CN110458778A - A kind of depth image denoising method, device and storage medium - Google Patents
A kind of depth image denoising method, device and storage medium Download PDFInfo
- Publication number
- CN110458778A CN110458778A CN201910728854.4A CN201910728854A CN110458778A CN 110458778 A CN110458778 A CN 110458778A CN 201910728854 A CN201910728854 A CN 201910728854A CN 110458778 A CN110458778 A CN 110458778A
- Authority
- CN
- China
- Prior art keywords
- deep learning
- depth
- denoising
- depth image
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000003860 storage Methods 0.000 title claims abstract description 20
- 238000013135 deep learning Methods 0.000 claims abstract description 119
- 238000012549 training Methods 0.000 claims abstract description 91
- 230000006870 function Effects 0.000 claims abstract description 70
- 238000005259 measurement Methods 0.000 claims abstract description 31
- 230000004044 response Effects 0.000 claims abstract description 29
- 238000004088 simulation Methods 0.000 claims abstract description 16
- 230000015654 memory Effects 0.000 claims description 25
- 238000013527 convolutional neural network Methods 0.000 claims description 8
- 238000003384 imaging method Methods 0.000 claims description 6
- 230000004927 fusion Effects 0.000 claims description 5
- 238000004891 communication Methods 0.000 claims description 3
- 238000004590 computer program Methods 0.000 description 8
- 238000012545 processing Methods 0.000 description 6
- 230000008901 benefit Effects 0.000 description 5
- 230000003287 optical effect Effects 0.000 description 5
- 238000009826 distribution Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 238000007499 fusion processing Methods 0.000 description 2
- 238000002310 reflectometry Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 230000001360 synchronised effect Effects 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 210000001367 artery Anatomy 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000000151 deposition Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000004080 punching Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000009131 signaling function Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000005309 stochastic process Methods 0.000 description 1
- 238000002366 time-of-flight method Methods 0.000 description 1
- 210000003462 vein Anatomy 0.000 description 1
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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
-
- 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]
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of depth image denoising method, device and storage medium, method includes: that simulation obtains depth transducer to the original response data of different 3D scenes as training sample set;The deep learning network for denoising to depth image and the loss function for training the deep learning network are constructed, wherein loss function includes space loss item and the loss of time;Training sample set is input to deep learning network, after carrying out deep learning training to deep learning network by loss function, and model verifying is carried out by the assessment sample set of actual measurement and obtains denoising acoustic model;By depth transducer actual acquisition to depth image data be input to the denoising acoustic model, the depth image after generating denoising.The embodiment of the present invention is trained deep learning network by the loss function with depth measurement signal characteristic and verifies junior scholar's acquistion and denoises to denoising model realization, eliminates interference of the environment light to signal light, improves the signal-to-noise ratio of depth image.
Description
Technical field
The present invention relates to image sensing technical fields more particularly to a kind of depth image denoising method, device and storage to be situated between
Matter.
Background technique
Optical triangle method refers to the method for optically obtaining scene three-dimensional information.Optical triangle method technology exists
The fields such as industrial detection, recognition of face, smart home, machine vision and virtual reality are of great importance with wide before
Scape.According to whether active light source, optical triangle method method can be divided into passive three-dimensional sensing and active three-dimensional sensing two
Major class, currently used three-dimensional sensing method have: flight time (Time of Flight, ToF), structure light scheme, binocular vision
Feel and light field record etc..Wherein, pulsed-beam time-of-flight methods are received in recent years due to advantages such as its cost, precision, detection ranges
Industry is more and more paid close attention to.
Depth sensing module based on flight time (dToF) scheme direct in the flight time has been widely used for laser
The equipment such as radar and mobile phone.As known to the flying speed of optical signal and constant in dToF scheme, system can pass through signal
The time interval being emitted between receiving accurately calculates the distance of testee.One typical dTOF sensor-based system needs makes
With high performance photodetector to capture optical signal, such as single-photon avalanche diode (single photon avalanche
Diode, SPAD), because its have the advantages that system power dissipation it is low, needed for laser energy it is low, be a kind of great potential and practical value
Scheme.But in the weaker situation of signal light, photon numbers are extremely limited, signal is easy to be flooded by noise, and SPAD is passed
Sensor is easy the interference by environment light, is unable to judge accurately the position where signal light due to the highly sensitive characteristic of itself,
So as to cause the measurement result inaccuracy of depth, the noise for how removing depth image is also urgently to be resolved.
Therefore, the existing technology needs to be improved and developed.
Summary of the invention
In view of above-mentioned deficiencies of the prior art, the purpose of the present invention is to provide a kind of depth image denoising methods, device
And storage medium, it is intended to which solving depth transducer in the prior art by the interference of environment light is caused depth measurements inaccurate
The problem of.
Technical scheme is as follows:
A kind of depth image denoising method comprising following steps:
Simulation obtains depth transducer to the original response data of different 3D scenes as training sample set;
Construct the deep learning network for denoising to depth image and the damage for training the deep learning network
Function is lost, wherein the loss function includes space loss item and the loss of time;
The training sample set is input to deep learning network, by the loss function to the deep learning network
Deep learning training is carried out, and obtains denoising acoustic model after carrying out model verifying by the assessment sample set of actual measurement;
By depth transducer actual acquisition to depth image data be input to the denoising acoustic model, after generating denoising
Depth image.
In the depth image denoising method, the simulation obtains depth transducer to the original response of different 3D scenes
Data are as training sample set, comprising:
Obtain the impulse response of the data set and depth transducer of different 3D scenes;
Different 3D scenes are emulated according to the impulse response of the data set and depth transducer of the difference 3D scene
Simulation obtains depth transducer to the original response data of different 3D scenes as training sample set.
In the depth image denoising method, the deep learning network for denoise to depth image for suitable for
The convolutional neural networks of image, semantic segmentation, wherein the input data of the convolutional neural networks is three-dimensional data.
In the depth image denoising method, the loss function constructed for training the deep learning network,
Include:
The loss of time is constructed, the loss of time item isWherein
DKLFor KL divergence,It is the triggering times measured value of position k point, n is time, h(k)[n] is the target value of position k point;
Space loss item is constructed, the space loss item is TV (soft argmax (h(k)),Wherein TV is Quan Bianfen, obtains triggering times by soft argmax function
Time location where peak value;
The loss function for training the deep learning network is constructed, the loss function isλTVFor hyper parameter.
It is described that the training sample set is input to deep learning network in the depth image denoising method, pass through
The loss function carries out deep learning training to the deep learning network, and carries out model by the assessment sample set of actual measurement
Denoising acoustic model is obtained after verifying, comprising:
The training sample set is input to deep learning network and carries out deep learning training to it, according to deep learning net
The output valve of network calculates the value of corresponding loss function;
Continue deep learning training after adjusting the parameter of the deep learning network according to the value of current loss function,
Adjusting parameter until the value of loss function is converged to, less than preset threshold, then complete by training repeatedly;
The deep learning network that the assessment sample set of actual measurement is input to training completion is subjected to model verifying, if authentication failed
It then readjusts parameter and carries out deep learning training, until being proved to be successful;
Using the deep learning network being proved to be successful as denoising acoustic model.
In the depth image denoising method, the depth that the assessment sample set of actual measurement is input to training completion
It practises network and carries out model verifying, parameter is readjusted if authentication failed and carries out deep learning training and is wrapped until being proved to be successful
It includes:
The assessment sample set of actual measurement is input to the deep learning network of training completion;
The output valve of the deep learning network is compared with corresponding actual depth in assessment sample set;
The then judgment models authentication failed when comparing result is greater than preset difference value continues depth after readjusting parameter
Learning training, until model is proved to be successful.
In the depth image denoising method, the depth image data input that depth transducer actual acquisition is arrived
To the denoising acoustic model, before the depth image after generating denoising, further includes:
Two-dimensional image data corresponding with the depth image data is obtained by imaging sensor.
In the depth image denoising method, the depth image data input that depth transducer actual acquisition is arrived
To the denoising acoustic model, the depth image after denoising is generated, comprising:
The depth image data that depth transducer actual acquisition is arrived and the two-dimensional image data carry out sensor fusion
Processing;
Data after fusion treatment are input to the denoising acoustic model, the depth image after generating denoising.
In the depth image denoising method, the depth image data input that depth transducer actual acquisition is arrived
To the denoising acoustic model, after the depth image after generating denoising, further includes:
The depth map after denoising is up-sampled according to the two-dimensional image data, depth transducer is obtained and does not detect
Point depth information.
Further embodiment of this invention additionally provides a kind of depth image denoising device, and described device includes at least one processing
Device;And
The memory being connect at least one described processor communication;Wherein,
The memory is stored with the instruction that can be executed by least one described processor, and described instruction is by described at least one
A processor executes, so that at least one described processor is able to carry out above-mentioned depth image denoising method.
Another embodiment of the present invention additionally provides a kind of non-volatile computer readable storage medium storing program for executing, described non-volatile
Computer-readable recording medium storage has computer executable instructions, and the computer executable instructions are by one or more processors
When execution, one or more of processors may make to execute above-mentioned depth image denoising method.
Another embodiment of the present invention additionally provides a kind of computer program product, and the computer program product includes depositing
The computer program on non-volatile computer readable storage medium storing program for executing is stored up, the computer program includes program instruction, works as institute
When stating program instruction and being executed by processor, the processor is made to execute above-mentioned depth image denoising method.
The utility model has the advantages that the invention discloses a kind of depth image denoising method, device and storage mediums, compared to existing skill
Art, the embodiment of the present invention is by constructing the deep learning network denoised to depth image and having depth measurement signal special
The loss function of sign is trained to the deep learning network by the loss function and obtains after verifying denoising acoustic model,
And then the collected depth image data of depth transducer is denoised using the denoising acoustic model, when eliminating depth measurement
Interference of the environment light to signal light, improves the signal-to-noise ratio of depth image.
Detailed description of the invention
Present invention will be further explained below with reference to the attached drawings and examples, in attached drawing:
Fig. 1 is the flow chart of depth image denoising method preferred embodiment provided by the invention;
Fig. 2 is the structural representation of depth sensing system in depth image denoising method preferred embodiment provided by the invention
Figure;
Fig. 3 be depth image denoising method preferred embodiment provided by the invention in depth transducer triggering times at any time
The statistic histogram of variation;
Fig. 4 is the hardware structural diagram that depth image provided by the invention denoises device preferred embodiment.
Specific embodiment
To make the purpose of the present invention, technical solution and effect clearer, clear and definite, below to the present invention further specifically
It is bright.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not intended to limit the present invention.Below
The embodiment of the present invention is introduced in conjunction with attached drawing.
Referring to Fig. 1, Fig. 1 is the flow chart of depth image denoising method preferred embodiment provided by the invention.Such as Fig. 1 institute
Show comprising following steps:
S100, simulation obtain depth transducer to the original response data of different 3D scenes as training sample set;
S200 constructs deep learning network for denoising to depth image and for training the deep learning network
Loss function, wherein the loss function includes space loss item and the loss of time;
The training sample set is input to deep learning network, by the loss function to the depth by S300
It practises network and carries out deep learning training, and obtain denoising acoustic model after carrying out model verifying by the assessment sample set of actual measurement;
S400, by depth transducer actual acquisition to depth image data be input to the denoising acoustic model, generation is gone
Depth image after making an uproar.
As shown in Fig. 2, (single spy can also be used certainly so that SPAD array obtains the depth sensing system of depth image as an example
Survey device and depth image obtained by mechanical scanning), on the object in pulsed laser irradiation to target scene, in scene while also
Background ambient light (may be regarded as constant homogeneous, do not influenced by object distance in scene), and SPAD array received is anti-in scene
Be emitted back towards come pulse laser and environment light, when the photon of return successfully triggers SPAD, system record triggering occur when
Between, above procedure is repeated as many times, each SPAD sensor is recorded at the time of trigger each time, obtains triggering times at any time
The statistic histogram (histogram) of variation, as shown in figure 3, the original signal that i.e. each SPAD sensor receives is not
The statistical distribution of the number of photons received between simultaneously, due to high sensitivity, SPAD is easy to (be different from master by the photon in environment
The signal photon that dynamic light source issues) excitation, therefore can have a degree of noise in the signal received, what is be detected
In the limited situation of photon numbers (it may be low by the reflectivity of object, the power of signal light is weak, and the testing time is short, object distance
Situations such as remote, is caused), since the number of photons of return obeys Poisson stochastic process, generate shot noise (shot noise), such as
Fruit is only simply regarded as signal position according to the time location where highest point on histogram, it is more likely that no
The time of return of signal photon can accurately be obtained, so as to cause the calculating mistake of distance.
Therefore, in the present embodiment, the depth map that depth transducer actual acquisition arrives is extracted using deep learning algorithm
As the signal light in data, the interference of filtering environmental light obtains the depth image of high s/n ratio.Specifically, due to depth
Habit needs huge sample when being trained, and first simulation obtains original response data conduct of the depth transducer to different 3D scenes
Training sample set constructs the deep learning network for denoising to depth image and for training the deep learning net later
The loss function of network, wherein the loss function includes space loss item and the loss of time, i.e., according to the sound of depth transducer
The rule of induction signal, building one, for training the loss function of deep learning network, later inputs the training sample set
To deep learning network, by it is above-mentioned embody loss function regular included in response signal to deep learning network into
After the enough training of row, further it is suitable for target signal to noise ratio by obtaining after the assessment sample set progress model verifying of actual measurement
Acoustic model is denoised, i.e., by showing that the denoising acoustic model currently obtained can be in practical feelings after enough training and model verifying
Noise is correctly handled in condition, thus the depth image number arrived by the denoising acoustic model after deep learning to actual acquisition
According to being denoised, the interference of environment light is effectively reduced, the depth image after generating denoising, realization is gone based on deep learning algorithm
It makes an uproar processing.
Further, the simulation obtains depth transducer to the original response data of different 3D scenes as training sample
Collection, comprising:
Obtain the impulse response of the data set and depth transducer of different 3D scenes;
Different 3D scenes are emulated according to the impulse response of the data set and depth transducer of the difference 3D scene
Simulation obtains depth transducer to the original response data of different 3D scenes as training sample set.
It needs huge sample data when being trained to deep learning network, in the present embodiment, passes through analog simulation
Training sample set is obtained, the data set for obtaining the different 3D scenes of actual measurement can be downloaded on the internet first, which includes fields
The distance and reflectivity of object in scape, and pass through the impulse response of experiment test acquisition depth transducer, by taking SPAD as an example, arteries and veins
Punching response includes the dark current and dark counting (response of SPAD when without photon incidence), environment light and shot noise of SPAD
Influence, structure based on depth sensing system as shown in Figure 2 carries out analogue simulation, obtains SPAD array corresponding to 3D scene
Data set in difference 3D scene original response data, the original response data that will acquire are as training sample set, by 3D
Known depth map is set as the truthful data/model answer demarcated in the data set of scene, wherein for obtained by each scene
To original response data be three-dimensional, be x-axis, y-axis and time t direction respectively, thus before completing deep learning training
Data preparation provides data basis for the training of subsequent deep learning, obtains more accurately and effectively denoising acoustic model.
Further, described for being the volume divided suitable for image, semantic to the deep learning network that depth image denoises
Product neural network, wherein the input data of the convolutional neural networks is three-dimensional data.
In the present embodiment, deep learning training is carried out by deep learning network and realizes denoising function, wherein the depth
Learning network is the convolutional neural networks (convolution neural network, CNN) divided suitable for image, semantic, right
The convolutional neural networks for being suitable for image, semantic segmentation carry out realizing denoising function after accordingly modifying, wherein by the convolution mind
Input data through network is revised as three-dimensional data, that is, includes x-axis, the three-dimensional data of y-axis and time t direction, in addition determine again
Justice constructs the loss function for training the deep learning network, embodies depth transducer response signal institute by loss function
The rule for including denoises so that the deep learning network after the completion of training has denoising function for subsequent depth image.
Further, the loss function constructed for training the deep learning network, comprising:
The loss of time is constructed, the loss of time item isWherein
DKLFor KL divergence,It is the triggering times measured value of position k point, n is time, h(k)[n] is the target value of position k point;
Space loss item is constructed, the space loss item is TV (soft argmax (h(k)),Wherein TV is Quan Bianfen, obtains triggering times by soft argmax function
Time location where peak value;
The loss function for training the deep learning network is constructed, the loss function isλTVFor hyper parameter.
In the present embodiment, loss function is the target as deep learning, for measuring the knot of deep learning network output
Difference between fruit and the truthful data demarcated, difference is the smaller the better, is reached with this to be trained to deep learning network
Target, the application construct corresponding loss function for the rule of the response signal of depth transducer, specifically, deep
The response signal of sensor is spent with the rule in terms of in terms of the time and space.On the one hand, each pixel of the image of natural scene
Between there is certain association, be not completely random, scene image be often it is smooth, be mainly made of low frequency component
, also have identical rule for depth, for example, other than object edge, a pixel more likely with the picture of surrounding
Element possess very close to depth value, therefore the reliability of each location point depth information can be evaluated by loss of time;Separately
On the one hand, the signal light and the characteristic of background ambient light in time that depth transducer receives have very big difference, and signal light exists
It is concentration on time, heterogeneous, and the intensity of signal is related at a distance from object, and environment light is equally distributed, and
Unrelated with scene, when photon numbers are very limited, signal light is usually concentrated in together, and triggering caused by environment light is random
Distribution, it, can be with when the signal to a SPAD pixel unit is handled and due to the relevance between adjacent pixel
The statistic histogram of pixel (such as 8 nearest points) around the pixel is combined together and is handled, such as is set in
One, in time window similar in pulse width, after the number of triggering reaches minimum number, can just be considered signal, otherwise
It will be handled as noise, therefore the reliable of each depth transducer unit triggers signal can be evaluated by space loss item
Property.
When it is implemented, the loss function is made of two parts, for embodying the above-mentioned depth transducer response referred to
It include two rules in signal, first part is the loss of time, the specially log-likelihood of the depth of each point and time-domain signal
Function, can be by KL divergence (Kullback-Leibler divergence) Lai Hengliang;The definition of KL divergence isWhereinIt is the triggering times measured value of position k point, n is that statistics is straight
The time of square figure, h(k)[n] is the target (ground truth) of position k point, i.e. truthful data;Second part is space loss
, specifically first with soft argmax functionFrom triggering times-time system
The time location where triggering times peak value is obtained in meter histogram, depth can be calculated according to the time, further to soft
Full variation (total variation) TV (soft argmax (h for the estimation of Depth that argmax function obtains(k)) it is used as space
Item is lost, the weight in loss function between above-mentioned two can pass through a hyper parameter λTVIt adjusts, therefore this is used to train
The complete loss function of Damage degree learning network are as follows:Utilizing analog simulation
Obtained training sample set and loss function can obtain corresponding denoising acoustic mode after carrying out enough training to deep learning network
Type.
Further, described that the training sample set is input to deep learning network, by the loss function to institute
It states deep learning network and carries out deep learning training, and obtain denoising after carrying out model verifying by the assessment sample set of actual measurement
Model, comprising:
The training sample set is input to deep learning network and carries out deep learning training to it, according to deep learning net
The output valve of network calculates the value of corresponding loss function;
Continue deep learning training after adjusting the parameter of the deep learning network according to the value of current loss function,
Adjusting parameter until the value of loss function is converged to, less than preset threshold, then complete by training repeatedly;
The deep learning network that the assessment sample set of actual measurement is input to training completion is subjected to model verifying, if authentication failed
It then readjusts parameter and carries out deep learning training, until being proved to be successful;
Using the deep learning network being proved to be successful as denoising acoustic model.
In the present embodiment, when being specifically trained, the training sample set that analog simulation is obtained is input to deep learning
The deep learning network is trained in network, the value of corresponding loss function is calculated according to the output valve of deep learning network,
With the difference between this result and truthful data demarcate to measure the output of current depth learning network, according to currently losing
The value of function continues to be trained after adjusting weight/parameter of the deep learning network, continues according to deep learning network
Output valve calculates the value and adjusting parameter of corresponding loss function, repeats the above process, i.e., is adjusted repeatedly according to the value of loss function
The parameter of deep learning network then indicates to train less than preset threshold completion until the value of loss function is converged to, specific default
Threshold value can be configured according to application requirement and target signal to noise ratio etc., i.e., when the value of loss function is converged to less than preset threshold
When, it is poor already less than preset threshold between the result and truthful data demarcate of current depth learning network output to illustrate,
The parameter of current depth learning network be suitable for denoising task, the output for being substantially equal to truthful data can be obtained as a result, it
Afterwards, to ensure that current model can correctly handle noise in actual treatment, pass through the assessment sample set of actual measurement in the present embodiment
Model verifying is carried out to the deep learning network that training is completed, shows that "current" model can in a practical situation just if being proved to be successful
Really processing noise data, need to return if authentication failed readjust parameter and carry out deep learning training, until verifying at
Function, with ensure deep learning network in subsequent actual treatment to the accuracy of denoising, therefore will training complete and
The deep learning network that model is proved to be successful is as denoising acoustic model.
Preferably, the deep learning network progress model that the assessment sample set of actual measurement is input to training completion is tested
Card readjusts parameter if authentication failed and carries out deep learning training, until being proved to be successful, comprising:
The assessment sample set of actual measurement is input to the deep learning network of training completion;
The output valve of the deep learning network is compared with corresponding actual depth in assessment sample set;
The then judgment models authentication failed when comparing result is greater than preset difference value continues depth after readjusting parameter
Learning training, until model is proved to be successful.
In the present embodiment, after training, the deep learning completed is trained to training using the assessment sample set of actual measurement
Network carries out assessment test, and the specific assessment sample set is that depth transducer carries out the scene of different known depth distributions
Survey obtained sample set, i.e. the depth value of each actual measurement has corresponded to an actual depth value, when carrying out model verifying,
The assessment sample set of actual measurement is input to the deep learning network of training completion, later by the output valve of the deep learning network
With assessment sample set in corresponding actual depth compare, compare size of the difference between the two, if comparing result be less than etc.
In preset difference value, show the difference after data of the "current" model to actual measurement denoise between actual depth value default
Range is judged as is proved to be successful that at this time "current" model successfully can carry out denoising to measured data;If deep learning network
Output valve and actual depth between deviation be greater than preset difference value, illustrate denoising inaccuracy, influence denoise effect, need
It returns and continues deep learning training after readjusting parameter, until model is proved to be successful, is stablized and accurately denoised
Acoustic model, it is ensured that the reliability of denoising function.
Further, it is described by depth transducer actual acquisition to depth image data be input to the denoising acoustic mode
Type, generate denoising after depth image before, further includes:
Two-dimensional image data corresponding with the depth image data is obtained by imaging sensor.
In the present embodiment, although it also has array sizes since depth transducer can obtain the three-dimensional information of scene
The small, defects such as pixel quantity is few, resolution ratio is low, and imaging sensor can obtain high-resolution image data, but dimension only limits
In 2-D data, it is lost depth direction information, therefore is directed to the advantage and disadvantage of both sensors, this example can be to depth map
Two-dimensional image data corresponding with depth image data is also obtained before as carrying out denoising, wherein the high-resolution two dimension
Image data can be RGB image and be also possible to light distribution image, and the spy of two kinds of image datas is combined in subsequent processing
Point exports to advanced optimize as a result, obtaining the better depth image of effect.
Specifically, in one embodiment, it is described by depth transducer actual acquisition to depth image data be input to
The denoising acoustic model, the depth image after generating denoising, comprising:
The depth image data that depth transducer actual acquisition is arrived and the two-dimensional image data carry out sensor fusion
Processing;
Data after fusion treatment are input to the denoising acoustic model, the depth image after generating denoising.
In the present embodiment, after getting high-resolution two-dimensional image data using traditional imaging sensor, first will
The depth image data and the two-dimensional image data that depth transducer actual acquisition arrives carry out sensor fusion treatment, will merge
Data that treated are input to denoising acoustic model, i.e., two kinds of obtained information of sensor, In are combined in deep learning network
To when for example the collected depth image of SPAD array even depth sensor-based system denoises, two-dimensional image data is introduced, to go
Model of making an uproar provides additional reference information, further increases the accuracy of depth image denoising.
Further, in another embodiment, the depth image data that depth transducer actual acquisition is arrived is defeated
Enter to the denoising acoustic model, after the depth image after generation denoising, further includes:
The depth map after denoising is up-sampled according to the two-dimensional image data, depth transducer is obtained and does not detect
Point depth information.
In the present embodiment, since SPAD sensor is difficult to accomplish large-scale sensing unit array, array sizes are limited to be led
It causes depth image lower in lateral resolution ratio, in order to improve the resolution ratio of depth image, is obtaining high-resolution two dimension
After image data, depth image can be subjected to characteristic matching according to high-resolution two-dimensional image data, utilize two dimensional image
Data are up-sampled (upsampling) to the depth map after the denoising of low resolution, accurately to speculate that obtaining depth passes with this
Sensor array fails the depth information of the point detected, in conjunction with the advantages of two kinds of sensors, makes the resolution ratio of depth image significantly
It improves, further realizes super-resolution imaging on the basis of depth image denoising, it is preferable that the present embodiment is in deep learning network
In, loss function can be modified accordingly to embody by up-sampling the error of depth and actual depth being calculated,Wherein,For deep learning net
The depth of network output, zhFor the real depth data demarcated, λupFor the hyper parameter for adjusting weight between the two, so that instruction
Denoising acoustic model after white silk is more applicable for the depth sensing system up-sampled by two-dimensional image data, both realizes depth
The accurate denoising of degree image also improves the lateral resolution of depth image.
Another embodiment of the present invention provides a kind of depth images to denoise device, as shown in figure 4, device 10 includes:
One or more processors 110 and memory 120 are introduced in Fig. 4 by taking a processor 110 as an example, are located
Reason device 110 can be connected with memory 120 by bus or other modes, in Fig. 4 for being connected by bus.
Processor 110 is used for the various control logics of finishing device 10, can be general processor, Digital Signal Processing
Device (DSP), specific integrated circuit (ASIC), field programmable gate array (FPGA), single-chip microcontroller, ARM (Acorn RISC
) or other programmable logic device, discrete gate or transistor logic, discrete hardware component or these components Machine
Any combination.In addition, processor 110 can also be any conventional processors, microprocessor or state machine.Processor 110 can also
To be implemented as calculating the combination of equipment, for example, the combination of DSP and microprocessor, multi-microprocessor, one or more micro- places
Manage device combination DSP core or any other this configuration.
Memory 120 is used as a kind of non-volatile computer readable storage medium storing program for executing, can be used for storing non-volatile software journey
Sequence, non-volatile computer executable program and module, as the depth image denoising method in the embodiment of the present invention is corresponding
Program instruction.Non-volatile software program, instruction and the unit that processor 110 is stored in memory 120 by operation, from
And the various function application and data processing of executive device 10, i.e. depth image denoising side in realization above method embodiment
Method.
Memory 120 may include storing program area and storage data area, wherein storing program area can store operation dress
It sets, application program required at least one function;Storage data area, which can be stored, uses created data etc. according to device 10.
It can also include nonvolatile memory in addition, memory 120 may include high-speed random access memory, for example, at least one
A disk memory, flush memory device or other non-volatile solid state memory parts.In some embodiments, memory 120 can
Choosing includes the memory remotely located relative to processor 110, these remote memories can pass through network connection to device 10.
The example of above-mentioned network includes but is not limited to internet, intranet, local area network, mobile radio communication and combinations thereof.
One or more unit is stored in memory 120, when being executed by one or more processor 110, is held
Depth image denoising method in the above-mentioned any means embodiment of row, for example, executing the method and step in Fig. 1 described above
S100 to step S400.
The embodiment of the invention provides a kind of non-volatile computer readable storage medium storing program for executing, computer readable storage medium is deposited
Computer executable instructions are contained, which is executed by one or more processors, for example, executing above retouch
Method and step S100 to step S400 in the Fig. 1 stated.
As an example, non-volatile memory medium can include that read-only memory (ROM), programming ROM (PROM), electricity can
Programming ROM (EPROM), electrically erasable ROM (EEPROM) or flash memory.Volatile memory can include as external high speed
The random access memory (RAM) of buffer memory.By illustrate it is beautiful unrestricted, RAM can with such as synchronous random access memory (SRAM),
Dynamic ram, (DRAM), synchronous dram (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM
(ESDRAM), many forms of Synchlink DRAM (SLDRAM) and directly Rambus (Lan Basi) RAM (DRRAM) etc
It obtains.The disclosed memory assembly or memory of operating environment described herein be intended to include these and/or it is any
Other are suitble to one or more of the memory of type.
Another embodiment of the invention provides a kind of computer program product, and computer program product includes being stored in
Computer program on non-volatile computer readable storage medium storing program for executing, computer program include program instruction, when program instruction quilt
When processor executes, the processor is made to execute the depth image denoising method of above method embodiment.For example, executing above retouch
Method and step S100 to step S400 in the Fig. 1 stated.
In conclusion in depth image denoising method disclosed by the invention, device and storage medium, which comprises
Simulation obtains depth transducer to the original response data of different 3D scenes as training sample set;Building is for depth image
The deep learning network of denoising and loss function for training the deep learning network, wherein loss function includes space
Lose item and the loss of time;Training sample set is input to deep learning network, by loss function to deep learning network
Deep learning training is carried out, and obtains denoising acoustic model after carrying out model verifying by the assessment sample set of actual measurement;Depth is passed
Sensor actual acquisition to depth image data be input to the denoising acoustic model, the depth image after generating denoising.The present invention
Loss of the embodiment by constructing the deep learning network denoised to depth image and with depth measurement signal characteristic
Function is trained the deep learning network by the loss function and obtains denoising acoustic model after verifying, and then utilizes
The denoising acoustic model denoises the collected depth image data of depth transducer, environment light pair when eliminating depth measurement
The interference of signal light improves the signal-to-noise ratio of depth image.
Embodiments described above is only schematical, wherein as illustrated by the separation member unit can be or
It may not be and be physically separated, component shown as a unit may or may not be physical unit, it can
It is in one place, or may be distributed over multiple network units.Can select according to actual needs part therein or
Person's whole module achieves the purpose of the solution of this embodiment.
By the description of above embodiment, those skilled in the art can be understood that each embodiment can be by
Software adds the mode of general hardware platform to realize, naturally it is also possible to pass through hardware realization.Based on this understanding, above-mentioned technology
Scheme substantially in other words can be embodied in the form of software products the part that the relevant technologies contribute, the computer
Software product can reside in computer readable storage medium, such as ROM/RAM, magnetic disk, CD, including some instructions to
So that a computer electronic equipment (can be personal computer, server or network electronic devices etc.) executes each reality
The method for applying certain parts of example or embodiment.
Among other things, such as " can ', " energy ", " possibility " or " can be with " etc conditional statement unless in addition specific
Ground is stated or is otherwise understood in context as used, is otherwise generally intended to convey particular implementation energy
Including (however other embodiments do not include) special characteristic, element and/or operation.Therefore, such conditional statement is generally
It is not intended to imply that feature, element and/or operation are all needed one or more embodiments or one anyway
Or multiple embodiments must include for determining these features, element in the case where being with or without student's input or prompt
And/or the logic whether operation is included or will be performed in any particular implementation.
The content described in the present description and drawings herein include be capable of providing depth image denoising method,
The example of device and storage medium.Certainly, can not for description the disclosure various features purpose come describe element and/or
The combination that each of method is envisioned that, it can be appreciated that, many other combinations and displacement of disclosed feature are
It is possible., it will thus be apparent that can be made without departing from the scope or spirit of the present disclosure to the disclosure respectively
Kind modification.In addition, or in alternative solution, the considerations of other embodiments of the disclosure are to the specification and drawings and such as this
It may be obvious in the practice of the disclosure presented in text.It is intended that showing proposed in the specification and drawings
Example is considered illustrative and not restrictive in all respects.Although using specific term herein, they
The purpose of limitation is used and is not used in general and descriptive sense.
Claims (11)
1. a kind of depth image denoising method, which comprises the steps of:
Simulation obtains depth transducer to the original response data of different 3D scenes as training sample set;
Construct the deep learning network for denoising to depth image and the loss letter for training the deep learning network
Number, wherein the loss function includes space loss item and the loss of time;
The training sample set is input to deep learning network, the deep learning network is carried out by the loss function
Deep learning training, and denoising acoustic model is obtained after carrying out model verifying by the assessment sample set of actual measurement;
By depth transducer actual acquisition to depth image data be input to the denoising acoustic model, the depth after generating denoising
Image.
2. depth image denoising method according to claim 1, which is characterized in that the simulation obtains depth transducer pair
The original response data of different 3D scenes are as training sample set, comprising:
Obtain the impulse response of the data set and depth transducer of different 3D scenes;
Analogue simulation is carried out to different 3D scenes according to the impulse response of the data set and depth transducer of the difference 3D scene,
Depth transducer is obtained to the original response data of different 3D scenes as training sample set.
3. depth image denoising method according to claim 1, which is characterized in that described for being denoised to depth image
Deep learning network is the convolutional neural networks divided suitable for image, semantic, wherein the input data of the convolutional neural networks
For three-dimensional data.
4. depth image denoising method according to claim 1, which is characterized in that the building is for training the depth
The loss function of learning network, comprising:
The loss of time is constructed, the loss of time item isWherein DKLFor KL
Divergence,It is the triggering times measured value of position k point, n is time, h(k)[n] is the target value of position k point;
Space loss item is constructed, the space loss item is TV (soft argmax (h(k)),Wherein TV is Quan Bianfen, obtains triggering times by soft argmax function
Time location where peak value;
The loss function for training the deep learning network is constructed, the loss function isλTVFor hyper parameter.
5. depth image denoising method according to claim 1, which is characterized in that described to input the training sample set
To deep learning network, deep learning training is carried out to the deep learning network by the loss function, and passes through actual measurement
Assessment sample set carry out model verifying after obtain denoising acoustic model, comprising:
The training sample set is input to deep learning network and carries out deep learning training to it, according to deep learning network
Output valve calculates the value of corresponding loss function;
Continue deep learning training after adjusting the parameter of the deep learning network according to the value of current loss function, repeatedly
Adjusting parameter until the value of loss function is converged to, less than preset threshold, then complete by training;
The deep learning network that the assessment sample set of actual measurement is input to training completion is subjected to model verifying, is weighed if authentication failed
New adjusting parameter simultaneously carries out deep learning training, until being proved to be successful;
Using the deep learning network being proved to be successful as denoising acoustic model.
6. depth image denoising method according to claim 5, which is characterized in that the assessment sample set by actual measurement is defeated
Enter to the deep learning network that training is completed and carry out model verifying, parameter is readjusted if authentication failed and carries out deep learning
Training, until being proved to be successful, comprising:
The assessment sample set of actual measurement is input to the deep learning network of training completion;
The output valve of the deep learning network is compared with corresponding actual depth in assessment sample set;
The then judgment models authentication failed when comparing result is greater than preset difference value continues deep learning after readjusting parameter
Training, until model is proved to be successful.
7. depth image denoising method according to claim 1, which is characterized in that described by depth transducer actual acquisition
To depth image data be input to the denoising acoustic model, before the depth image after generating denoising, further includes:
Two-dimensional image data corresponding with the depth image data is obtained by imaging sensor.
8. depth image denoising method according to claim 7, which is characterized in that described by depth transducer actual acquisition
To depth image data be input to the denoising acoustic model, the depth image after generating denoising, comprising:
The depth image data that depth transducer actual acquisition is arrived and the two-dimensional image data carry out sensor fusion treatment;
Data after fusion treatment are input to the denoising acoustic model, the depth image after generating denoising.
9. depth image denoising method according to claim 7, which is characterized in that described by depth transducer actual acquisition
To depth image data be input to the denoising acoustic model, after the depth image after generating denoising, further includes:
The depth map after denoising is up-sampled according to the two-dimensional image data, obtains the point that depth transducer does not detect
Depth information.
10. a kind of depth image denoises device, which is characterized in that described device includes at least one processor;And
The memory being connect at least one described processor communication;Wherein,
The memory is stored with the instruction that can be executed by least one described processor, and described instruction is by described at least one
It manages device to execute, so that at least one described processor is able to carry out the described in any item depth image denoising sides claim 1-9
Method.
11. a kind of non-volatile computer readable storage medium storing program for executing, which is characterized in that the non-volatile computer readable storage medium
Matter is stored with computer executable instructions, when which is executed by one or more processors, may make institute
It states one or more processors perform claim and requires the described in any item depth image denoising methods of 1-9.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910728854.4A CN110458778A (en) | 2019-08-08 | 2019-08-08 | A kind of depth image denoising method, device and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910728854.4A CN110458778A (en) | 2019-08-08 | 2019-08-08 | A kind of depth image denoising method, device and storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110458778A true CN110458778A (en) | 2019-11-15 |
Family
ID=68485377
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910728854.4A Pending CN110458778A (en) | 2019-08-08 | 2019-08-08 | A kind of depth image denoising method, device and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110458778A (en) |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111160522A (en) * | 2019-12-12 | 2020-05-15 | 第四范式(北京)技术有限公司 | Anti-noise method, device and system for machine learning |
CN111208509A (en) * | 2020-01-15 | 2020-05-29 | 中国人民解放军国防科技大学 | Ultra-wideband radar human body target posture visualization enhancing method |
CN111428374A (en) * | 2020-03-30 | 2020-07-17 | 苏州惟信易量智能科技有限公司 | Part defect detection method, device, equipment and storage medium |
CN111603191A (en) * | 2020-05-29 | 2020-09-01 | 上海联影医疗科技有限公司 | Voice noise reduction method and device in medical scanning and computer equipment |
CN112149617A (en) * | 2020-10-13 | 2020-12-29 | 中国工程物理研究院计算机应用研究所 | Pulse waveform denoising method based on deep learning |
CN112346076A (en) * | 2020-11-25 | 2021-02-09 | Oppo(重庆)智能科技有限公司 | Control method of electronic device, and computer-readable storage medium |
CN112461360A (en) * | 2020-10-26 | 2021-03-09 | 北京理工大学 | High-resolution single photon imaging method and system combined with physical noise model |
CN113009508A (en) * | 2019-12-20 | 2021-06-22 | 舜宇光学(浙江)研究院有限公司 | Multipath interference correction method for TOF module, system and electronic equipment thereof |
CN113050120A (en) * | 2021-03-24 | 2021-06-29 | Oppo广东移动通信有限公司 | Imaging method, electronic apparatus, control apparatus, and storage medium |
CN113313653A (en) * | 2021-06-23 | 2021-08-27 | 北京鼎泰智源科技有限公司 | Image denoising method and device based on generative countermeasure network |
CN113327209A (en) * | 2021-06-29 | 2021-08-31 | Oppo广东移动通信有限公司 | Depth image generation method and device, electronic equipment and storage medium |
CN113469898A (en) * | 2021-06-02 | 2021-10-01 | 北京邮电大学 | Image distortion removal method based on deep learning and related equipment |
CN113628190A (en) * | 2021-08-11 | 2021-11-09 | 跃维(佛山)智能科技有限公司 | Depth map denoising method and device, electronic equipment and medium |
CN113867659A (en) * | 2020-06-30 | 2021-12-31 | 西安诺瓦星云科技股份有限公司 | Image output method and device and module controller |
CN113901927A (en) * | 2021-10-12 | 2022-01-07 | 大连海事大学 | Underwater object shape recognition method based on flow field pressure time course |
CN114723797A (en) * | 2021-01-06 | 2022-07-08 | 南京大学 | TOF depth imaging method based on depth learning |
CN114764803A (en) * | 2022-06-16 | 2022-07-19 | 深圳深知未来智能有限公司 | Noise evaluation method and device based on real noise scene and storage medium |
US20220352997A1 (en) * | 2019-09-18 | 2022-11-03 | Zte Corporation | Interference simulation system, method and apparatus, interference test system, method and apparatus, and computer readable storage medium |
CN116681618A (en) * | 2023-06-13 | 2023-09-01 | 强联智创(北京)科技有限公司 | Image denoising method, electronic device and storage medium |
WO2023216057A1 (en) * | 2022-05-09 | 2023-11-16 | Shanghai United Imaging Healthcare Co., Ltd. | System and method for medical imaging |
CN117876263A (en) * | 2024-03-13 | 2024-04-12 | 之江实验室 | Astronomical image processing method and device |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150112897A1 (en) * | 2013-10-23 | 2015-04-23 | Mitsubishi Electric Research Laboratories, Inc. | Method for Estimating Parameters of a Graph Spectral Filter Using Training Data |
CN108447113A (en) * | 2018-03-22 | 2018-08-24 | 北京理工大学 | A kind of pulsed intensity correlation three-D imaging method based on deep learning model |
CN109003240A (en) * | 2018-07-16 | 2018-12-14 | 安徽理工大学 | A kind of image de-noising method based on multiple dimensioned parallel C NN |
CN109658344A (en) * | 2018-11-12 | 2019-04-19 | 哈尔滨工业大学(深圳) | Image de-noising method, device, equipment and storage medium based on deep learning |
-
2019
- 2019-08-08 CN CN201910728854.4A patent/CN110458778A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150112897A1 (en) * | 2013-10-23 | 2015-04-23 | Mitsubishi Electric Research Laboratories, Inc. | Method for Estimating Parameters of a Graph Spectral Filter Using Training Data |
CN108447113A (en) * | 2018-03-22 | 2018-08-24 | 北京理工大学 | A kind of pulsed intensity correlation three-D imaging method based on deep learning model |
CN109003240A (en) * | 2018-07-16 | 2018-12-14 | 安徽理工大学 | A kind of image de-noising method based on multiple dimensioned parallel C NN |
CN109658344A (en) * | 2018-11-12 | 2019-04-19 | 哈尔滨工业大学(深圳) | Image de-noising method, device, equipment and storage medium based on deep learning |
Non-Patent Citations (4)
Title |
---|
DAVID B. LINDELL 等: "Single-Photon 3D Imaging with Deep Sensor Fusion", 《ACM》 * |
何东健 等: "Kinect获取植物三维点云数据的去噪方法", 《农业机械学报》 * |
岳建平 等: "《安全监测技术与应用》", 31 August 2018 * |
黄孝平: "《当代机器深度学习方法与应用研究》", 30 November 2017 * |
Cited By (31)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20220352997A1 (en) * | 2019-09-18 | 2022-11-03 | Zte Corporation | Interference simulation system, method and apparatus, interference test system, method and apparatus, and computer readable storage medium |
CN111160522A (en) * | 2019-12-12 | 2020-05-15 | 第四范式(北京)技术有限公司 | Anti-noise method, device and system for machine learning |
CN111160522B (en) * | 2019-12-12 | 2023-11-24 | 第四范式(北京)技术有限公司 | Noise-resistant method, device and system for machine learning |
CN113009508B (en) * | 2019-12-20 | 2023-11-07 | 舜宇光学(浙江)研究院有限公司 | Multipath interference correction method for TOF module, system and electronic equipment thereof |
CN113009508A (en) * | 2019-12-20 | 2021-06-22 | 舜宇光学(浙江)研究院有限公司 | Multipath interference correction method for TOF module, system and electronic equipment thereof |
CN111208509A (en) * | 2020-01-15 | 2020-05-29 | 中国人民解放军国防科技大学 | Ultra-wideband radar human body target posture visualization enhancing method |
CN111208509B (en) * | 2020-01-15 | 2020-12-29 | 中国人民解放军国防科技大学 | Ultra-wideband radar human body target posture visualization enhancing method |
CN111428374A (en) * | 2020-03-30 | 2020-07-17 | 苏州惟信易量智能科技有限公司 | Part defect detection method, device, equipment and storage medium |
CN111603191A (en) * | 2020-05-29 | 2020-09-01 | 上海联影医疗科技有限公司 | Voice noise reduction method and device in medical scanning and computer equipment |
CN111603191B (en) * | 2020-05-29 | 2023-10-20 | 上海联影医疗科技股份有限公司 | Speech noise reduction method and device in medical scanning and computer equipment |
CN113867659B (en) * | 2020-06-30 | 2024-02-06 | 西安诺瓦星云科技股份有限公司 | Image output method and device and module controller |
CN113867659A (en) * | 2020-06-30 | 2021-12-31 | 西安诺瓦星云科技股份有限公司 | Image output method and device and module controller |
CN112149617A (en) * | 2020-10-13 | 2020-12-29 | 中国工程物理研究院计算机应用研究所 | Pulse waveform denoising method based on deep learning |
CN112461360B (en) * | 2020-10-26 | 2021-10-12 | 北京理工大学 | High-resolution single photon imaging method and system combined with physical noise model |
CN112461360A (en) * | 2020-10-26 | 2021-03-09 | 北京理工大学 | High-resolution single photon imaging method and system combined with physical noise model |
CN112346076A (en) * | 2020-11-25 | 2021-02-09 | Oppo(重庆)智能科技有限公司 | Control method of electronic device, and computer-readable storage medium |
CN114723797B (en) * | 2021-01-06 | 2024-04-12 | 南京大学 | TOF depth imaging method based on deep learning |
CN114723797A (en) * | 2021-01-06 | 2022-07-08 | 南京大学 | TOF depth imaging method based on depth learning |
CN113050120A (en) * | 2021-03-24 | 2021-06-29 | Oppo广东移动通信有限公司 | Imaging method, electronic apparatus, control apparatus, and storage medium |
CN113469898A (en) * | 2021-06-02 | 2021-10-01 | 北京邮电大学 | Image distortion removal method based on deep learning and related equipment |
CN113313653A (en) * | 2021-06-23 | 2021-08-27 | 北京鼎泰智源科技有限公司 | Image denoising method and device based on generative countermeasure network |
CN113327209A (en) * | 2021-06-29 | 2021-08-31 | Oppo广东移动通信有限公司 | Depth image generation method and device, electronic equipment and storage medium |
CN113628190A (en) * | 2021-08-11 | 2021-11-09 | 跃维(佛山)智能科技有限公司 | Depth map denoising method and device, electronic equipment and medium |
CN113628190B (en) * | 2021-08-11 | 2024-03-15 | 跨维(深圳)智能数字科技有限公司 | Depth map denoising method and device, electronic equipment and medium |
CN113901927B (en) * | 2021-10-12 | 2023-02-03 | 大连海事大学 | Underwater object shape recognition method based on flow field pressure time course |
CN113901927A (en) * | 2021-10-12 | 2022-01-07 | 大连海事大学 | Underwater object shape recognition method based on flow field pressure time course |
WO2023216057A1 (en) * | 2022-05-09 | 2023-11-16 | Shanghai United Imaging Healthcare Co., Ltd. | System and method for medical imaging |
CN114764803A (en) * | 2022-06-16 | 2022-07-19 | 深圳深知未来智能有限公司 | Noise evaluation method and device based on real noise scene and storage medium |
CN116681618A (en) * | 2023-06-13 | 2023-09-01 | 强联智创(北京)科技有限公司 | Image denoising method, electronic device and storage medium |
CN117876263A (en) * | 2024-03-13 | 2024-04-12 | 之江实验室 | Astronomical image processing method and device |
CN117876263B (en) * | 2024-03-13 | 2024-05-17 | 之江实验室 | Astronomical image processing method and device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110458778A (en) | A kind of depth image denoising method, device and storage medium | |
US11402496B2 (en) | Method and apparatus for enhancing semantic features of SAR image oriented small set of samples | |
EP3293705B1 (en) | 3d reconstruction of a real object from a depth map | |
Guo et al. | Tackling 3d tof artifacts through learning and the flat dataset | |
EP3924879B1 (en) | Systems and methods involving creation and/or utilization of image mosaics | |
US20110200249A1 (en) | Surface detection in images based on spatial data | |
CN109285149A (en) | Appraisal procedure, device and the calculating equipment of quality of human face image | |
US11373396B2 (en) | Systems and methods for improved feature extraction using polarization information | |
EP3480782A1 (en) | Method and device for reducing noise in a depth image | |
JP2013101113A (en) | Method for performing 3d reconfiguration of object in scene | |
CN111429515A (en) | Learning method of robot obstacle avoidance behavior based on deep learning | |
JP7293488B2 (en) | How to simulate a continuous wave lidar sensor | |
WO2023075885A1 (en) | Denoising depth image data using neural networks | |
Cifuentes et al. | Modeling 3D canopy structure and transmitted PAR using terrestrial LiDAR | |
CN110490058A (en) | Training method, device, system and the computer-readable medium of pedestrian detection model | |
CN113009508B (en) | Multipath interference correction method for TOF module, system and electronic equipment thereof | |
JP7277342B2 (en) | Curved object segmentation using geometric antecedents | |
WO2021074502A1 (en) | Improved location of an acoustic source | |
JP7197003B2 (en) | Depth estimation device, depth estimation method, and depth estimation program | |
CN111753887A (en) | Point source target image control point detection model training method and device | |
US20240126952A1 (en) | Arithmetic operation system, training method, and non-transitory computer readable medium storing training program | |
KR20210114238A (en) | Method and apparatus for camera calibration | |
US20240125935A1 (en) | Arithmetic operation system, training method, and non-transitory computer readable medium storing training program | |
US20240126953A1 (en) | Arithmetic operation system, training method, and non-transitory computer readable medium storing training program | |
Chen et al. | Hyperspectral target detection via multiple instance LSTM target localization network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191115 |