CN110211057A - A kind of image processing method based on full convolutional network, device and computer equipment - Google Patents
A kind of image processing method based on full convolutional network, device and computer equipment Download PDFInfo
- Publication number
- CN110211057A CN110211057A CN201910402733.0A CN201910402733A CN110211057A CN 110211057 A CN110211057 A CN 110211057A CN 201910402733 A CN201910402733 A CN 201910402733A CN 110211057 A CN110211057 A CN 110211057A
- Authority
- CN
- China
- Prior art keywords
- input picture
- image
- low resolution
- linear relationship
- input
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- 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
- G06T2207/10004—Still image; Photographic image
-
- 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]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
Abstract
This application involves a kind of image processing method based on full convolutional network, device and computer equipments, the described method includes: reducing the resolution ratio of the first input picture, it obtains the first low resolution input picture and is input in full convolutional network, obtain low resolution output image；The first input picture and the first low resolution input picture are converted into the second input picture and the second low resolution input picture respectively；Using top sampling method, processing image is obtained by the second low resolution input picture and low resolution output image, the first linear relationship between processing image and the second low resolution input picture is calculated；Second input picture is obtained by the first linear relationship and exports the second linear relationship between image, output image is obtained by the second linear relationship and the second input picture.This method can the existing housebroken network frame of seamless interfacing, have versatility, have high speed-up ratio while not influencing the quality of image procossing.
Description
Technical field
This application involves image processing technologies, more particularly to a kind of image processing method based on full convolutional network
Method, device and computer equipment.
Background technique
Depth learning technology promotes the development of many image procossings, such as image recognition, target detection etc., is based on depth
The image procossings such as image super-resolution rebuilding, defogging, HDR, the details enhancing of habit technology achieve especially gratifying effect, still
Depth learning technology needs to handle more data, deeper network and then leads to slower processing speed；Due to deep learning
Processing speed it is slow, cause its application to have significant limitation, such as mostly use traditional figure in the image procossing of mobile phone terminal
As processing technique, effect is obviously inferior to the image procossing based on depth learning technology.
Therefore, the prior art has much room for improvement.
Summary of the invention
Based on this, it is necessary in view of the above technical problems, provide one kind and be improved deeply while guaranteeing quality of image processing
Spend the image processing method based on full convolutional network, device and the computer equipment of learning efficiency.
A kind of image processing method based on full convolutional network, which comprises
A, the resolution ratio for reducing the first input picture, obtains the first low resolution input picture；
B, the first low resolution input picture is input in full convolutional network, obtains low resolution output image；
C, first input picture and the first low resolution input picture are converted to respectively defeated with the low resolution
Identical second input picture of image channel and the second low resolution input picture out；
D, using top sampling method, processing figure is obtained by the second low resolution input picture and low resolution output image
The first linear relationship between the processing image and the second low resolution input picture is calculated in picture；
E, the second input picture is obtained by the first linear relationship and exports the second linear relationship between image, pass through the second line
Sexual intercourse and the second input picture obtain output image.
Optionally, the step C includes:
First input picture is input in the guidance network being made of convolutional layer, is obtained and the low resolution output figure
As identical second input picture in channel；
The first low resolution input picture is input in the guidance network being made of convolutional layer, is obtained and the low resolution
Rate exports the identical second resolution input picture of image channel.
Second input picture and low resolution output image are local near line sexual intercourse；The second low resolution
Rate input picture and low resolution output image are local near line sexual intercourse.
Optionally, the second input picture obtained by the first linear relationship in the step E and export between image the
The method of two wires sexual intercourse includes:
The parameter in first linear relationship is amplified to obtain the parameter of the second linear relationship using bilinear interpolation；
Wherein, the parameter of the second linear relationship is corresponding with the parameter of the first linear relationship, the second input of the second linear relationship
Image is corresponding with the second low resolution input picture of the first linear relationship；The output image and First Line of second linear relationship
The processing image of sexual intercourse is corresponding.
Optionally, the step A includes:
The resolution ratio of the high-definition picture is reduced using Image Zooming Algorithm.
Optionally, the joint up-sampling system is guiding filtering or bilateral filtering.
A kind of image processing apparatus based on full convolutional network, described device include:
It reduces module resolution and obtains the first low resolution input picture for reducing the resolution ratio of the first input picture；
Full convolutional network module obtains low point for the first low resolution input picture to be input in full convolutional network
Resolution exports image；
Channel conversion module, for respectively by first input picture and the first low resolution input picture be converted to
The low resolution exports identical second input picture of image channel and the second low resolution input picture；
Low-resolution image processing module, for utilizing top sampling method, by the second low resolution input picture and low point
Resolution output image obtains processing image, is calculated between the processing image and the second low resolution input picture
First linear relationship；
Image output module, it is second linear between image for obtaining the second input picture by the first linear relationship and exporting
Relationship obtains output image by the second linear relationship and the second input picture.
Optionally, described image output module further include: the second linear relationship generation unit；
The second linear relationship generation unit, for using bilinear interpolation to the parameter in first linear relationship into
Row amplification obtains the parameter of the second linear relationship；
Wherein, the parameter of the second linear relationship is corresponding with the parameter of the first linear relationship, the second input of the second linear relationship
Image is corresponding with the second low resolution input picture of the first linear relationship；The output image and First Line of second linear relationship
The processing image of sexual intercourse is corresponding.
A kind of computer equipment, including memory and processor, the memory are stored with computer program, the processing
Device performs the steps of when executing the computer program
A, the resolution ratio for reducing the first input picture, obtains the first low resolution input picture；
B, the first low resolution input picture is input in full convolutional network, obtains low resolution output image；
C, first input picture and the first low resolution input picture are converted to respectively defeated with the low resolution
Identical second input picture of image channel and the second low resolution input picture out；
D, using top sampling method, processing figure is obtained by the second low resolution input picture and low resolution output image
The first linear relationship between the processing image and the second low resolution input picture is calculated in picture；
E, the second input picture is obtained by the first linear relationship and exports the second linear relationship between image, pass through the second line
Sexual intercourse and the second input picture obtain output image.
A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor
It is performed the steps of when row
A, the resolution ratio for reducing the first input picture, obtains the first low resolution input picture；
B, the first low resolution input picture is input in full convolutional network, obtains low resolution output image；
C, first input picture and the first low resolution input picture are converted to respectively defeated with the low resolution
Identical second input picture of image channel and the second low resolution input picture out；
D, using top sampling method, processing figure is obtained by the second low resolution input picture and low resolution output image
The first linear relationship between the processing image and the second low resolution input picture is calculated in picture；
E, the second input picture is obtained by the first linear relationship and exports the second linear relationship between image, pass through the second line
Sexual intercourse and the second input picture obtain output image.
A kind of above-mentioned image processing method based on full convolutional network, device and computer equipment, which comprises drop
The resolution ratio of low first input picture obtains the first low resolution input picture and is input in full convolutional network, obtains low point
Resolution exports image；The first input picture and the first low resolution input picture are converted into the second input picture and second respectively
Low resolution input picture；Using top sampling method, obtained by the second low resolution input picture and low resolution output image
Image is handled, the first linear relationship between processing image and the second low resolution input picture is calculated；It is linear by first
Relationship obtains the second input picture and exports the second linear relationship between image, is schemed by the second linear relationship and the second input
As obtaining output image.This method can the existing housebroken network frame of seamless interfacing, have versatility, do not influencing image
There is high speed-up ratio while the quality of processing, existing image processing network can also be integrated again, obtained therewith after training
Suitable image processing effect.
Detailed description of the invention
Fig. 1 is a kind of flow diagram of the image processing method based on full convolutional network of the present invention；
Fig. 2 is a kind of functional block diagram of the image processing method based on full convolutional network of the present invention；
Fig. 3 is the schematic diagram of one embodiment of the invention；
Fig. 4 is the structural block diagram of kind of the image processing method based on full convolutional network of the invention；
Fig. 5 is the internal structure chart of computer equipment in one embodiment of the invention.
Specific embodiment
It is with reference to the accompanying drawings and embodiments, right in order to which the objects, technical solutions and advantages of the application are more clearly understood
The application is further elaborated.It should be appreciated that specific embodiment described herein is only used to explain the application, not
For limiting the application.
Please refer to Fig. 1 and Fig. 2, a kind of image processing method based on full convolutional network, which comprises
S1, the resolution ratio for reducing the first input picture, obtain the first low resolution input picture.
Untreated image is the image that RAW image lens sensors directly acquire, and the RAW picture got is four-way
Picture, it is described relative to the first low resolution input picture after reduction resolution ratio using RAW image as the first input picture
First input picture is high-definition picture, if directly to handle high-definition picture data to be treated various for deep learning,
Meeting is time-consuming for a long time, and therefore, this method first reduces the resolution ratio of the input picture.
Optionally, the step of reducing the resolution ratio of first input picture in the step S1 includes:
The resolution ratio of first input picture is reduced using Image Zooming Algorithm.
Optionally, described image scaling algorithm includes: closest interpolation algorithm, bilinear interpolation, bicubic interpolation calculation
Method.
Preferably, the resolution ratio that first input picture is reduced using bicubic interpolation algorithm, is calculated with bicubic interpolation
The first low resolution input picture that method obtains is available more preferably compared to other two kinds of image scaling interpolation algorithm details
Image.
Assuming that the size of the first input picture g1 is 640 × 640, the first low resolution input picture g2's after scaling is big
It is small be 320 × 320, g1 each pixel it is unknown, g2 is unknown；The value of each pixel (x2, y2) in g2 is obtained, is needed
The each pixel for first finding g2 corresponds to the pixel (x1, y1) of g1, by 16 nearest pixels of distance (x1, y1) as meter
The pixel value parameter at (x2, the y2) of g2 is calculated, the weight of 16 pixels is found out by function, the value of (x2, y2) is equal to 16
The weighted superposition of a pixel.
S2, the first low resolution input picture is input in full convolutional network, obtains low resolution output image；
Optionally, the full convolutional network includes: that Super-resolution reconstruction establishing network, image denoising, image defogging, image remove rain, thin
Existing image processing network, the full convolutional network such as section enhancing network and night scene enhancing constitute end-to-end of image to image
It practises, these networks only include the standard module of Conv, without connecting class layer entirely；Application of the invention can be integrated with seamless interfacing
Any network structure trained has extremely strong versatility, is not limited only to several image processing networks listed.
Select one of full convolutional network or a variety of image processing networks to the first low resolution input figure as needed
As being handled, for example, it is desired to carry out super-resolution processing to the first low resolution input picture, then full convolution net is selected
First low resolution input picture is input to Super-resolution reconstruction establishing network, obtains oversubscription by the Super-resolution reconstruction establishing network in network
Low resolution afterwards exports image；For example, it is desired to be carried out at rain processing and denoising to the first low resolution input picture
Reason, then be input to image denoising network for the first low resolution input picture, the image after obtaining denoising, then will be after denoising
Image be input to image and remove rain network, go the output layer of rain network is available by denoising network and to remove rain network in image
Low resolution export image.
Since full convolutional network is handled the first low resolution input picture, compared with directly handling original image, place
The time-consuming of reason low resolution input picture greatly reduces.
S3, first input picture and the first low resolution input picture be converted to and the low resolution respectively
Rate exports identical second input picture of image channel and the second low resolution input picture.
In step sl, the first low resolution input picture is obtained by the first input picture.First low resolution is defeated
Entering image and the first input picture is all RAW image, and port number is identical, port number 4；In step s 2, described first low point
Resolution input picture is handled to obtain low resolution output image by full convolutional network, and the low resolution output image is RGB tri-
Channel image.
Therefore, it is necessary to the first low resolution input picture and first input picture are converted to RGB triple channel
Image can be realized using traditional channel conversion method, such as realized by python, realized by color interpolation method.
It may be implemented first input picture being converted to the second of RGB triple channel using existing channel conversion method
The first low resolution input picture is converted to the second low resolution input picture of RGB triple channel by input picture.
Preferably, the step S3 includes:
First input picture is input in the guidance network being made of convolutional layer, is obtained and the low resolution output figure
As identical second input picture in channel；
The first low resolution input picture is input in the guidance network being made of convolutional layer, is obtained and the low resolution
Rate exports the identical second resolution input picture of image channel.
The application provides a kind of guidance network different from conventional channels conversion method as channel switching network, described to draw
Wire guide network includes the part convolutional layer in convolutional network, and the guidance network can guarantee the input and output of full convolutional network
Between port number it is identical, guidance network can denoise RAW image, demosaicing and feature extraction, provide preferably
Feature, so that output and input image meets local near line sexual intercourse as far as possible.
Guidance network is made of convolutional layer, needs in advance to carry out guidance network using frame proposed by the present invention end-to-end
Training, training after could use, likewise, the untrained net of end-to-end training may be implemented in method provided by the invention
Network, then integrated use, further increase quality of image processing.
Such as in going rain application, the first low resolution input picture and low resolution output image are not able to satisfy near-linear
Relationship, carries out joint training end to end to guidance network, and the guidance network after addition training guides the extracted feature of network
It is exported with the low resolution obtained by full convolutional network between image and has better near line sexual intercourse；First input picture warp
The guidance available RGB triple channel image after denoising, demosaicing and feature extraction of network is crossed to input to get to second
Image.
S4, using top sampling method, obtained everywhere by the second low resolution input picture and low resolution output image
Image is managed, the first linear relationship between the processing image and the second low resolution input picture is calculated.
Optionally, the top sampling method includes: guiding filtering or bilateral filtering.
Guiding filtering is a kind of image filtering technology, is schemed by a guidance, is filtered to target figure, so that most
Output image afterwards is generally similar to target image, but texture part is similar to navigational figure, obtained output image
There are linear relationships with target image.
In one embodiment, the low resolution output image is the first low resolution input picture by super-resolution
It is obtained after reason, guiding filtering assumes exist between the second low resolution input picture and low resolution output image in regional area
Near line sexual intercourse is schemed the second low resolution input picture as target image, low resolution output image as guidance, input
Processing image is obtained into guiding filtering, and the processing image and the first low resolution input picture are calculated by guiding filtering
Between the first linear relationship.
The method that least square can be used in first linear relationship is calculated, and sees formula (1):
(1)
O in formula1Represent processing image, I1The second low resolution input picture is represented, a, b are the parameters of linear relation, are being schemed
Parameter a and b as can directly calculate the first linear relationship in localized mass by least square method mode, each pixel
Specifically there is an a and b.
S5, the second input picture is obtained by the first linear relationship and exports the second linear relationship between image, pass through
Second linear relationship and the second input picture obtain output image；
The second input picture is obtained by the first linear relationship in the step S5 and exports the second linear relationship between image
Method include:
The parameter in first linear relationship is amplified to obtain the parameter of the second linear relationship using bilinear interpolation；
Wherein, the parameter of the second linear relationship is corresponding with the parameter of the first linear relationship, the second input of the second linear relationship
Image is corresponding with the second low resolution input picture of the first linear relationship；The output image and First Line of second linear relationship
The processing image of sexual intercourse is corresponding.
In the application of full convolutional network, there can be local near line sexual intercourse between the input and output of full convolutional network, it is right
Local the first input picture of near line sexual intercourse and output image between identical application, the input of the first low resolution and output
Between near line sexual intercourse be approximately uniform, therefore can by calculating this relationship in low-resolution image, directly will
This near-linear relation and function obtains output image to input picture.
There is an a and b by each pixel of the second low resolution input picture, corresponding second input picture
Each pixel has A and B, and the second linear relationship between available second input picture and output image can be expressed as public affairs
Formula (2)
(2)
In formula (2), O2Represent output image, I2It represents the second input picture and formula (2) is applied to input picture, can obtain
Output image after to superresolution processing.
In one embodiment, referring to Fig. 3, using a kind of image processing method based on full convolutional network to untreated figure
Handle as carrying out rain, to the raw image of camera lens sensor 640 × 640 sizes obtained, i.e. the first input picture 100 into
Row reduces resolution processes, obtains the first low resolution input picture 101, the first low resolution input picture 101 is input to
Rain network is removed, low resolution output image 201 is obtained；
First low resolution input picture 101 is input to guidance network and obtains the second low resolution input picture, by second low point
Resolution input picture and low resolution output image 201 input wave filter, obtain processing image, and processing image and the
Two low the first linear relationships differentiated between input picture；
First input picture 100 is input to guidance network, obtains the second input picture, obtains by first linear relationship
Two wires sexual intercourse, i.e. linear relationship between the second input picture and output image, second input picture is input to and is drawn
Waveguide filter obtains output image according to second input picture and second linear relationship.
By actual measurement, when input picture size is 640X640, by going rain network directly to handle input picture time-consuming
15.2784s, time-consuming using the present processes: 4.0742s, obtained imaging effect go rain network processes defeated with direct use
The imaging effect for entering image is suitable, it is intimate it is identical go rain quality under can obtain 3.75 speed-up ratio；
In others processing networks, speed-up ratio be will be different, for example, by go rain network directly handle size for 640 ×
640 input picture time-consuming 15.3972s, time-consuming using the present processes: 3.8844s, in intimate identical defogging quality
The speed-up ratio of lower acquirement is 3.96；By actual measurement, a kind of image processing method based on full convolutional network provided by the invention
Method can obtain nearly 4 times of speed-up ratio relative to conventional method.
Versatility of the invention is verified by integrating several different network applications, in one embodiment, works as integrated image
Super-resolution rebuilding application, the evaluation index PSNR and SSIM of image are (26.4101/ respectively after original EDSR network oversubscription
0.8168), using the proposed frame index of the present invention is (26.5781/0.8303) respectively, and visually available indifference
Specific effects.
In one embodiment, as integrated HDR in application, using frame provided by the present invention visually available nothing
Otherness effect.
In one embodiment, based on a kind of image processing method based on full convolutional network, the present invention also provides one
Image processing apparatus of the kind based on full convolutional network, referring to Fig. 4, described device includes:
It reduces module resolution 501 and obtains the first low resolution input picture for reducing the resolution ratio of the first input picture；
Full convolutional network module 502, for the first low resolution input picture to be input in full convolutional network, obtains low
Resolution output image；
Channel conversion module 503, for respectively converting first input picture and the first low resolution input picture
For the second input picture identical with low resolution output image channel and the second low resolution input picture；
Low-resolution image processing module 504, for utilizing top sampling method, by the second low resolution input picture and low
Resolution output image obtains processing image, is calculated between the processing image and the second low resolution input picture
The first linear relationship；
Image output module 505, for obtaining second between the second input picture and output image by the first linear relationship
Linear relationship obtains output image by the second linear relationship and the second input picture.
Optionally, the channel conversion module 503 further includes guidance network unit；
The guidance network unit is obtained for being input to first input picture in the guidance network being made of convolutional layer
To the second input picture identical with low resolution output image channel；
The first low resolution input picture is input in the guidance network being made of convolutional layer, is obtained and the low resolution
Rate exports the identical second resolution input picture of image channel.
Optionally, described image output module 505 further includes the second linear relationship generation unit；
The second linear relationship generation unit, for using bilinear interpolation to the parameter in first linear relationship into
Row amplification obtains the parameter of the second linear relationship；
Wherein, the parameter of the second linear relationship is corresponding with the parameter of the first linear relationship, the second input of the second linear relationship
Image is corresponding with the second low resolution input picture of the first linear relationship；The output image and First Line of second linear relationship
The processing image of sexual intercourse is corresponding.
In one embodiment, the present invention provides a kind of computer equipment, which can be terminal, and internal structure is such as
Shown in Fig. 5.The computer equipment includes processor, memory, network interface, display screen and the input connected by system bus
Device.Wherein, the processor of the computer equipment is for providing calculating and control ability.The memory of the computer equipment includes
Non-volatile memory medium, built-in storage.The non-volatile memory medium is stored with operating system and computer program.The memory
Reservoir provides environment for the operation of operating system and computer program in non-volatile memory medium.The net of the computer equipment
Network interface is used to communicate with external terminal by network connection.It is a kind of non-to realize when the computer program is executed by processor
The processing method of even grain.The display screen of the computer equipment can be liquid crystal display or electric ink display screen, should
The input unit of computer equipment can be the touch layer covered on display screen, be also possible to be arranged on computer equipment shell
Key, trace ball or Trackpad can also be external keyboard, Trackpad or mouse etc..
It will be understood by those skilled in the art that shown in fig. 5 is only the frame of part-structure relevant to application scheme
Figure, does not constitute the restriction for the computer equipment being applied thereon to application scheme, specific computer equipment can wrap
It includes than more or fewer components as shown in the figure, perhaps combines certain components or with different component layouts.
In one embodiment, a kind of computer equipment, including memory and processor, the memory storage are provided
There is computer program, the processor performs the steps of when executing the computer program
The resolution ratio for reducing the first input picture obtains the first low resolution input picture；
The first low resolution input picture is input in full convolutional network, low resolution output image is obtained；
First input picture and the first low resolution input picture are converted to respectively and exported with the low resolution
Identical second input picture of image channel and the second low resolution input picture；
Using top sampling method, processing image is obtained by the second low resolution input picture and low resolution output image,
The first linear relationship between the processing image and the second low resolution input picture is calculated；
The second input picture is obtained by the first linear relationship and exports the second linear relationship between image, it is linear by second
Relationship and the second input picture obtain output image.
In one embodiment, a kind of computer readable storage medium is provided, computer program is stored thereon with, it is described
It is performed the steps of when computer program is executed by processor
The resolution ratio for reducing the first input picture obtains the first low resolution input picture；
The first low resolution input picture is input in full convolutional network, low resolution output image is obtained；
First input picture and the first low resolution input picture are converted to respectively and exported with the low resolution
Identical second input picture of image channel and the second low resolution input picture；
Using top sampling method, processing image is obtained by the second low resolution input picture and low resolution output image,
The first linear relationship between the processing image and the second low resolution input picture is calculated；
The second input picture is obtained by the first linear relationship and exports the second linear relationship between image, it is linear by second
Relationship and the second input picture obtain output image.
Each technical characteristic of above embodiments can be combined arbitrarily, for simplicity of description, not to above-described embodiment
In each technical characteristic it is all possible combination be all described, as long as however, the combination of these technical characteristics be not present lance
Shield all should be considered as described in this specification.
A kind of above-mentioned image processing method based on full convolutional network, device and computer equipment, which comprises drop
The resolution ratio of low first input picture obtains the first low resolution input picture and is input in full convolutional network, obtains low point
Resolution exports image；The first input picture and the first low resolution input picture are converted into the second input picture and second respectively
Low resolution input picture；Using top sampling method, obtained by the second low resolution input picture and low resolution output image
Image is handled, the first linear relationship between processing image and the second low resolution input picture is calculated；It is linear by first
Relationship obtains the second input picture and exports the second linear relationship between image, is schemed by the second linear relationship and the second input
As obtaining output image.This method can the existing housebroken network frame of seamless interfacing, have versatility, do not influencing image
There is high speed-up ratio while the quality of processing, existing image processing network can also be integrated again, obtained therewith after training
Suitable image processing effect.
Claims (10)
1. a kind of image processing method based on full convolutional network, which is characterized in that the described method includes:
A, the resolution ratio for reducing the first input picture, obtains the first low resolution input picture；
B, the first low resolution input picture is input in full convolutional network, obtains low resolution output image；
C, first input picture and the first low resolution input picture are converted to respectively defeated with the low resolution
Identical second input picture of image channel and the second low resolution input picture out；
D, using top sampling method, processing figure is obtained by the second low resolution input picture and low resolution output image
The first linear relationship between the processing image and the second low resolution input picture is calculated in picture；
E, the second input picture is obtained by the first linear relationship and exports the second linear relationship between image, pass through the second line
Sexual intercourse and the second input picture obtain output image.
2. the method according to claim 1, wherein the step C includes:
First input picture is input in the guidance network being made of convolutional layer, is obtained and the low resolution output figure
As identical second input picture in channel；
The first low resolution input picture is input in the guidance network being made of convolutional layer, is obtained and the low resolution
Rate exports the identical second resolution input picture of image channel.
3. the method according to claim 1, wherein second input picture and low resolution output are schemed
As being local near line sexual intercourse；The second low resolution input picture and low resolution output image are local near-linear
Relationship.
4. the method according to claim 1, wherein obtaining second by the first linear relationship in the step E
Input picture and the method for exporting the second linear relationship between image include:
The parameter in first linear relationship is amplified to obtain the parameter of the second linear relationship using bilinear interpolation；
Wherein, the parameter of the second linear relationship is corresponding with the parameter of the first linear relationship, the second input of the second linear relationship
Image is corresponding with the second low resolution input picture of the first linear relationship；The output image and First Line of second linear relationship
The processing image of sexual intercourse is corresponding.
5. the method according to claim 1, wherein the step A includes:
The resolution ratio of first input picture is reduced using Image Zooming Algorithm.
6. the method according to claim 1, wherein the joint up-sampling system is guiding filtering or bilateral filter
Wave.
7. a kind of image processing apparatus based on full convolutional network, which is characterized in that described device includes:
It reduces module resolution and obtains the first low resolution input picture for reducing the resolution ratio of the first input picture；
Full convolutional network module obtains low point for the first low resolution input picture to be input in full convolutional network
Resolution exports image；
Channel conversion module, for respectively by first input picture and the first low resolution input picture be converted to
The low resolution exports identical second input picture of image channel and the second low resolution input picture；
Low-resolution image processing module, for utilizing top sampling method, by the second low resolution input picture and low point
Resolution output image obtains processing image, is calculated between the processing image and the second low resolution input picture
First linear relationship；
Image output module, it is second linear between image for obtaining the second input picture by the first linear relationship and exporting
Relationship obtains output image by the second linear relationship and the second input picture.
8. device according to claim 7, which is characterized in that described image output module further include: the second linear relationship
Generation unit；
The second linear relationship generation unit, for using bilinear interpolation to the parameter in first linear relationship into
Row amplification obtains the parameter of the second linear relationship；
Wherein, the parameter of the second linear relationship is corresponding with the parameter of the first linear relationship, the second input of the second linear relationship
Image is corresponding with the second low resolution input picture of the first linear relationship；The output image and First Line of second linear relationship
The processing image of sexual intercourse is corresponding.
9. a kind of computer equipment, including memory and processor, the memory are stored with computer program, feature exists
In the step of processor realizes any one of claims 1 to 6 the method when executing the computer program.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
The step of method described in any one of claims 1 to 6 is realized when being executed by processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910402733.0A CN110211057B (en) | 2019-05-15 | 2019-05-15 | Image processing method and device based on full convolution network and computer equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910402733.0A CN110211057B (en) | 2019-05-15 | 2019-05-15 | Image processing method and device based on full convolution network and computer equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110211057A true CN110211057A (en) | 2019-09-06 |
| CN110211057B CN110211057B (en) | 2023-08-29 |
Family
ID=67787254
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910402733.0A Active CN110211057B (en) | 2019-05-15 | 2019-05-15 | Image processing method and device based on full convolution network and computer equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110211057B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110766610A (en) * | 2019-10-28 | 2020-02-07 | 维沃移动通信有限公司 | Super-resolution image reconstruction method and electronic equipment |
| CN111709879A (en) * | 2020-06-17 | 2020-09-25 | Oppo广东移动通信有限公司 | Image processing method, image processing device and terminal equipment |
| CN112488916A (en) * | 2019-09-12 | 2021-03-12 | 武汉Tcl集团工业研究院有限公司 | Training method of image hyper-resolution reconstruction model and computer equipment |
| CN113744167A (en) * | 2021-09-02 | 2021-12-03 | 厦门美图之家科技有限公司 | Image data conversion method and device |
| CN113766090A (en) * | 2020-06-02 | 2021-12-07 | 武汉Tcl集团工业研究院有限公司 | Image processing method, terminal and storage medium |
| CN117097993A (en) * | 2023-10-20 | 2023-11-21 | 荣耀终端有限公司 | Image processing method and related device |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130028538A1 (en) * | 2011-07-29 | 2013-01-31 | Simske Steven J | Method and system for image upscaling |
| CN106228512A (en) * | 2016-07-19 | 2016-12-14 | 北京工业大学 | Based on learning rate adaptive convolutional neural networks image super-resolution rebuilding method |
| WO2018035805A1 (en) * | 2016-08-25 | 2018-03-01 | Intel Corporation | Coupled multi-task fully convolutional networks using multi-scale contextual information and hierarchical hyper-features for semantic image segmentation |
| CN108288270A (en) * | 2018-02-05 | 2018-07-17 | 南京邮电大学 | A kind of object detection method based on channel trimming and full convolution deep learning |
| CN108647775A (en) * | 2018-04-25 | 2018-10-12 | 陕西师范大学 | Super-resolution image reconstruction method based on full convolutional neural networks single image |
| US20190080456A1 (en) * | 2017-09-12 | 2019-03-14 | Shenzhen Keya Medical Technology Corporation | Method and system for performing segmentation of image having a sparsely distributed object |
-
2019
- 2019-05-15 CN CN201910402733.0A patent/CN110211057B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130028538A1 (en) * | 2011-07-29 | 2013-01-31 | Simske Steven J | Method and system for image upscaling |
| CN106228512A (en) * | 2016-07-19 | 2016-12-14 | 北京工业大学 | Based on learning rate adaptive convolutional neural networks image super-resolution rebuilding method |
| WO2018035805A1 (en) * | 2016-08-25 | 2018-03-01 | Intel Corporation | Coupled multi-task fully convolutional networks using multi-scale contextual information and hierarchical hyper-features for semantic image segmentation |
| US20190080456A1 (en) * | 2017-09-12 | 2019-03-14 | Shenzhen Keya Medical Technology Corporation | Method and system for performing segmentation of image having a sparsely distributed object |
| CN108288270A (en) * | 2018-02-05 | 2018-07-17 | 南京邮电大学 | A kind of object detection method based on channel trimming and full convolution deep learning |
| CN108647775A (en) * | 2018-04-25 | 2018-10-12 | 陕西师范大学 | Super-resolution image reconstruction method based on full convolutional neural networks single image |
Non-Patent Citations (2)
| Title |
|---|
| LIU RISHENG等: "Deep Layer Prior Optimization for Single Image Rain Streaks Removal" * |
| 马冬冬;李金宗;朱兵;穆立胜;: "并行图像复原与超分辨处理系统的设计与实现" * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112488916A (en) * | 2019-09-12 | 2021-03-12 | 武汉Tcl集团工业研究院有限公司 | Training method of image hyper-resolution reconstruction model and computer equipment |
| CN112488916B (en) * | 2019-09-12 | 2023-12-26 | 武汉Tcl集团工业研究院有限公司 | Training method and computer equipment for image super-resolution reconstruction model |
| CN110766610A (en) * | 2019-10-28 | 2020-02-07 | 维沃移动通信有限公司 | Super-resolution image reconstruction method and electronic equipment |
| CN113766090A (en) * | 2020-06-02 | 2021-12-07 | 武汉Tcl集团工业研究院有限公司 | Image processing method, terminal and storage medium |
| CN113766090B (en) * | 2020-06-02 | 2023-08-01 | 武汉Tcl集团工业研究院有限公司 | Image processing method, terminal and storage medium |
| CN111709879A (en) * | 2020-06-17 | 2020-09-25 | Oppo广东移动通信有限公司 | Image processing method, image processing device and terminal equipment |
| CN111709879B (en) * | 2020-06-17 | 2023-05-26 | Oppo广东移动通信有限公司 | Image processing method, image processing device and terminal equipment |
| CN113744167A (en) * | 2021-09-02 | 2021-12-03 | 厦门美图之家科技有限公司 | Image data conversion method and device |
| CN113744167B (en) * | 2021-09-02 | 2024-04-30 | 厦门美图之家科技有限公司 | Image data conversion method and device |
| CN117097993A (en) * | 2023-10-20 | 2023-11-21 | 荣耀终端有限公司 | Image processing method and related device |
| CN117097993B (en) * | 2023-10-20 | 2024-05-28 | 荣耀终端有限公司 | Image processing method and related device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110211057B (en) | 2023-08-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110211057A (en) | A kind of image processing method based on full convolutional network, device and computer equipment | |
| CN110136066B (en) | Video-oriented super-resolution method, device, equipment and storage medium | |
| CN110996170B (en) | Video file playing method and related equipment | |
| US20050248590A1 (en) | Method and system for scaling, filtering, scan conversion, panoramic scaling, YC adjustment, and color conversion in a display controller | |
| CN107169927B (en) | Image processing system, method and display device | |
| CN108109109B (en) | Super-resolution image reconstruction method, device, medium and computing equipment | |
| WO2020253103A1 (en) | Video image processing method, device, apparatus, and storage medium | |
| CN113850367B (en) | Network model training method, image processing method and related equipment thereof | |
| US20130243312A1 (en) | Color distance measurement apparatus, color distance measurement method, and program | |
| CN108242063B (en) | Light field image depth estimation method based on GPU acceleration | |
| CN110992265A (en) | Image processing method and model, model training method and electronic equipment | |
| CN112541868B (en) | Image processing method, device, computer equipment and storage medium | |
| US20080285868A1 (en) | Simple Adaptive Wavelet Thresholding | |
| CN110889809B9 (en) | Image processing method and device, electronic equipment and storage medium | |
| JPWO2013005727A1 (en) | Imaging apparatus and program | |
| WO2004077357A1 (en) | Apparatus and method for producing thumbnail images and for improving image quality of re-sized images | |
| WO2019090580A1 (en) | System and method for image dynamic range adjusting | |
| CN107220934B (en) | Image reconstruction method and device | |
| KR100852752B1 (en) | Method and apparatus for downscaling a digital matrix image | |
| CN110602410A (en) | Image processing method and device, aerial camera and storage medium | |
| KR102213765B1 (en) | An Image Sensor, A Camera Module And Optical Device Comprising A Camera Module | |
| CN114298889A (en) | Image processing circuit and image processing method | |
| US20230070778A1 (en) | Image processing circuit and image processing method | |
| CN112862717B (en) | Image denoising and blurring method, system and medium | |
| CN109685824B (en) | Motion judgment method and device based on singular value decomposition characteristics and electronic equipment |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |