

Description

一种暗光图像降噪方法及装置
本申请要求于2021年12月29日提交中国专利局、申请号为202111641889.8发明名称为“一种暗光图像降噪方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及图像处理技术领域，特别是涉及一种暗光图像降噪方法及装置。
背景技术
随着计算机视觉技术和光学影像技术的快速发展，利用视频图像采集设备获取图像或视频，并对图像或视频进行处理的方式，被广泛应用于安防、海防、智能交通等各个方面。
然而，当成像设备处于暗光条件下时，由于光线严重不足，会导致采集到的图像包含非常多的噪声，使得图像不清晰，画质差，进而，会降低后期图像处理结果的准确度。因此，此类图像在进行分析处理之前需要进行降噪处理，而在降噪过程中，图像存在的坏点，会严重影响图像降噪质量。
成像设备的成像元件通常是CCD(Charge-coupled Device，电荷耦合元件)或CMOS(Complementary Metal-Oxide-Semiconductor，互补金属氧化物半导体)，包含数百万个感光单元，如果某个感光单元损坏，则成为坏点，图像中对应像素位置的像素值会明显异于周围像素点。对于暗光图像，坏点通常为高亮坏点。
在基于深度学习的暗光图像RAW域降噪技术中，本身图像的像素值就相对较小，输入数据中的高亮坏点，经过卷积神经网络的感受野(receptive-field，RF)机制放大，会由单点影响扩散至几个像素甚至几十个像素范围，严重降低了图像的视觉质量。参见图1，图1为采用现有技术进行降噪处理后RAW域暗光图像中坏点影响的一种示意图，降噪处理前的原始暗光图像中有一个高亮坏点，如图1所示，经过降噪处理后，坏点已经影响周边的多个像素点。
虽然很多图像信号处理(Image Signal Processing，ISP)的算法中包含了坏点校正这一环节，但由于坏点的多样性和复杂性，基于传统的图像处理方 法一般不能完全消除，而且，去坏点的强度较大，会导致图像产生类似于中值滤波的平滑效果，损害图像的边缘细节。
发明内容
本申请实施例的目的在于提供一种暗光图像降噪方法及装置，以实现显著降低图像坏点对暗光图像降噪过程产生的影响，提高暗光图像的降噪质量。具体技术方案如下：
本申请提供了一种暗光图像降噪方法，所述方法包括：
获取RAW域暗光图像；
对所述RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，所述降噪网络模型是根据样本图像训练的，所述样本图像包括：模拟坏点暗光图像和无噪图像；
对所述输出图像进行所述预设图像增强变换的逆变换，得到降噪图像。
可选的，所述对所述RAW域暗光图像进行预设图像增强变换的步骤，包括：
依次对所述RAW域暗光图像进行归一化、gamma变换。
可选的，采用如下步骤训练所述降噪网络模型：
获取初始神经网络模型和所述样本图像；
将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型；
基于所述初始神经网络模型输出结果和进行所述预设图像增强变换后的无噪图像，计算损失值；
基于所述损失值调整所述初始神经网络的模型参数，并返回将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型的步骤，直到所述初始神经网络模型收敛；
将收敛的初始神经网络模型确定为所述降噪网络模型。
可选的，采用如下步骤获得所述模拟坏点暗光图像：
获取暗光噪声图像；
基于预先统计的传感器坏点比例，在所述暗光噪声图像中随机生成特定 数量的坏点，得到所述模拟坏点暗光图像，其中，所生成的坏点的像素值为高亮值或所生成的坏点的像素值与周边像素点的像素值不满足预设分布。
本申请实施例还提供了一种暗光图像降噪装置，所述装置包括：
获取模块，用于获取RAW域暗光图像；
降噪模块，用于对所述RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，所述降噪网络模型是根据样本图像训练的，所述样本图像包括：模拟坏点暗光图像和无噪图像；
逆变换模块，用于对所述输出图像进行所述预设图像增强变换的逆变换，得到降噪图像。
可选的，所述降噪模块包含增强变换子模块，所述增强变换子模块，具体用于：
依次对所述RAW域暗光图像进行归一化、gamma变换。
可选的，还包括训练模块，所述训练模块，具体用于：
获取初始神经网络模型和所述样本图像；
将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型；
基于所述初始神经网络模型输出结果和进行所述预设图像增强变换后的无噪图像，计算损失值；
基于所述损失值调整所述初始神经网络的模型参数，并返回将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型的步骤，直到所述初始神经网络模型收敛；
将收敛的初始神经网络模型确定为所述降噪网络模型。
可选的，还包括：生成模块，所述生成模块具体用于：
获取暗光噪声图像；
基于预先统计的传感器坏点比例，在所述暗光噪声图像中随机生成特定数量的坏点，得到所述模拟坏点暗光图像，其中，所生成的坏点的像素值为高亮值或所生成的坏点的像素值与周边像素点的像素值不满足预设分布。
本申请实施例还提供了一种电子设备，包括处理器、通信接口、存储器 和通信总线，其中，处理器，通信接口，存储器通过通信总线完成相互间的通信；
存储器，用于存放计算机程序；
处理器，用于执行存储器上所存放的程序时，实现上述任一暗光图像降噪方法步骤。
本申请实施例还提供了一种计算机可读存储介质，所述计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现上述任一暗光图像降噪方法步骤。
应用本申请实施例提供的暗光图像降噪方法及装置，获取RAW域暗光图像；对RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，降噪网络模型是根据样本图像训练的，样本图像包括：模拟坏点暗光图像和无噪图像；对输出图像进行预设图像增强变换的逆变换，得到降噪图像。
可见，通过模拟坏点的分布情况生成模拟坏点暗光图像，结合无噪图像训练降噪网络模型，在训练完成后，降噪网络模型能够对输入的RAW域图像进行降噪处理，且自动进行坏点抑制，即大幅降低图像中坏点对图像降噪的影响，提高了图像降噪的质量。此外，由于降噪网络模型在图像降噪过程中能够自动进行坏点抑制，因此无需专门进行坏点矫正这一环节，避免在坏点矫正过程中损失图像的边缘细节。
并且，在将RAW域暗光图像输入网络模型之前，先进行图像增强变换，对极大值坏点进行有效的抑制，增强图像暗部细节，更利于在后续的网络模型中进行降噪处理，进一步提高图像降噪质量。
当然，实施本申请的任一产品或方法并不一定需要同时达到以上所述的所有优点。
附图说明
此处所说明的附图用来提供对本申请的进一步理解，构成本申请的一部分，本申请的示意性实施例及其说明用于解释本申请，并不构成对本申请的不当限定。
图1为采用现有技术进行降噪处理后RAW域暗光图像中坏点影响的一种示意图；
图2为本申请实施例提供的暗光图像降噪方法的一种流程示意图；
图3(a)为采用现有技术进行降噪处理的结果示意图；
图3(b)为采用本申请实施例提供的暗光图像降噪方法进行降噪处理的结果示意图；
图4为本申请实施例提供的降噪网络模型的一种结构示意图；
图5为经过本申请实施例提供的暗光图像降噪方法处理后RAW域暗光图像中坏点影响的一种示意图；
图6为本申请实施例提供的暗光图像降噪装置的一种结构示意图；
图7为本申请实施例提供的电子设备的一种结构示意图。
具体实施方式
为使本申请的目的、技术方案、及优点更加清楚明白，以下参照附图并举实施例，对本申请进一步详细说明。显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
为了降低图像坏点对暗光图像降噪过程产生的影响，提高暗光图像的降噪质量，本申请实施例提供了一种暗光图像降噪方法及装置，该方法可以应用电子设备，例如电子设备可以是台式计算机、服务器、平板电脑或者手机等具备图像处理能力的设备。
参见图2，图2为本申请实施例提供的暗光图像降噪方法的一种流程示意图，如图2所示，方法包括以下步骤：
S201：获取RAW域暗光图像。
本申请实施例提供的暗光图像降噪方法针对RAW域暗光图像。本步骤中，获取需要降噪处理的RAW域暗光图像。
成像设备处于暗光条件下采集的图像即为暗光图像，由于光线严重不足，导致采集到的图像包含非常多的噪声，使得图像不清晰，画质差。因此需要进行降噪处理。
本领域技术人员可以理解，传感器输出的图像即为RAW域图像(也叫原始图像文件)，在对RAW域图像进行ISP处理之后，图像的噪声性质会变得更为复杂，难以处理。因此，通常在ISP处理之前进行降噪，也就是直接对RAW域图像进行降噪处理。
综上，成像设备在暗光条件下采集图像时，成像设备的传感器直接输出的图像即为RAW域暗光图像。其中，暗光条件是指拍摄环境的光线不足，即成像设备的进光量不足。
S202：对RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，降噪网络模型是根据样本图像训练的，样本图像包括：模拟坏点暗光图像和无噪图像。
本申请实施例中，采用深度学习的思想，训练神经网络模型，实现在降噪过程中自动进行坏点抑制。
为了增强RAW域暗光图像的暗部细节，在将RAW域暗光图像输入预先训练的降噪网络模型之前，先进行预设图像增强变换。
图像增强变换可以包括图像归一化和gamma变换。其中，图像归一化能够降低RAW域暗光图像中高亮部分的影响，并且便于后续进行gamma变换。gamma变换可以压缩图像中灰度级高的部分，拉伸灰度级低的部分，从而增强暗部细节。
在经过预设图像增强变换后，将图像输入降噪网络模型，由于降噪网络模型是基于样本图像预先训练完成的，因此能够对图像进行降噪处理，得到输出图像。
其中，用于训练降噪网络模型的样本图像包括大量的模拟坏点暗光图像和无噪图像。其中，模拟坏点暗光图像和无噪图像可以是一一对应的，例如，样本图像中包含1000张模拟坏点暗光图像和1000张无噪图像，每张模拟坏点暗光图像对应一张无噪图像，相互对应的模拟坏点暗光图像和无噪图像的图像内容相同，但图像质量不同。
样本图像中的无噪图像可以理解为理想状态下的图像，不包含噪声也不包含坏点，模拟坏点暗光图像则包含噪声与坏点。
本申请的一个实施例中，采用如下步骤获取模拟坏点暗光图像：
获取暗光噪声图像；基于预先统计的传感器坏点比例，在暗光噪声图像中随机生成特定数量的坏点，得到模拟坏点暗光图像，其中，所生成的坏点的像素值为高亮值或所生成的坏点的像素值与周边像素点的像素值不满足预设分布。
具体的，在模拟噪声图像的过程中，先获取暗光噪声图像，即包含噪声的暗光图像，再根据统计的CMOS或CCD传感器的坏点比例情况，随机生成特定数量且位置随机分布的坏点。
例如，通常一张1080p分辨率的图像中坏点数目范围为几百到几千，坏点数目和传感器的工艺有关。则本申请实施例中，可以在包含噪声的RAW域暗光图像上随机确定一些像素坐标，将这些位置的像素值设置为异常值，或明显不满足泊松分布及高斯分布的像素值，以模拟坏点分布情况，生成模拟坏点暗光图像。
由于本申请实施例中样本图像中的模拟坏点暗光图像和无噪图像可以是一一对应的，因此，可以先获取无噪图像，对无噪图像进行处理获得暗光噪声图像，然后在暗光噪声图像上生成坏点，从而得到每张无噪图像对应的模拟坏点暗光图像。
作为一个示例，在正常亮度下，把成像设备的ISO值(感光度)调到最低，拍摄得到无噪RAW图像。然后对无噪图像中像素点的像素值除以不同的倍数，例如10、100、200倍，获得不同级别的暗光无噪图像，即样本图像中的无噪图像。获得暗光无噪图像后，基于物理成像建模为暗光无噪图像添加噪声得到与暗光无噪图像一一对应的暗光噪声图像，由于这样获得的暗光噪声图像中坏点信息不足，所以还需要在每张暗光噪声图像中生成坏点，从而得到与无噪图像一一对应的模拟坏点暗光图像。
本申请实施例中，预先采用模拟坏点暗光图像和无噪图像训练降噪网络模型，训练完成的降噪网络模型能够对输入的RAW图像进行降噪处理。
本步骤中，将变换后的图像输入降噪网络模型，得到输出图像。
S203：对输出图像进行预设图像增强变换的逆变换，得到降噪图像。
在得到降噪网络模型输出的图像后，进行预设图像增强变换的逆变换，即可得到降噪后的RAW域图像。
由于预设图像增强变换依次包括：归一化和gamma变换，因此逆变换可以依次包括逆gamma变换和逆归一化变换。
应用本申请实施例提供的暗光图像降噪方法，获取RAW域暗光图像；对RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，降噪网络模型是根据样本图像训练的，样本图像包括：模拟坏点暗光图像和无噪图像；对输出图像进行预设图像增强变换的逆变换，得到降噪图像。
可见，通过模拟坏点的分布情况生成模拟坏点暗光图像，结合无噪图像训练降噪网络模型，在训练完成后，降噪网络模型能够对输入的RAW域图像进行降噪处理，且自动进行坏点抑制，即大幅降低图像中坏点对图像降噪的影响，提高了图像降噪的质量。此外，由于降噪网络模型在图像降噪过程中能够自动进行坏点抑制，因此无需专门进行坏点矫正这一环节，避免在坏点矫正过程中损失图像的边缘细节。
并且，在将RAW域暗光图像输入网络模型之前，先进行图像增强变换，对极大值坏点进行有效的抑制，增强图像暗部细节，更利于在后续的网络模型中进行降噪处理，进一步提高图像降噪质量。
作为一个示例，参见图3(a)和图3(b)，图3(a)对应的原始图像和图3(b)对应的原始图像相同，且均来自公开数据集。图3(a)为采用现有技术进行降噪处理的结果示意图，图3(b)为采用本申请实施例提供的暗光图像降噪方法进行降噪处理的结果示意图。
可见，图3(a)无法消除坏点的影响，导致图像中包含较多像素值异常的像素点，看上去较为模糊，即图像降噪质量差。而图3(b)采用了本申请实施例提供的暗光图像降噪方法，能够基本消除坏点对图像降噪的影响，因此基本不包含像素值异常的像素点，图像整体更为干净、清晰。
下面对降噪网络模型的训练过程进行介绍，具体的训练步骤可以包括：
步骤11：获取初始神经网络模型和样本图像。
具体的，初始神经网络模型可以是卷积神经网络模型，网络模型的结构可以为编码器-解码器结构。
作为一个示例，参见图4，图4为本申请实施例提供的降噪网络模型的一 种结构示意图。图4中，input表示输入图像，output表示输入图像；conv1表示第一卷积结构，kernel size(卷积核)＝3×3，stride(步长)为1；conv2表示第一卷积结构，卷积核＝2×2，步长为2；deconv表示反卷积(transposed convolution)结构，卷积核＝2×2，步长为2；relu表示线性整流函数。
样本图像包括大量相互对应的模拟坏点暗光图像和无噪图像，具体参见步骤S202中相关介绍。
步骤12：将进行预设图像增强变换后的模拟坏点暗光图像输入初始神经网络模型；基于初始神经网络模型输出结果和进行预设图像增强变换后的无噪图像，计算损失值。
本申请实施例中，可以分别对模拟坏点暗光图像和无噪图像进行预设图像增强变换。预设图像增强变换包括归一化和gamma变换。
以对模拟坏点暗光图像进行预设图像增强变换处理为例，设模拟坏点暗光图像为I，(i，j)为像素坐标，坐标(i，j)的像素点的像素值为x ij。
设模拟坏点暗光图像中像素点最大亮度为max_b，则对模拟坏点暗光图像中像素点x ij进行归一化处理，得到x‘ ij。
用公式可以表示为：
Figure PCTCN2022142230-appb-000001
通常，像素点占用存储空间为10bits，则max_b＝1023；像素点占用存储空间为12bits，则max_b＝4095。
归一化处理能够显著降低RAW域高亮部分的影响，归一化处理后的图像表示为I’，再对其进行gamma变换。
作为一个示例，对I’中像素点x‘ ij做gamma变换，获得x’‘ ij，其中gamma系数可以根据需求进行设定，当gamma系数为
Figure PCTCN2022142230-appb-000002
上述变换可以表示为：
Figure PCTCN2022142230-appb-000003
系数小于1的gamma变换，能够压缩图像中灰度级高的部分，拉伸灰度级低的部分，从而增强暗光图像的暗部细节。
将进行预设图像增强变换后的模拟坏点暗光图像输入初始神经网络模型，得到初始神经网络模型的输出结果，再结合进行预设图像增强变换后的无噪 图像，可以计算损失值。
用公式可以表示为：
Figure PCTCN2022142230-appb-000004
Figure PCTCN2022142230-appb-000005
其中，x为图像增强变换后的模拟坏点暗光图像，y为图像增强变换后的无噪图像，f为网络模型拟合的函数关系，
Figure PCTCN2022142230-appb-000006
为神经网络模型输出的结果，y与
Figure PCTCN2022142230-appb-000007
的差的绝对值可以作为损失值。
作为一个示例，分别计算y和
Figure PCTCN2022142230-appb-000008
中对应位置像素点的像素值的差值绝对值，再进行平均，得到损失值。
步骤13：基于损失值调整初始神经网络的模型参数，并返回将进行预设图像增强变换后的模拟坏点暗光图像输入初始神经网络模型的步骤，直到初始神经网络模型收敛。
根据损失值采用反向传播的方式，调整模型参数，完成一轮训练。
随后返回步骤12进行下一轮的训练，在经过多次迭代训练后能够得到稳定的模型参数，此时可以认为初始神经网络模型收敛。
步骤14：将收敛的初始神经网络模型确定为降噪网络模型。
可见，本申请实施例中，通过模拟坏点的分布情况生成模拟坏点暗光图像，结合无噪图像训练降噪网络模型。在训练过程中，不断调整网络模型的参数，使得网络模型能够对输入图像进行处理，且处理结果不断接近无噪图像，实现图像降噪。并且，由于输入图像是模拟的带有坏点的暗光图像，因此，训练完成的降噪网络模型能够很好的适用于带有坏点的暗光图像的降噪，即自动抑制坏点对图像降噪的影响，提高图像降噪质量。
作为一个示例，参见图5，图5为经过本申请实施例提供的暗光图像降噪方法处理后RAW域暗光图像中坏点影响的一种示意图。
图5对应的降噪处理前的原始暗光图像中有一个高亮坏点，经过本申请实施例提供的暗光图像降噪方法处理后，已看不到高亮坏点，可见，经过本申请提供的暗光图像降噪方法处理后，对暗光图像中高亮坏点起到明显的抑制作用，高亮坏点不再如图1所示那样扩散至几个甚至几十个像素范围。
参见图6，图6为本申请实施例提供的暗光图像降噪装置的一种结构示意 图，如图6所示，装置可以包括：
获取模块601，用于获取RAW域暗光图像；
降噪模块602，用于对RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，降噪网络模型是根据样本图像训练的，样本图像包括：模拟坏点暗光图像和无噪图像；
逆变换模块603，用于对输出图像进行预设图像增强变换的逆变换，得到降噪图像。
在本申请的一个实施例中，降噪模块602包含增强变换子模块，所述增强变换子模块，具体用于：
依次对所述RAW域暗光图像进行归一化、gamma变换。
在本申请的一个实施例中，在图6所示装置基础上，还可以包括训练模块，所述训练模块，具体用于：
获取初始神经网络模型和所述样本图像；
将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型；
基于所述初始神经网络模型输出结果和进行所述预设图像增强变换后的无噪图像，计算损失值；
基于所述损失值调整所述初始神经网络的模型参数，并返回将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型的步骤，直到所述初始神经网络模型收敛；
将收敛的初始神经网络模型确定为所述降噪网络模型。
在本申请的一个实施例中，在图6所示装置基础上，还可以包括生成模块，所述生成模块具体用于：
获取暗光噪声图像；
基于预先统计的传感器坏点比例，在所述暗光噪声图像中随机生成特定数量的坏点，得到模拟坏点暗光图像，其中，所生成的坏点的像素值为高亮值或所生成的坏点的像素值与周边像素点的像素值不满足预设分布。
应用本申请实施例提供的暗光图像降噪装置，获取RAW域暗光图像；对RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降 噪网络模型，得到输出图像；其中，降噪网络模型是根据样本图像训练的，样本图像包括：模拟坏点暗光图像和无噪图像；对输出图像进行预设图像增强变换的逆变换，得到降噪图像。
可见，通过模拟坏点的分布情况生成模拟坏点暗光图像，结合无噪图像训练降噪网络模型，在训练完成后，降噪网络模型能够对输入的RAW域图像进行降噪处理，且自动进行坏点抑制，即大幅降低图像中坏点对图像降噪的影响，提高了图像降噪的质量。此外，由于降噪网络模型在图像降噪过程中能够自动进行坏点抑制，因此无需专门进行坏点矫正这一环节，避免在坏点矫正过程中损失图像的边缘细节。
并且，在将RAW域暗光图像输入网络模型之前，先进行图像增强变换，对极大值坏点进行有效的抑制，增强图像暗部细节，更利于在后续的网络模型中进行降噪处理，进一步提高图像降噪质量。
本申请实施例还提供了一种电子设备，如图7所示，包括处理器701、通信接口702、存储器703和通信总线704，其中，处理器701，通信接口702，存储器703通过通信总线704完成相互间的通信，
存储器703，用于存放计算机程序；
处理器701，用于执行存储器703上所存放的程序时，实现如下步骤：
获取RAW域暗光图像；
对所述RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，所述降噪网络模型是根据样本图像训练的，所述样本图像包括：模拟坏点暗光图像和无噪图像；
对所述输出图像进行所述预设图像增强变换的逆变换，得到降噪图像。
上述电子设备提到的通信总线可以是外设部件互连标准(Peripheral Component Interconnect，PCI)总线或扩展工业标准结构(Extended Industry Standard Architecture，EISA)总线等。该通信总线可以分为地址总线、数据总线、控制总线等。为便于表示，图中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
通信接口用于上述电子设备与其他设备之间的通信。
存储器可以包括随机存取存储器(Random Access Memory，RAM)，也可 以包括非易失性存储器(Non-Volatile Memory，NVM)，例如至少一个磁盘存储器。可选的，存储器还可以是至少一个位于远离前述处理器的存储装置。
上述的处理器可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、网络处理器(Network Processor，NP)等；还可以是数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。
应用本申请实施例提供的电子设备，获取RAW域暗光图像；对RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，降噪网络模型是根据样本图像训练的，样本图像包括：模拟坏点暗光图像和无噪图像；对输出图像进行预设图像增强变换的逆变换，得到降噪图像。
可见，通过模拟坏点的分布情况生成模拟坏点暗光图像，结合无噪图像训练降噪网络模型，在训练完成后，降噪网络模型能够对输入的RAW域图像进行降噪处理，且自动进行坏点抑制，即大幅降低图像中坏点对图像降噪的影响，提高了图像降噪的质量。此外，由于降噪网络模型在图像降噪过程中能够自动进行坏点抑制，因此无需专门进行坏点矫正这一环节，避免在坏点矫正过程中损失图像的边缘细节。
并且，在将RAW域暗光图像输入网络模型之前，先进行图像增强变换，对极大值坏点进行有效的抑制，增强图像暗部细节，更利于在后续的网络模型中进行降噪处理，进一步提高图像降噪质量。
在本申请提供的又一实施例中，还提供了一种计算机可读存储介质，该计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现上述任一暗光图像降噪方法的步骤。
在本申请提供的又一实施例中，还提供了一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行上述实施例中任一暗光图像降噪方法。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意 组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘Solid State Disk(SSD))等。
需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
本说明书中的各个实施例均采用相关的方式描述，各个实施例之间相同相似的部分互相参见即可，每个实施例重点说明的都是与其他实施例的不同之处。尤其，对于暗光图像降噪装置、电子设备、计算机可读存储介质、计算机程序产品实施例而言，由于其基本相似于暗光图像降噪方法实施例，所以描述的比较简单，相关之处参见暗光图像降噪方法实施例的部分说明即可。
以上所述仅为本申请的较佳实施例，并非用于限定本申请的保护范围。凡在本申请的精神和原则之内所作的任何修改、等同替换、改进等，均包含 在本申请的保护范围内。

Claims (10)

  1. 一种暗光图像降噪方法，其特征在于，所述方法包括：
    获取RAW域暗光图像；
    对所述RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，所述降噪网络模型是根据样本图像训练的，所述样本图像包括：模拟坏点暗光图像和无噪图像；
    对所述输出图像进行所述预设图像增强变换的逆变换，得到降噪图像。
  2. 根据权利要求1所述的方法，其特征在于，所述对所述RAW域暗光图像进行预设图像增强变换的步骤，包括：
    依次对所述RAW域暗光图像进行归一化、gamma变换。
  3. 根据权利要求1所述的方法，其特征在于，采用如下步骤训练所述降噪网络模型：
    获取初始神经网络模型和所述样本图像；
    将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型；
    基于所述初始神经网络模型输出结果和进行所述预设图像增强变换后的无噪图像，计算损失值；
    基于所述损失值调整所述初始神经网络的模型参数，并返回将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型的步骤，直到所述初始神经网络模型收敛；
    将收敛的初始神经网络模型确定为所述降噪网络模型。
  4. 根据权利要求1所述的方法，其特征在于，采用如下步骤获得所述模拟坏点暗光图像：
    获取暗光噪声图像；
    基于预先统计的传感器坏点比例，在所述暗光噪声图像中随机生成特定数量的坏点，得到所述模拟坏点暗光图像，其中，所生成的坏点的像素值为高亮值或所生成的坏点的像素值与周边像素点的像素值不满足预设分布。
  5. 一种暗光图像降噪装置，其特征在于，所述装置包括：
    获取模块，用于获取RAW域暗光图像；
    降噪模块，用于对所述RAW域暗光图像进行预设图像增强变换，将变换后的图像输入预先训练的降噪网络模型，得到输出图像；其中，所述降噪网络模型是根据样本图像训练的，所述样本图像包括：模拟坏点暗光图像和无噪图像；
    逆变换模块，用于对所述输出图像进行所述预设图像增强变换的逆变换，得到降噪图像。
  6. 根据权利要求5所述的装置，其特征在于，所述降噪模块包含增强变换子模块，所述增强变换子模块，具体用于：
    依次对所述RAW域暗光图像进行归一化、gamma变换。
  7. 根据权利要求5所述的装置，其特征在于，还包括训练模块，所述训练模块，具体用于：
    获取初始神经网络模型和所述样本图像；
    将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型；
    基于所述初始神经网络模型输出结果和进行所述预设图像增强变换后的无噪图像，计算损失值；
    基于所述损失值调整所述初始神经网络的模型参数，并返回将进行所述预设图像增强变换后的模拟坏点暗光图像输入所述初始神经网络模型的步骤，直到所述初始神经网络模型收敛；
    将收敛的初始神经网络模型确定为所述降噪网络模型。
  8. 根据权利要求5所述的装置，其特征在于，还包括：生成模块，所述生成模块具体用于：
    获取暗光噪声图像；
    基于预先统计的传感器坏点比例，在所述暗光噪声图像中随机生成特定数量的坏点，得到所述模拟坏点暗光图像，其中，所生成的坏点的像素值为高亮值或所生成的坏点的像素值与周边像素点的像素值不满足预设分布。
  9. 一种电子设备，其特征在于，包括处理器、通信接口、存储器和通信总线，其中，处理器，通信接口，存储器通过通信总线完成相互间的通信；
    存储器，用于存放计算机程序；
    处理器，用于执行存储器上所存放的程序时，实现权利要求1-4任一所述的方法步骤。
  10. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现权利要求1-4任一所述的方法步骤。
PCT/CN2022/142230 2021-12-29 2022-12-27 一种暗光图像降噪方法及装置 WO2023125503A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111641889.8 2021-12-29
CN202111641889.8A CN114418873B (zh) 2021-12-29 2021-12-29 一种暗光图像降噪方法及装置

Publications (1)

Publication Number Publication Date
WO2023125503A1 true WO2023125503A1 (zh) 2023-07-06

Family

ID=81268846

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/142230 WO2023125503A1 (zh) 2021-12-29 2022-12-27 一种暗光图像降噪方法及装置

Country Status (2)

Country Link
CN (1) CN114418873B (zh)
WO (1) WO2023125503A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114418873B (zh) * 2021-12-29 2022-12-20 英特灵达信息技术(深圳)有限公司 一种暗光图像降噪方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111064904A (zh) * 2019-12-26 2020-04-24 深圳深知未来智能有限公司 一种暗光图像增强方法
CN111402153A (zh) * 2020-03-10 2020-07-10 上海富瀚微电子股份有限公司 一种图像处理方法及系统
CN113850741A (zh) * 2021-10-10 2021-12-28 杭州知存智能科技有限公司 图像降噪方法、装置、电子设备以及存储介质
CN114418873A (zh) * 2021-12-29 2022-04-29 英特灵达信息技术(深圳)有限公司 一种暗光图像降噪方法及装置

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9934557B2 (en) * 2016-03-22 2018-04-03 Samsung Electronics Co., Ltd Method and apparatus of image representation and processing for dynamic vision sensor
CN112529775A (zh) * 2019-09-18 2021-03-19 华为技术有限公司 一种图像处理的方法和装置
CN110766621B (zh) * 2019-10-09 2022-03-25 Oppo广东移动通信有限公司 图像处理方法、装置、存储介质及电子设备
CN110992272B (zh) * 2019-10-18 2023-03-14 深圳大学 基于深度学习的暗光图像增强方法、装置、设备及介质
CN113052768B (zh) * 2019-12-27 2024-03-19 武汉Tcl集团工业研究院有限公司 一种处理图像的方法、终端及计算机可读存储介质
CN113052814B (zh) * 2021-03-23 2024-05-10 浙江工业大学 基于Retinex和注意力机制的暗光图像增强方法
CN113822812A (zh) * 2021-09-15 2021-12-21 维沃移动通信有限公司 图像降噪方法和电子设备

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111064904A (zh) * 2019-12-26 2020-04-24 深圳深知未来智能有限公司 一种暗光图像增强方法
CN111402153A (zh) * 2020-03-10 2020-07-10 上海富瀚微电子股份有限公司 一种图像处理方法及系统
CN113850741A (zh) * 2021-10-10 2021-12-28 杭州知存智能科技有限公司 图像降噪方法、装置、电子设备以及存储介质
CN114418873A (zh) * 2021-12-29 2022-04-29 英特灵达信息技术(深圳)有限公司 一种暗光图像降噪方法及装置

Also Published As

Publication number Publication date
CN114418873B (zh) 2022-12-20
CN114418873A (zh) 2022-04-29

Similar Documents

Publication Publication Date Title
US11836898B2 (en) Method and apparatus for generating image, and electronic device
WO2023125503A1 (zh) 一种暗光图像降噪方法及装置
CN111340735B (zh) 一种led屏体校正方法、装置及终端
WO2023125440A1 (zh) 一种降噪方法、装置、电子设备及介质
Hu et al. Source camera identification using large components of sensor pattern noise
CN108229583B (zh) 一种基于主方向差分特征的快速模板匹配的方法及装置
CN115496668A (zh) 图像处理方法、装置、电子设备及存储介质
Wu et al. Reflectance-guided histogram equalization and comparametric approximation
Tan et al. A real-time video denoising algorithm with FPGA implementation for Poisson–Gaussian noise
WO2021013139A1 (zh) 图像处理的方法和装置
CN110689496B (zh) 降噪模型的确定方法、装置、电子设备和计算机存储介质
CN112634166A (zh) 一种图像处理方法、装置、电子设备及存储介质
Fry et al. Validation of modulation transfer functions and noise power spectra from natural scenes
US20230245276A1 (en) Method and apparatus for acquiring raw image, and electronic device
WO2023284236A1 (zh) 图像盲去噪方法、装置、电子设备和存储介质
Sazonova et al. Fast and efficient iris image enhancement using logarithmic image processing
CN115330612A (zh) 基于自适应中值滤波的辐照图像去噪方法和系统
CN114494080A (zh) 一种图像生成方法、装置、电子设备及存储介质
CN111737519B (zh) 识别机器人账号的方法、装置、电子设备及计算机可读存储介质
CN114119377A (zh) 一种图像处理方法及装置
US20240169497A1 (en) Airy-Disk Correction for Deblurring an Image
CN115953329B (zh) 基于视觉的图像处理方法、系统、电子设备及存储介质
CN113066023B (zh) 一种基于自校准卷积神经网络的sar图像去斑方法
TWI689890B (zh) 雜訊等化方法與雜訊去除方法
CN117710221A (zh) 一种图像处理方法和装置

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: 22914757

Country of ref document: EP

Kind code of ref document: A1