WO2021022685A1 - 一种神经网络训练方法、装置及终端设备 - Google Patents

一种神经网络训练方法、装置及终端设备 Download PDF

Info

Publication number
WO2021022685A1
WO2021022685A1 PCT/CN2019/114946 CN2019114946W WO2021022685A1 WO 2021022685 A1 WO2021022685 A1 WO 2021022685A1 CN 2019114946 W CN2019114946 W CN 2019114946W WO 2021022685 A1 WO2021022685 A1 WO 2021022685A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
image
noise
generate
update
Prior art date
Application number
PCT/CN2019/114946
Other languages
English (en)
French (fr)
Inventor
孙振鉷
Original Assignee
合肥图鸭信息科技有限公司
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 合肥图鸭信息科技有限公司 filed Critical 合肥图鸭信息科技有限公司
Publication of WO2021022685A1 publication Critical patent/WO2021022685A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/44Decoders specially adapted therefor, e.g. video decoders which are asymmetric with respect to the encoder
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/90Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using coding techniques not provided for in groups H04N19/10-H04N19/85, e.g. fractals
    • H04N19/91Entropy coding, e.g. variable length coding [VLC] or arithmetic coding

Definitions

  • This application relates to the field of image deep learning, in particular to a neural network training method, device and terminal equipment.
  • One of the objectives of the embodiments of the present application is to provide a neural network training method, device, and terminal equipment, aiming to solve the problem of slow decoding in the deep learning image compression algorithm.
  • a neural network training method including:
  • Step A Generate image noise
  • Step B Input the image noise into the neural network to generate the corresponding noise generated image
  • Step C Adjust the weight parameters of the neural network according to the noise generated image and the original image, and update the neural network in step B according to the adjusted weight parameters;
  • Step D Repeat steps B to C until the neural network meets the preset condition.
  • said inputting image noise into a neural network to generate a corresponding noise generated image includes:
  • the image noise is convolved in a neural network to generate a corresponding noise generated image.
  • the adjusting the weight parameters of the neural network according to the noise generated image and the original image includes:
  • the weight parameter of the neural network is adjusted through the gradient update.
  • the repeating step B to step C until the neural network meets a preset condition includes:
  • step B Repeat step B until the number of steps C reaches the preset number.
  • step D it further includes:
  • the image noise is input into the neural network after the reconstruction weight parameter is updated to generate a reconstructed image.
  • a neural network training device including:
  • Image noise generation module used to generate image noise
  • Convolution module used to input image noise into the neural network to generate the corresponding noise generated image
  • a neural network update module configured to adjust the weight parameter of the neural network according to the noise generated image and the original image, and update the neural network in step B according to the adjusted weight parameter;
  • the loop module is used to repeat step B to step C until the neural network meets the preset condition.
  • the convolution module includes:
  • the convolution operation unit is configured to perform a convolution operation on the image noise in a neural network to generate a corresponding noise generated image.
  • the neural network update module includes:
  • a loss function unit for generating a loss function based on the noise generated image and the original image
  • a gradient update unit for performing gradient update according to the loss function
  • the parameter adjustment unit is configured to adjust the weight parameter of the neural network through the gradient update.
  • a neural network training terminal device including a memory, a processor, and a computer program stored in the above-mentioned memory and capable of running on the above-mentioned processor.
  • the above-mentioned processor executes the above-mentioned computer program, the above-mentioned computer program is implemented as described in the first aspect. Steps of the provided method.
  • the beneficial effect of the neural network training method provided by the embodiment of this application is that: this embodiment adjusts the weight parameters of the neural network by generating noise until the image compression effect of the neural network reaches the expected index, according to the neural network after adjusting the weight parameters To perform image compression, the image compression effect is improved, and the problem of slow decoding in the image compression algorithm is solved.
  • the beneficial effect of the neural network training device is that: in this embodiment, the image noise generation module 41 generates noise, the convolution module 42 generates the noise to generate an image, and the neural network update module 43 uses the noise to generate the image to adjust The weight parameters of the neural network until the image compression effect of the neural network reaches the expected index, the image compression is performed according to the neural network after adjusting the weight parameters, which improves the image compression effect and solves the problem of slow decoding in the image compression algorithm.
  • the beneficial effect of the neural network training terminal device is that: this embodiment generates random noise, and calculates the random noise through the noise generated image generated by the neural network and the original image to update the weight parameters in the neural network. , And repeatedly update the weight parameters in the neural network to make the neural network reach the most ideal state.
  • Fig. 1 is a schematic diagram of an implementation process of a neural network training method provided by an embodiment of the present application
  • Fig. 2 is a schematic diagram of a convolutional neural network structure provided by another embodiment of the present application.
  • FIG. 3 is a schematic diagram of a process of adjusting weight parameters according to another embodiment of the present application.
  • FIG. 4 is a schematic diagram of a neural network training device provided by another embodiment of the present application.
  • Fig. 5 is a schematic diagram of a neural network training terminal device provided by another embodiment of the present application.
  • FIG. 1 shows the implementation process of the neural network training method provided in the first embodiment of the present invention.
  • the execution subject of the method may be a terminal device, including but not limited to a smart phone, a tablet computer, a personal computer, a server, and the like. It should be noted that the number of terminal devices is not fixed and can be deployed according to actual conditions. Further, the implementation process of the neural network training method provided in the first embodiment above is described in detail as follows:
  • Step S101 generating image noise.
  • the aforementioned image noise may be randomly generated image noise, or preset image noise, which is not limited here.
  • image noise refers to unnecessary or redundant interference information existing in image data, including but not limited to Gaussian noise, Poisson noise, multiplicative noise, salt and pepper noise, etc.
  • a Gaussian noise whose probability density function obeys a Gaussian distribution is randomly generated.
  • the above-mentioned Gaussian noise is a matrix with a size of 1*H*W*C, where H is the height of the noise matrix, and W is the noise matrix. Wide, C is the number of channels of the noise matrix.
  • Step S102 input image noise into the neural network to generate a corresponding noise generated image.
  • the foregoing neural network is a convolutional neural network
  • the convolutional neural network may include at least one convolutional layer.
  • the above-mentioned convolutional layer may include a convolution kernel, and the image input to the convolutional layer undergoes a convolution operation with the convolution kernel to remove redundant image information, and output an image containing feature information. If the size of the above convolution kernel is greater than 1 ⁇ 1, the convolutional layer can output multiple feature maps whose size is smaller than the input image. After multiple convolutional layers are processed, the size of the image input to the convolutional neural network is Multi-level shrinking, to obtain multiple feature maps whose size is smaller than the image size of the input neural network.
  • inputting image noise into the neural network to generate the corresponding noise image may be a deconvolution operation, which is opposite to the above-described process of removing redundant information from the input image to generate a feature image .
  • the aforementioned convolutional neural network may also include a pooling layer, an Inception module, a fully connected layer, etc., which are not limited here.
  • the convolutional neural network may include four convolutional layers, and the dimensions of the weight parameter matrix corresponding to the four convolutional layers are H 1 *W 1 *N 1 *C 1 , H 2 * W 2 *N 2 *C 2 , H 3 *W 3 *N 3 *C 3 , H 4 *W 4 *N 4 *C 4 , the steps are S 1 , S 2 , S 3 , S 4, respectively .
  • H is the height of the weight parameter matrix
  • W is the width of the weight parameter matrix
  • N is the number of output channels of the weight parameter matrix
  • C is the number of input channels of the weight parameter matrix.
  • the number of input channels of the weight parameters of the four convolutional layers is respectively associated with the number of output channels of the weight parameters of the previous layer.
  • C 1 C
  • C 2 N 1
  • C 3 N 2
  • C 4 N 3
  • the randomly generated Gaussian noise is a matrix of 1*H*W*C.
  • Step S103 Adjust the weight parameters of the neural network according to the noise generated image, and update the neural network in step B according to the adjusted weight parameters.
  • FIG. 3 shows the process of adjusting the weight parameters of the aforementioned neural network according to the noise generated image:
  • the loss function between the noise generated image and the original image may use MSE (Mean Square Error).
  • MSE Mel Square Error
  • formula (1) the formula of MSE is shown in formula (1):
  • H is the height of the noise-generated image
  • W is the width of the noise-generated image
  • C is the number of channels of the noise-generated image
  • X' represents the noise-generated image
  • X represents the original image
  • X'i, j, m represents the noise-generated image
  • the values in the i-th row and j-th column of the m-th channel, Xi ,j,m represent the values in the i-th row and j-th column of the m-th channel in the original image.
  • Step S302 Perform gradient update according to the aforementioned loss function.
  • the gradient update formula is shown in formula (2):
  • W represents the weight parameter of the neural network
  • W' represents the weight parameter after update
  • is the preset learning rate
  • ⁇ W is the calculated gradient.
  • an existing adaptive gradient optimizer can be used for calculation.
  • the Adam optimizer can be used. Further, input the above-mentioned MSE calculation result, the weight parameter of the neural network, and the preset learning rate into the Adam optimizer to obtain the updated weight parameter.
  • Step S303 Adjust the weight parameters of the neural network through the above gradient update.
  • Step S104 Repeat steps B to C until the aforementioned neural network meets the preset conditions.
  • repeating step B to step C until the neural network meets a preset condition includes:
  • step B Repeat step B until the number of steps C reaches the preset number.
  • step B is repeatedly executed until the number of times of step C reaches a preset number, where the preset number is manually preset in the neural network training program or preset in the terminal device loaded with the neural network training program.
  • step B to step C are repeatedly executed until the performance index of the image generated by the neural network reaches the preset threshold.
  • the performance indicators of the image generated by the neural network include the peak signal to noise ratio (PSNR) (Peak Signal to Noise Ratio) and the pixel bit BPP (bits per pixel).
  • PSNR peak signal to noise ratio
  • pixel bit BPP bits per pixel
  • the test atlas is put into the neural network after updating the weight parameters to test the performance indicators of the neural network, that is, the peak signal-to-noise ratio PSNR and the pixel bit BPP.
  • PSNR peak signal-to-noise ratio
  • the aforementioned test atlas may include 24 Kodak standard test atlases, which are not limited here.
  • noise is generated to adjust the weight parameters of the neural network until the image compression effect of the neural network reaches the expected index, and the image compression is performed according to the neural network after adjusting the weight parameters, which improves the image compression effect and solves the problem of image compression.
  • the problem of slow decoding in the algorithm is generated to adjust the weight parameters of the neural network until the image compression effect of the neural network reaches the expected index, and the image compression is performed according to the neural network after adjusting the weight parameters, which improves the image compression effect and solves the problem of image compression.
  • FIG. 4 shows a schematic diagram of a neural network training device provided by an embodiment of the present invention. For ease of description, only parts related to the embodiment of the present invention are shown.
  • the neural network training device 4 includes: an image noise generation module 41, a convolution module 42, a neural network update module 43, and a loop module 44.
  • the image noise generation module 41 is used to generate image noise
  • the convolution module 42 is used to input image noise to the neural network to generate a corresponding noise generated image
  • the neural network update module 43 is configured to adjust the weight parameter of the neural network according to the noise generated image and the original image, and update the neural network in step B according to the adjusted weight parameter;
  • the loop module 44 is used to repeat step B to step C until the above neural network meets the preset condition.
  • step B repeating step B to step C until the neural network meets the preset condition includes:
  • step B Repeat step B until the number of steps C reaches the preset number.
  • the aforementioned convolution module 42 includes:
  • the convolution operation unit is configured to perform convolution operation on the above image noise in a neural network to generate a corresponding noise generated image.
  • the aforementioned neural network update module 43 includes:
  • the loss function unit is used to generate a loss function based on the noise generated image and the original image
  • the gradient update unit is used to update the gradient according to the above loss function
  • the parameter adjustment unit is configured to adjust the weight parameters of the neural network through the gradient update.
  • the aforementioned neural network training device 4 further includes:
  • the image compression module is used to extract the weight parameters of the neural network, and use the weight parameters as the feature image;
  • the above-mentioned image noise is input into the neural network after the above-mentioned reconstruction weight parameter is updated to generate a reconstructed image.
  • the image noise generation module 41 generates noise
  • the convolution module 42 generates noise to generate an image.
  • the neural network update module 43 uses the noise to generate the image to adjust the weight parameters of the neural network until the image compression effect of the neural network reaches the expected index. So far, image compression is performed based on the neural network that has adjusted the weight parameters, which improves the image compression effect and solves the problem of slow decoding in the image compression algorithm.
  • Fig. 5 is a schematic diagram of a neural network training terminal device provided by an embodiment of the present invention.
  • the neural network training terminal device 5 of this embodiment includes: a processor 50, a memory 51, and a computer program 52 stored in the memory 51 and running on the processor 50, such as a neural network Training program.
  • the processor 50 executes the computer program 52, the steps in the foregoing neural network training method embodiments, such as steps 101 to 104 shown in FIG. 1, are implemented.
  • the processor 50 executes the computer program 52, the functions of the modules in the foregoing device embodiments, such as the functions of the modules 41 to 44 shown in FIG. 4, are realized.
  • the computer program 52 may be divided into one or more modules/units, and the one or more modules/units are stored in the memory 51 and executed by the processor 50 to complete this invention.
  • the one or more modules/units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program 52 in the neural network training terminal device 5.
  • the computer program 52 can be divided into an image noise generation module, a convolution module, a neural network update module, and a loop module.
  • the specific functions of each module are as follows:
  • Image noise generation module used to generate image noise
  • Convolution module used to input image noise into the neural network to generate the corresponding noise generated image
  • Neural network update module used to adjust the weight parameters of the neural network according to the noise generated image and the original image, and update the neural network in step B according to the adjusted weight parameters;
  • Loop module used to repeat step B to step C until the neural network meets the preset condition.
  • the network-trained terminal device 5 can be a computing device such as a desktop computer, a notebook, a palmtop computer, and a cloud server.
  • the network-trained terminal device may include, but is not limited to, a processor 50 and a memory 51.
  • FIG. 5 is only an example of the network-trained terminal device 5, and does not constitute a limitation on the network-trained terminal device 5. It may include more or less components than shown in the figure, or a combination of some Components, or different components, for example, the network-trained terminal device may also include input and output devices, network access devices, buses, and so on.
  • the so-called processor 50 may be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the memory 51 may be an internal storage unit of the network-trained terminal device 5, such as a hard disk or memory of the network-trained terminal device 5.
  • the memory 51 may also be an external storage device of the network-trained terminal device 5, for example, a plug-in hard disk equipped on the network-trained terminal device 5, a smart media card (SMC), and a secure digital (Secure Digital, SD) card, Flash Card, etc.
  • the memory 51 may also include both an internal storage unit of the network-trained terminal device 5 and an external storage device.
  • the memory 51 is used to store the computer program and other programs and data required by the network-trained terminal device.
  • the memory 51 can also be used to temporarily store data that has been output or will be output.
  • the present invention updates the weight parameters in the neural network by generating random noise, calculating the random noise through the noise generated image generated by the neural network and the original image, and repeatedly updating the weight parameters in the neural network to make the neural network Achieve the most ideal state.
  • the disclosed device/terminal device and method may be implemented in other ways.
  • the device/terminal device embodiments described above are only illustrative.
  • the division of the modules or units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units.
  • components can be combined or integrated into another system, or some features can be omitted or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • the functional units in the various embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.
  • the integrated module/unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the present invention implements all or part of the processes in the above-mentioned embodiments and methods, and can also be completed by instructing relevant hardware through a computer program.
  • the computer program can be stored in a computer-readable storage medium. When the program is executed by the processor, the steps of the foregoing method embodiments can be implemented.
  • the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file, or some intermediate forms.
  • the computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory) , Random Access Memory (RAM, Random Access Memory), electrical carrier signal, telecommunications signal, and software distribution media, etc.
  • ROM Read-Only Memory
  • RAM Random Access Memory
  • electrical carrier signal telecommunications signal
  • software distribution media etc.
  • the content contained in the computer-readable medium can be appropriately added or deleted in accordance with the requirements of the legislation and patent practice in the jurisdiction.
  • the computer-readable medium Does not include electrical carrier signals and telecommunication signals.

Landscapes

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

Abstract

一种神经网络训练方法、装置及终端设备,该神经网络训练方法包括步骤A:生成图像噪声;步骤B:将图像噪声输入到神经网络生成对应的噪声生成图像;步骤C:根据所述噪声生成图像和原始图像调整所述神经网络的权重参数,并根据所述调整后的权重参数更新步骤B所述的神经网络;步骤D:重复执行步骤B到步骤C直至所述神经网络满足预设条件为止。根据噪声生成图像和原始图像来调整神经网络的权重参数,通过调整完权重参数的神经网络来进行图像压缩,提高了图像压缩效果,解决了图像压缩算法中解码过慢的问题。

Description

[根据细则26改正04.12.2019] 一种神经网络训练方法、装置及终端设备
本申请要求于2019年08月08日在中国专利局提交的、申请号为201910727890.9、发明名称为“一种神经网络训练方法、装置及终端设备”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
本申请涉及图像深度学习领域,具体涉及一种神经网络训练方法、装置及终端设备。
背景技术
这里的陈述仅提供与本申请有关的背景信息,而不必然构成现有技术。
传统的图像压缩算法,例如jpeg、jpeg2000等,在获得极高压缩率的同时会大幅丢失图像中的高频信息,导致图像信息大幅丢失,引起图像失真。目前网站、社交媒体上的高清图片日益增多,随之而来的带宽消耗也不断增大,如不进行压缩的话会占用过多的资源空间,如运用传统的图像压缩算法,则会导致图像经压缩后不清晰等问题。
技术问题
本申请实施例的目的之一在于:提供一种神经网络训练方法、装置及终端设备,旨在解决深度学习图像压缩算法中解码过慢的问题。
技术解决方案
为解决上述技术问题,本申请实施例采用的技术方案是:
第一方面,提供了一种神经网络训练方法,包括:
步骤A:生成图像噪声;
步骤B:将图像噪声输入到神经网络生成对应的噪声生成图像;
步骤C:根据所述噪声生成图像和原始图像调整所述神经网络的权重参数,并根据所述调整后的权重参数更新步骤B所述的神经网络;
步骤D:重复执行步骤B到步骤C直至所述神经网络满足预设条件为止。
在一个实施例中,所述将图像噪声输入到神经网络生成对应的噪声生成图像包括:
将所述图像噪声在神经网络中进行卷积操作生成对应的噪声生成图像。
在一个实施例中,所述根据所述噪声生成图像和原始图像调整所述神经网络的权重参数包括:
根据所述噪声生成图像与原始图像生成损失函数;
根据所述损失函数进行梯度更新;
通过所述梯度更新调整所述神经网络的权重参数。
在一个实施例中,所述重复执行步骤B到步骤C直至所述神经网络满足预设条件为止包括:
重复执行步骤B到步骤C直至所述神经网络生成图像的性能指标达到预设阈值为止
重复执行步骤B到步骤C的次数达到预设次数为止。
在一个实施例中,在步骤D后,还包括:
提取所述神经网络的权重参数,将所述权重参数作为特征图像;
将所述特征图像通过熵编码得到编码数据;
将所述编码数据通过熵解码生成重构权重参数;
根据所述重构权重参数更新所述神经网络;
将所述图像噪声输入所述重构权重参数更新后的神经网络中,生成重构图 像。
第二方面,提供了一种神经网络训练装置,包括:
图像噪声生成模块,用于生成图像噪声;
卷积模块,用于将图像噪声输入到神经网络生成对应的噪声生成图像;
神经网络更新模块,用于根据所述噪声生成图像和原始图像调整所述神经网络的权重参数,并根据所述调整后的权重参数更新步骤B所述的神经网络;
循环模块,用于重复执行步骤B到步骤C直至所述神经网络满足预设条件为止。
在一个实施例中,所述卷积模块包括:
卷积操作单元,用于将所述图像噪声在神经网络中进行卷积操作生成对应的噪声生成图像。
在一个实施例中,所述神经网络更新模块包括:
损失函数单元,用于根据所述噪声生成图像与原始图像生成损失函数;
梯度更新单元,用于根据所述损失函数进行梯度更新;
参数调整单元,用于通过所述梯度更新调整所述神经网络的权重参数。
第三方面,提供一种神经网络训练终端设备,包括存储器、处理器以及存储在上述存储器中并可在上述处理器上运行的计算机程序,上述处理器执行上述计算机程序时实现如上第一方面所提供的方法的步骤。
本申请实施例提供的一种神经网络训练方法的有益效果在于:本实施例通过生成噪声来调整神经网络的权重参数直至神经网络的图像压缩效果达到预期指标为止,根据调整完权重参数的神经网络来进行图像压缩,提高了图像压缩效果,解决了图像压缩算法中解码过慢的问题。
本申请实施例提供的一种神经网络训练装置的有益效果在于:本实施例通 过图像噪声生成模块41生成噪声,根据卷积模块42生成噪声生成图像,神经网络更新模块43通过噪声生成图像来调整神经网络的权重参数直至神经网络的图像压缩效果达到预期指标为止,根据调整完权重参数的神经网络来进行图像压缩,提高了图像压缩效果,解决了图像压缩算法中解码过慢的问题。
本申请实施例提供的一种神经网络训练终端设备的有益效果在于:本实施例通过生成随机噪声,将随机噪声通过神经网络生成的噪声生成图像与原始图像进行计算来更新神经网络中的权重参数,并一直重复更新神经网络中的权重参数来使得神经网络达到最理想的状态。从而提高了神经网络对图像压缩的效果,提高了压缩算法解码的速度。
附图说明
为了更清楚地说明本申请实施例中的技术方案,下面将对实施例或示范性技术描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其它的附图。
图1是本申请一实施例提供的神经网络训练方法的实现流程的示意图;
图2是本申请另一实施例提供的卷积神经网络结构的示意图;
图3是本申请另一实施例提供的调整权重参数流程的示意图;
图4是本申请另一实施例提供的神经网络训练装置的示意图;
图5是本申请另一实施例提供的神经网络训练终端设备的示意图。
本发明的实施方式
为了使本申请的目的、技术方案及优点更加清楚明白,以下结合附图及实 施例,对本申请进行进一步详细说明。应当理解,此处所描述的具体实施例仅用以解释本发明,并不用于限定本申请。
需说明的是,当部件被称为“固定于”或“设置于”另一个部件,它可以直接在另一个部件上或者间接在该另一个部件上。当一个部件被称为是“连接于”另一个部件,它可以是直接或者间接连接至该另一个部件上。术语“上”、“下”、“左”、“右”等指示的方位或位置关系为基于附图所示的方位或位置关系,仅是为了便于描述,而不是指示或暗示所指的装置或元件必须具有特定的方位、以特定的方位构造和操作,因此不能理解为对本申请的限制,对于本领域的普通技术人员而言,可以根据具体情况理解上述术语的具体含义。术语“第一”、“第二”仅用于便于描述目的,而不能理解为指示或暗示相对重要性或者隐含指明技术特征的数量。“多个”的含义是两个或两个以上,除非另有明确具体的限定。
为了说明本申请所述的技术方案,以下结合具体附图及实施例进行详细说明。
实施例一
图1示出了本发明实施例一提供的神经网络训练方法的实现流程,该方法的执行主体可以是终端设备,包括但不限于智能手机、平板电脑、个人电脑、服务器等。需要说明的是,该终端设备的数目并不是固定的,可以根据实际情况进行部署。进一步地,上述实施例一提供的神经网络训练方法的实现流程详述如下:
步骤S101,生成图像噪声。
可选地,上述图像噪声可以是随机生成的图像噪声,也可以是预先设置好 的图像噪声,此处不作限定。其中,图像噪声是指存在于图像数据中的不必要的或多余的干扰信息,包括但不限于高斯噪声、泊松噪声、乘性噪声、椒盐噪声等。
示例地,在本发明实施例中随机生成一个概率密度函数服从高斯分布的高斯噪声,上述高斯噪声为1*H*W*C尺寸的矩阵,其中H为噪声矩阵的高,W为噪声矩阵的宽,C为噪声矩阵的通道数。
步骤S102,将图像噪声输入到神经网络生成对应的噪声生成图像。
可选地,上述神经网络为卷积神经网络,该卷积神经网络可以包括至少一个卷积层。进一步地,上述卷积层可包括卷积核,输入卷积层的图像经过与卷积核的卷积运算后去除冗余的图像信息,输出包含特征信息的图像。如果上述卷积核的尺寸大于1×1,则卷积层可以输出多幅尺寸小于输入图像的特征图,在经过多个卷积层的处理后,输入卷积神经网络的图像的尺寸经过了多级收缩,得到多幅尺寸小于输入神经网络的图像尺寸的特征图。进一步地,在本发明实施例中,将图像噪声输入到神经网络生成对应的噪声图像可以是反卷积操作,反卷积操作则与上述描述的输入图像去除冗余信息生成特征图像的过程相反。可选地,上述卷积神经网络还可以包括池化层、Inception模块、全连接层等,此处不作限定。
示例地,如图2所示,卷积神经网络可以包括四个卷积层,四个卷积层对应的权重参数矩阵的尺寸分别为H 1*W 1*N 1*C 1,H 2*W 2*N 2*C 2,H 3*W 3*N 3*C 3,H 4*W 4*N 4*C 4,步长分别为S 1、S 2、S 3、S 4。其中,H为权重参数矩阵的高,W为权重参数矩阵的宽,N为权重参数矩阵的输出通道数,C为权重参数矩阵的输入通道数。且四个卷积层的权重参数的输入通道数分别与前一层的权重参数的输出通道数相关联。示例地,在图2中,C 1=C,C 2=N 1,C 3=N 2,C 4=N 3。 在步骤S101示例中随机生成的高斯噪声为1*H*W*C的矩阵,经过上述卷积神经网络进行卷积操作后,对应生成尺寸为1*H’*W’*C’的噪声生成图像,其中H’=H*S 1*S 2*S 3*S 4,W’=W*S 1*S 2*S 3*S 4,C’=C 4
步骤S103,根据上述噪声生成图像调整上述神经网络的权重参数,并根据上述调整后的权重参数更新步骤B上述的神经网络。
可选地,图3示出了根据噪声生成图像调整上述神经网络的权重参数的过程:
步骤:S301:根据上述噪声生成图像与原始图像生成损失函数。
可选地,噪声生成图像和原始图像之间的损失函数可以使用MSE(均方误差)。具体的,MSE的公式如公式(1)所示:
Figure PCTCN2019114946-appb-000001
其中,H为噪声生成图像的高,W为噪声生成图像的宽,C为噪声生成图像通道数,X’代表噪声生成图像,X代表原始图像,X’ i,j,m代表噪声生成图像中第m通道第i行第j列的数值,X i,j,m代表原始图像中第m通道第i行第j列的数值。
步骤S302:根据上述损失函数进行梯度更新。
可选地,梯度更新的公式如公式(2)所示:
W′=W-αΔW             (2)
其中,W代表神经网络的权重参数,W’代表更新后的权重参数,α是预 先设定的学习率,ΔW是计算梯度。
可选地,在进行梯度更新的时候,可以使用现有的自适应梯度优化器来进行计算。具体地,可以使用Adam优化器。进一步地,在Adam优化器中输入上述MSE计算结果、神经网络的权重参数、预先设定的学习率,即可得到更新后的权重参数。
步骤S303:通过上述梯度更新调整神经网络的权重参数。
可选地,将上述计算得到的更新后的权重参数替换掉神经网络中原有的权重参数,成为新的神经网络。上述新的神经网络的权重参数即为步骤S302中计算出的更新后的权重参数。
步骤S104,重复执行步骤B到步骤C直至上述神经网络满足预设条件为止。
可选地,上述重复执行步骤B到步骤C直至上述神经网络满足预设条件为止包括:
重复执行步骤B到步骤C直至上述神经网络生成图像的性能指标达到预设阈值为止
重复执行步骤B到步骤C的次数达到预设次数为止。
进一步地,重复执行步骤B到步骤C的次数达到预设次数为止,其中预设次数为人工预先设置在神经网络训练程序中或者预先设置在装载神经网络训练程序的终端设备中。
进一步地,重复执行步骤B到步骤C直至上述神经网络生成图像的性能指标达到预设阈值为止。其中,上述神经网络生成图像的性能指标包括峰值信 噪比PSNR(Peak Signal to Noise Ratio)和像素比特BPP(bits per pixel)。
具体地,将测试图集放入到上述更新完权重参数后的神经网络中测试上述神经网络的性能指标,即峰值信噪比PSNR和像素比特BPP。可选地,在固定的像素比特BPP下,判断峰值信噪比PSNR是否达到预设阈值,峰值信噪比PSNR越高则代表图片压缩中损失的信息越少。可选地,上述测试图集可以包括24张柯达标准测试图集,此处不作限定。
本实施例中,通过生成噪声来调整神经网络的权重参数直至神经网络的图像压缩效果达到预期指标为止,根据调整完权重参数的神经网络来进行图像压缩,提高了图像压缩效果,解决了图像压缩算法中解码过慢的问题。
应理解,上述实施例中各步骤的序号的大小并不意味着执行顺序的先后,各过程的执行顺序应以其功能和内在逻辑确定,而不应对本发明实施例的实施过程构成任何限定。
实施例二
图4示出了本发明实施例提供的神经网络训练装置的示意图,为了便于说明,仅示出了与本发明实施例相关的部分。该神经网络训练装置4包括:图像噪声生成模块41,卷积模块42,神经网络更新模块43,循环模块44。
其中,图像噪声生成模块41,用于生成图像噪声;
卷积模块42,用于将图像噪声输入到神经网络生成对应的噪声生成图像;
神经网络更新模块43,用于根据上述噪声生成图像和原始图像调整上述神经网络的权重参数,并根据上述调整后的权重参数更新步骤B上述的神经网络;
循环模块44,用于重复执行步骤B到步骤C直至上述神经网络满足预设 条件为止。
具体地,上述重复执行步骤B到步骤C直至上述神经网络满足预设条件为止包括:
重复执行步骤B到步骤C直至上述神经网络生成图像的性能指标达到预设阈值为止
重复执行步骤B到步骤C的次数达到预设次数为止。
可选地,上述卷积模块42包括:
卷积操作单元,用于将上述图像噪声在神经网络中进行卷积操作生成对应的噪声生成图像。
可选地,上述神经网络更新模块43包括:
损失函数单元,用于根据上述噪声生成图像与原始图像生成损失函数;
梯度更新单元,用于根据上述损失函数进行梯度更新;
参数调整单元,用于通过上述梯度更新调整上述神经网络的权重参数。
可选地,上述神经网络训练装置4还包括:
图片压缩模块,用于提取上述神经网络的权重参数,将上述权重参数作为特征图像;
将上述特征图像通过熵编码得到编码数据;
将上述编码数据通过熵解码生成重构权重参数;
根据上述重构权重参数更新上述神经网络;
将上述图像噪声输入上述重构权重参数更新后的神经网络中,生成重构图像。
本实施例中,通过图像噪声生成模块41生成噪声,根据卷积模块42生成 噪声生成图像,神经网络更新模块43通过噪声生成图像来调整神经网络的权重参数直至神经网络的图像压缩效果达到预期指标为止,根据调整完权重参数的神经网络来进行图像压缩,提高了图像压缩效果,解决了图像压缩算法中解码过慢的问题。
实施例三
图5是本发明一实施例提供的神经网络训练终端设备的示意图。如图5所示,该实施例的神经网络训练终端设备5包括:处理器50、存储器51以及存储在所述存储器51中并可在所述处理器50上运行的计算机程序52,例如神经网络训练程序。所述处理器50执行所述计算机程序52时实现上述各个神经网络训练方法实施例中的步骤,例如图1所示的步骤101至104。或者,所述处理器50执行所述计算机程序52时实现上述各装置实施例中各模块的功能,例如图4所示模块41至44的功能。
示例性的,所述计算机程序52可以被分割成一个或多个模块/单元,所述一个或者多个模块/单元被存储在所述存储器51中,并由所述处理器50执行,以完成本发明。所述一个或多个模块/单元可以是能够完成特定功能的一系列计算机程序指令段,该指令段用于描述所述计算机程序52在所述神经网络训练终端设备5中的执行过程。例如,所述计算机程序52可以被分割成图像噪声生成模块、卷积模块、神经网络更新模块和循环模块,各模块具体功能如下:
图像噪声生成模块:用于生成图像噪声;
卷积模块:用于将图像噪声输入到神经网络生成对应的噪声生成图像;
神经网络更新模块:用于根据所述噪声生成图像和原始图像调整所述神经网络的权重参数,并根据所述调整后的权重参数更新步骤B所述的神经网络;
循环模块:用于重复执行步骤B到步骤C直至所述神经网络满足预设条件为止。
所述经网络训练终端设备5可以是桌上型计算机、笔记本、掌上电脑及云端服务器等计算设备。所述经网络训练终端设备可包括,但不仅限于,处理器50、存储器51。本领域技术人员可以理解,图5仅仅是经网络训练终端设备5的示例,并不构成对经网络训练终端设备5的限定,可以包括比图示更多或更少的部件,或者组合某些部件,或者不同的部件,例如所述经网络训练终端设备还可以包括输入输出设备、网络接入设备、总线等。
所称处理器50可以是中央处理单元(Central Processing Unit,CPU),还可以是其他通用处理器、数字信号处理器(Digital Signal Processor,DSP)、专用集成电路(Application Specific Integrated Circuit,ASIC)、现成可编程门阵列(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、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器51可以是所述经网络训练终端设备5的内部存储单元,例如经网络训练终端设备5的硬盘或内存。所述存储器51也可以是所述经网络训练终端设备5的外部存储设备,例如所述经网络训练终端设备5上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。进一步地,所述存储器51还可以既包括所述经网络训练终端设备5的内部存储单元也包括外部存储设备。所述存储器51用于存储所述计算机程序以及所述经网络训练终端设备所需的其他程序和数据。所述存储器51还可以用于暂时地存储已经输出或者将要输出的数据。
由上可见,本发明通过生成随机噪声,将随机噪声通过神经网络生成的噪 声生成图像与原始图像进行计算来更新神经网络中的权重参数,并一直重复更新神经网络中的权重参数来使得神经网络达到最理想的状态。从而提高了神经网络对图像压缩的效果,提高了压缩算法解码的速度。
所属领域的技术人员可以清楚地了解到,为了描述的方便和简洁,仅以上述各功能单元、模块的划分进行举例说明,实际应用中,可以根据需要而将上述功能分配由不同的功能单元、模块完成,即将所述装置的内部结构划分成不同的功能单元或模块,以完成以上描述的全部或者部分功能。实施例中的各功能单元、模块可以集成在一个处理单元中,也可以是各个单元单独物理存在,也可以两个或两个以上单元集成在一个单元中,上述集成的单元既可以采用硬件的形式实现,也可以采用软件功能单元的形式实现。另外,各功能单元、模块的具体名称也只是为了便于相互区分,并不用于限制本申请的保护范围。上述系统中单元、模块的具体工作过程,可以参考前述方法实施例中的对应过程,在此不再赘述。
在上述实施例中,对各个实施例的描述都各有侧重,某个实施例中没有详述或记载的部分,可以参见其它实施例的相关描述。
本领域普通技术人员可以意识到,结合本文中所公开的实施例描述的各示例的单元及算法步骤,能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行,取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能,但是这种实现不应认为超出本发明的范围。
在本发明所提供的实施例中,应该理解到,所揭露的装置/终端设备和方法,可以通过其它的方式实现。例如,以上所描述的装置/终端设备实施例仅仅是示意性的,例如,所述模块或单元的划分,仅仅为一种逻辑功能划分,实 际实现时可以有另外的划分方式,例如多个单元或组件可以结合或者可以集成到另一个系统,或一些特征可以忽略,或不执行。另一点,所显示或讨论的相互之间的耦合或直接耦合或通讯连接可以是通过一些接口,装置或单元的间接耦合或通讯连接,可以是电性,机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的,作为单元显示的部件可以是或者也可以不是物理单元,即可以位于一个地方,或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外,在本发明各个实施例中的各功能单元可以集成在一个处理单元中,也可以是各个单元单独物理存在,也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现,也可以采用软件功能单元的形式实现。
所述集成的模块/单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时,可以存储在一个计算机可读取存储介质中。基于这样的理解,本发明实现上述实施例方法中的全部或部分流程,也可以通过计算机程序来指令相关的硬件来完成,所述的计算机程序可存储于一计算机可读存储介质中,该计算机程序在被处理器执行时,可实现上述各个方法实施例的步骤。其中,所述计算机程序包括计算机程序代码,所述计算机程序代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述计算机可读介质可以包括:能够携带所述计算机程序代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器(ROM,Read-Only Memory)、随机存取存储器(RAM,Random Access Memory)、电载波信号、电信信号以及软件分发介质等。需要说明的是,所述计算机可读介质包含的内容可以根据 司法管辖区内立法和专利实践的要求进行适当的增减,例如在某些司法管辖区,根据立法和专利实践,计算机可读介质不包括是电载波信号和电信信号。
以上仅为本申请的可选实施例而已,并不用于限制本申请。对于本领域的技术人员来说,本申请可以有各种更改和变化。凡在本申请的精神和原则之内,所作的任何修改、等同替换、改进等,均应包含在本申请的权利要求范围之内。

Claims (10)

  1. 一种神经网络训练方法,其特征在于,包括:
    步骤A:生成图像噪声;
    步骤B:将图像噪声输入到神经网络生成对应的噪声生成图像;
    步骤C:根据所述噪声生成图像和原始图像调整所述神经网络的权重参数,并根据所述调整后的权重参数更新步骤B所述的神经网络;
    步骤D:重复执行步骤B到步骤C直至所述神经网络满足预设条件为止。
  2. 如权利要求1所述的神经网络训练方法,其特征在于,所述将图像噪声输入到神经网络生成对应的噪声生成图像包括:
    将所述图像噪声在神经网络中进行卷积操作生成对应的噪声生成图像。
  3. 如权利要求1所述的神经网络训练方法,其特征在于,所述根据所述噪声生成图像和原始图像调整所述神经网络的权重参数包括:
    根据所述噪声生成图像与原始图像生成损失函数;
    根据所述损失函数进行梯度更新;
    通过所述梯度更新调整所述神经网络的权重参数。
  4. 如权利要求1所述的神经网络训练方法,其特征在于,所述重复执行步骤B到步骤C直至所述神经网络满足预设条件为止包括:
    重复执行步骤B到步骤C直至所述神经网络生成图像的性能指标达到预设阈值为止
    重复执行步骤B到步骤C的次数达到预设次数为止。
  5. 如权利要求1所述的神经网络训练方法,其特征在于,在步骤D后, 还包括:
    提取所述神经网络的权重参数,将所述权重参数作为特征图像;
    将所述特征图像通过熵编码得到编码数据;
    将所述编码数据通过熵解码生成重构权重参数;
    根据所述重构权重参数初始化所述神经网络;
    将所述图像噪声输入所述重构权重参数更新后的神经网络中,生成重构图像。
  6. 一种神经网络训练装置,其特征在于,包括:
    图像噪声生成模块,用于生成图像噪声;
    卷积模块,用于将图像噪声输入到神经网络生成对应的噪声生成图像;
    神经网络更新模块,用于根据所述噪声生成图像和原始图像调整所述神经网络的权重参数,并根据所述调整后的权重参数更新步骤B所述的神经网络;
    循环模块,用于重复执行步骤B到步骤C直至所述神经网络满足预设条件为止。
  7. 如权利要求6所述的神经网络训练装置,其特征在于,所述卷积模块包括:
    卷积操作单元,用于将所述图像噪声在神经网络中进行卷积操作生成对应的噪声生成图像。
  8. 如权利要求6所述的神经网络训练装置,其特征在于,所述神经网络更新模块包括:
    损失函数单元,用于根据所述噪声生成图像与原始图像生成损失函数;
    梯度更新单元,用于根据所述损失函数进行梯度更新;
    参数调整单元,用于通过所述梯度更新调整所述神经网络的权重参数。
  9. 一种神经网络训练终端设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序,其特征在于,所述处理器执行所述计算机程序时实现如权利要求1至5任一项所述方法的步骤。
  10. 一种计算机可读存储介质,所述计算机可读存储介质存储有计算机程序,其特征在于,所述计算机程序被处理器执行时实现如权利要求1至5任一项所述方法的步骤。
PCT/CN2019/114946 2019-08-08 2019-11-01 一种神经网络训练方法、装置及终端设备 WO2021022685A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910727890.9 2019-08-08
CN201910727890.9A CN110569961A (zh) 2019-08-08 2019-08-08 一种神经网络训练方法、装置及终端设备

Publications (1)

Publication Number Publication Date
WO2021022685A1 true WO2021022685A1 (zh) 2021-02-11

Family

ID=68774786

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/114946 WO2021022685A1 (zh) 2019-08-08 2019-11-01 一种神经网络训练方法、装置及终端设备

Country Status (2)

Country Link
CN (1) CN110569961A (zh)
WO (1) WO2021022685A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113052301A (zh) * 2021-03-29 2021-06-29 商汤集团有限公司 神经网络生成方法、装置、电子设备及存储介质
CN113570056A (zh) * 2021-06-23 2021-10-29 上海交通大学 容错神经网络结构的优化方法、系统、介质及电子设备
CN113657576A (zh) * 2021-07-21 2021-11-16 浙江大华技术股份有限公司 卷积神经网络模型轻量化方法及设备、图像识别方法
US11356305B2 (en) * 2020-02-24 2022-06-07 Qualcomm Incorporated Method to convey the TX waveform distortion to the receiver
CN115761448A (zh) * 2022-12-02 2023-03-07 美的集团(上海)有限公司 神经网络的训练方法、训练装置和可读存储介质
US11711449B2 (en) 2021-12-07 2023-07-25 Capital One Services, Llc Compressing websites for fast data transfers
US11770510B2 (en) 2018-09-21 2023-09-26 Andrew Sviridenko Video information compression using sketch-video

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111556316B (zh) * 2020-04-08 2022-06-03 北京航空航天大学杭州创新研究院 一种基于深度神经网络加速的快速块分割编码方法和装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108062780A (zh) * 2017-12-29 2018-05-22 百度在线网络技术(北京)有限公司 图像压缩方法和装置
CN108090521A (zh) * 2018-01-12 2018-05-29 广州视声智能科技有限公司 一种生成式对抗网络模型的图像融合方法和判别器
CN108985464A (zh) * 2018-07-17 2018-12-11 重庆科技学院 基于信息最大化生成对抗网络的人脸连续特征生成方法
US20190096038A1 (en) * 2017-09-26 2019-03-28 Samsung Electronics Co., Ltd. Image denoising neural network architecture and method of training the same

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018018470A1 (zh) * 2016-07-27 2018-02-01 华为技术有限公司 一种去除图像噪声的方法、装置、设备及卷积神经网络
CN110062246B (zh) * 2018-01-19 2021-01-05 杭州海康威视数字技术股份有限公司 对视频帧数据进行处理的方法和装置
CN109872288B (zh) * 2019-01-31 2023-05-23 深圳大学 用于图像去噪的网络训练方法、装置、终端及存储介质
CN109919864A (zh) * 2019-02-20 2019-06-21 重庆邮电大学 一种基于稀疏去噪自编码网络的图像压缩感知方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190096038A1 (en) * 2017-09-26 2019-03-28 Samsung Electronics Co., Ltd. Image denoising neural network architecture and method of training the same
CN108062780A (zh) * 2017-12-29 2018-05-22 百度在线网络技术(北京)有限公司 图像压缩方法和装置
CN108090521A (zh) * 2018-01-12 2018-05-29 广州视声智能科技有限公司 一种生成式对抗网络模型的图像融合方法和判别器
CN108985464A (zh) * 2018-07-17 2018-12-11 重庆科技学院 基于信息最大化生成对抗网络的人脸连续特征生成方法

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11770510B2 (en) 2018-09-21 2023-09-26 Andrew Sviridenko Video information compression using sketch-video
US11356305B2 (en) * 2020-02-24 2022-06-07 Qualcomm Incorporated Method to convey the TX waveform distortion to the receiver
CN113052301A (zh) * 2021-03-29 2021-06-29 商汤集团有限公司 神经网络生成方法、装置、电子设备及存储介质
CN113052301B (zh) * 2021-03-29 2024-05-28 商汤集团有限公司 神经网络生成方法、装置、电子设备及存储介质
CN113570056A (zh) * 2021-06-23 2021-10-29 上海交通大学 容错神经网络结构的优化方法、系统、介质及电子设备
CN113657576A (zh) * 2021-07-21 2021-11-16 浙江大华技术股份有限公司 卷积神经网络模型轻量化方法及设备、图像识别方法
US11711449B2 (en) 2021-12-07 2023-07-25 Capital One Services, Llc Compressing websites for fast data transfers
CN115761448A (zh) * 2022-12-02 2023-03-07 美的集团(上海)有限公司 神经网络的训练方法、训练装置和可读存储介质
CN115761448B (zh) * 2022-12-02 2024-03-01 美的集团(上海)有限公司 神经网络的训练方法、训练装置和可读存储介质

Also Published As

Publication number Publication date
CN110569961A (zh) 2019-12-13

Similar Documents

Publication Publication Date Title
WO2021022685A1 (zh) 一种神经网络训练方法、装置及终端设备
WO2021098362A1 (zh) 视频分类模型构建、视频分类的方法、装置、设备及介质
CN109816615B (zh) 图像修复方法、装置、设备以及存储介质
CN109087273B (zh) 基于增强的神经网络的图像复原方法、存储介质及系统
CN109671026B (zh) 基于空洞卷积及自动编解码神经网络的灰度图像降噪方法
WO2021164269A1 (zh) 基于注意力机制的视差图获取方法和装置
WO2021035629A1 (zh) 获取画质增强网络的方法、图像画质增强方法、装置、可移动平台、相机及存储介质
CN111784699B (zh) 一种对三维点云数据进行目标分割方法、装置及终端设备
WO2020186888A1 (zh) 一种图像处理模型的构建方法、装置及终端设备
US11176404B2 (en) Method and apparatus for detecting object in image, and storage medium thereof
WO2021022686A1 (zh) 一种视频压缩方法、装置及终端设备
CN111383232A (zh) 抠图方法、装置、终端设备及计算机可读存储介质
WO2020165848A1 (en) Quality assessment of an image
CN111210004B (zh) 卷积计算方法、卷积计算装置及终端设备
CN115643001A (zh) 一种基于位平面的图像加密方法、系统及可读存储介质
WO2016101663A1 (zh) 图片压缩方法及装置
WO2021218140A1 (zh) 基于可变形卷积的图像识别方法、装置、计算机设备
CN113744159A (zh) 一种遥感图像去雾方法、装置及电子设备
WO2021128820A1 (zh) 数据处理方法、装置、设备、存储介质和程序产品
CN111083482A (zh) 一种视频压缩网络训练方法、装置及终端设备
CN110677671A (zh) 一种图像压缩方法、装置及终端设备
CN116739938A (zh) 色调映射方法、装置和显示设备
CN109308690B (zh) 一种图像亮度均衡方法及终端
CN113362229B (zh) 图像处理模型的训练方法、图像处理方法、装置及设备
WO2021139179A1 (zh) 基于局部阴影特效的图像样本生成方法及装置

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: 19940202

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: 19940202

Country of ref document: EP

Kind code of ref document: A1