CN107610192A - Adaptive observation compressed sensing image reconstructing method based on deep learning - Google Patents
Adaptive observation compressed sensing image reconstructing method based on deep learning Download PDFInfo
- Publication number
- CN107610192A CN107610192A CN201710923137.8A CN201710923137A CN107610192A CN 107610192 A CN107610192 A CN 107610192A CN 201710923137 A CN201710923137 A CN 201710923137A CN 107610192 A CN107610192 A CN 107610192A
- Authority
- CN
- China
- Prior art keywords
- network
- full articulamentum
- observation
- reconstructed
- file
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of adaptive observation compressed sensing image reconstructing method based on deep learning, mainly solves the problems, such as the observation that prior art can not obtain being adaptive to data set.Its implementation is：1. prepare reconstructed network DR2 network structures file and associated documents；2. addition output dimension obtains adaptive observation network less than the second full articulamentum of input on the basis of reconstructed network DR2 network structures, file needed for modification training, and adaptive observation network is trained using amended file, the model trained；3. carry out image observation and reconstruct using the model trained.The reconstruction result of the present invention is substantially better than the reconstruction result of existing random Gaussian observation, and strong adaptability, and real-time is good, available for radar imagery.
Description
Technical field
The invention belongs to technical field of image processing, relates generally to a kind of reconstruct side of adaptive observation compressed sensing image
Method, available for radar imagery.
Background technology
In information age now, people are to the demand of information also more and more higher.After the analog signal sampling of real world
It is necessary in this day and age to be converted into data signal and handled, and the theoretical foundation sampled is then that nyquist sampling is determined
Reason, it shows in the transfer process of analog signal and data signal is carried out to intactly retain the information in primary signal,
In sampling, sample frequency is at least needed as twice of signal highest frequency.But with information requirement more and more higher, broader bandwidth
The processing of signal becomes difficulty.In many application scenarios, too high sample rate causes people to have to sampled result
Stored and transmitted again after compression.
Compressive sensing theory shows to sample signal with low rate under certain condition, high probability reconstruct, big so as to save
Measure resource.A large amount of restructing algorithms are suggested in recent years, and such as orthogonal matching pursuit, base are followed the trail of, but these algorithms are excellent by solving
The characteristics of change problem is reconstructed makes the real-time that its difficulty has had.It is preferable that the characteristic trained under deep learning method line allows it to have
Real-time, therefore many image reconstructing methods based on deep learning are suggested, such as SDA, DR2, wherein:
SDA is only made up of full articulamentum, more parameters is needed when input data increases, so as to cause bigger calculating
Amount.
DR2 is made up of full articulamentum and convolutional layer, can be carried out piecemeal reconstruct, be reduced amount of calculation, it is not easy to over-fitting,
What DR2 was used skips formula connection so that quality reconstruction is more preferable；But because the DR2 observations used are random Gaussian observation, it is not easy to carry
Useful information of the access in so that the Y-PSNR PSNR of reconstructed image is low, influences the definition of reconstructed image.
The content of the invention
It is an object of the invention to the deficiency for above-mentioned existing compressed sensing Image Reconstruction deep learning method, proposes
A kind of compressed sensing Image Reconstruction deep learning method of adaptive observation, preferably to extract the useful information in data, is carried
The Y-PSNR PSNR of high reconstructed image, and then improve reconstructed image quality.
To achieve the above object, technical scheme includes as follows：
1) adaptive observation network is built：
Reconstructed network DR2 network structure file, parameter setting file, generation training 1a) is carried above and below github websites
Code, code, test code and the test set picture of generation checking collection of collection, download training set picture from SRCNN websites
Collect picture with checking；
Caffe platforms 1b) are built under Linux system, and matlab business softwares are installed；
Reconstructed network structured file, the i.e. full connection in reconstructed network DR2 1c) are changed on the basis of reconstructed network DR2
Addition output is the second full articulamentum of low-dimensional before layer, obtains adaptive observation network；
2) change data form starts to train：
2a) by 1a) in training set picture and checking collection picture be put into respectively Training_Data/Train files and
Under Training_Data/Test files, change and operation code generates the hdf5 forms of training set and the hdf5 lattice of checking collection
Formula file；
2b) use 1c) amended network structure file and parameter setting file and 2a) in training set hdf5 lattice
The hdf5 formatted files of formula and checking collection
2c) adaptive observation network is trained on caffe platforms：
2c1) only train reconstructed network DR2 full articulamentum and the second full articulamentum newly added；
2c2) in 2c1) on the basis of the second full articulamentum for training reconstructed network DR2 all layers and newly adding, that is, instruct
Practice adaptive observation network, the model trained, the model trained includes the second full articulamentum and reconstructed network
DR2；
3) image observation and reconstruct are carried out using the model trained：
3a) from 1a) in test set picture in choose an image as artwork, be divided into fixed-size several images
Block, the function in caffe is called by matlab business softwares, these image blocks of artwork are input to the model trained
In, each image block obtains a group observations after the second full articulamentum；
Reconstructed image block 3b) is obtained after reconstructed network DR2 per group observations, these reconstructed image blocks are suitable by artwork
Sequence merges, and obtains reconstructed image.
The present invention has advantages below compared with prior art：
1. strong adaptability：
Traditional random Gaussian observation is in order to which high probability meets RIP conditions, but limitation itself is very big, because it is not
Design to obtain according to training set；The observation that the present invention uses is obtained according to training set by training, available for different type
Training set, strong adaptability；
2. quality reconstruction is good：
The adaptive observation that the present invention uses can observe more information from artwork, and observation obtains after reconstruct
Reconstructed image Y-PSNR PSNR be obviously improved；
3. real-time is good：
The present invention reduces test and taken as a result of being deep learning method, has good reality in actual use
Shi Xing.
Brief description of the drawings
Fig. 1 is the implementation process figure of the present invention；
Fig. 2 is the adaptive observation network structure in the present invention；
Fig. 3 is artwork used in adaptive observation network in the test present invention；
Fig. 4 be with of the invention and existing random Gaussian observation procedure respectively when observation rate is 25% to Fig. 3 reconstruct image
Picture and reconstructed image Y-PSNR PSNR.
Embodiment
The present invention is described in detail with example below in conjunction with the accompanying drawings.
Reference picture 1, step is as follows for of the invention realizing：
Step 1, reconstructed network DR2 network structures file and associated documents are prepared.
Reconstructed network DR2 relevant document 1a) is carried above and below github websites, it includes：Network structure text needed for training
Part DR2_stage1.prototxt, DR2_stage2.prototxt, parameter setting file DR2_stage1_solver.proto
Txt and DR2_stage2_solver.prototxt, code file store2hdf5.m, life needed for making hdf5 formatted files
Into code file generate_train.m, generate_test.m needed for training set and checking collection, test code file
Network structure file reconnet_0_10.prototxt, reconnet_0_ needed for test_everything.m, test
25.prototxt and test set picture；
Training set picture and checking collection picture 1b) are downloaded from SRCNN websites；
Caffe platforms 1c) are built under Linux system, and matlab business softwares are installed.
Step 2, file needed for modification training.
2a) modification obtains adaptive observation network structure, such as Fig. 2 on the basis of reconstructed network DR2 network structure files
Shown, step is as follows：
2a1) open preserve network structure file DR2_stage2.prototxt, this document include full articulamentum fc1,
Reshape layers, 4 identical residual error study blocks, first residual error study block include the first convolutional layer conv1, volume Two
Lamination conv2, the 3rd convolutional layer conv3, the first BN layers, the first scale layers, the first relu layers, second residual error learn block bag
Include Volume Four lamination conv4, the 5th convolutional layer conv5, the 6th convolutional layer conv6, the 2nd BN layers, the 2nd scale layers, second
Relu layers, the 3rd residual error study block include the 7th convolutional layer conv7, the 8th convolutional layer conv8, the 9th convolutional layer conv9, the
Three BN layers, the 3rd scale layers, the 3rd relu layers, the 4th residual error study block include the tenth convolutional layer conv10, the 11st convolution
Layer conv11, the 12nd convolutional layer conv12, the of one layer of output dimension less than input dimension is added before full articulamentum fc1
Two full articulamentums, i.e., add a segment table in DR2_stage1.prototxt files according to specified format and show the second full articulamentum
Word；In the DR2_stage1.prototxt network structure files each layer all comprising layer title name and this layer it is previous
The title bottom of layer；
2a2) name of the second full articulamentum is changed to " fc1 ", the name of full articulamentum is changed to " fc2 ", and bottom is changed to
" fc1 ", the bottom of reshape layers are changed to " fc2 ", due to the preceding layer that full articulamentum is reshape layers, so reshape
The bottom of layer is also " fc2 "；
2a3) according to the dimension of input picture block and observation rate, the output dimension of the second full articulamentum of calculating：
It is less than requirement and the practicality of input dimension according to output dimension, the dimension that this example fixes input picture block is
1089, observation rate of the selection less than 50%；Both is multiplied again, obtains the output dimension of the second full articulamentum；Such as observe rate
For 25% when, output image block dimension is 1089x 25%=272；
2b) open preserve network structure file DR2_stage1.prototxt, this document include full articulamentum fc1,
Reshape layers, according to 2a) identical mode adds the second full articulamentum；
2c) the code file generate_train.m needed for generation training set and checking collection is opened using matlab softwares
And generate_test.m, modify, that is, delete the observation process code in two files；
Step 3, change data form starts to train.
3a) by 1b) in training set picture be put under the Training_Data/Train of path, will checking collection picture be put into
Under the Training_Data/Test of path, 2c is separately operable by matlab softwares) in amended code file generate_
The hdf5 forms of train.m, generate_test.m generation training set and the hdf5 formatted files of checking collection；
3b) adaptive observation network is trained on caffe platforms：
3b1) only train reconstructed network DR2 full articulamentum and the second full articulamentum newly added：
Linux system terminal is opened under the caffe_dr2-master/DR2/train of path, starts to instruct to terminal input
Experienced order ../../../build/tools/caffe train--solver=DR2_stage1_solver.prototxt,
The model 1 trained；
3b2) in 3a1) on the basis of train reconstructed network DR2 and the second full articulamentum for newly adding, i.e., training is adaptively
Observation grid：
Linux system terminal is opened under the caffe_dr2-master/DR2/train of path, starts to instruct to terminal input
Experienced order ../../../build/tools/caffe train--solver=DR2_stage2_
Solver.prototxt--weights=model/DR2_stage1_iter_1000000.c affemodel, are trained
Adaptive observation network model, wherein -- weights=model/DR2_stage1_iter_1000000.caffemodel
The full articulamentum and the second full articulamentum for representing adaptive observation network are trained on the basis of the model 1 trained.
Step 4, image observation and reconstruct are carried out using the model trained.
4a) from step 1a1) in test set picture in pick out a pictures as artwork, as shown in Figure 3；
4b) according to 2a) identical mode under path/test/deploy_prototxt_files/ file carry out
Modification, different file is changed according to the different choice of observation rate, when such as observation rate being 10%, change reconnet_0_10.p
Rototxt files, when observation rate is 25%, change reconnet_0_25.prototxt；
Matlab softwares are opened under the caffe_dr2-master/DR2/ of path, path is opened in matlab softwares
Test_everything.m code files under test/ are simultaneously modified to the code file, and the code file is run after modification
Complete test；
After the completion of 4c) testing, the reconstructed image of artwork and the Y-PSNR PSNR of reconstructed image are exported, such as Fig. 4 (a) institutes
Show；
4d) input order net.blobs (' fc1 ') .get_data () in matlab software command windows, in the order
Fc1 represent the second full articulamentum, the effect of the order is the observation of the output, i.e. original image of taking out the second full articulamentum.
The effect of the present invention can further illustrate that the emulation experiment step of existing method is such as by the comparison with existing method
Under：
The first step, prepare reconstructed network DR2 network structures file and associated documents.
1.1) reconstructed network DR2 relevant document is carried above and below github websites, it includes：Network structure text needed for training
Part DR2_stage1.prototxt, DR2_stage2.prototxt, parameter setting file DR2_stage1_solver.pro
Totxt and DR2_stage2_solver.prototxt, make hdf5 formatted files needed for code file store2hdf5.m,
Generate training set and verify the required code file generate_train.m, generate_test.m of collection, test code file
Network structure file reconnet_0_10.prototxt, reconnet_0_ needed for test_everything.m, test
25.prototxt and test set picture；
1.2) training set picture and checking collection picture are downloaded from SRCNN websites；
1.3) caffe platforms are built under Linux system, and matlab business softwares are installed.
Second step, change data form start to train.
Change data form starts to train.
2.1) the training set picture in 1.2) is put under the Training_Data/Train of path, checking collection picture is put
To under the Training_Data/Test of path, the code file generate_t in being separately operable 1.1) by matlab softwares
The hdf5 forms of rain.m, generate_test.m generation training set and the hdf5 formatted files of checking collection；
2.2) reconstructed network DR2 is trained on caffe platforms：
2.2.1 reconstructed network DR2 full articulamentum) is only trained：
Linux system terminal is opened under the caffe_dr2-master/DR2/train of path, starts to instruct to terminal input
Experienced order ../../../build/tools/caffe train--solver=DR2_stage1_solver.prototxt,
The model 1 trained；
2.2.2 2.2.1)) on the basis of train reconstructed network DR2：
Linux system terminal is opened under the caffe_dr2-master/DR2/train/ of path, starts to instruct to terminal input
Experienced order ../../../build/tools/caffe train--solver=DR2_stage2_
Solver.prototxt--weights=model/DR2_stage1_iter_1000000.c affemodel, are trained
Reconstructed network DR2 model, wherein -- weights=model/DR2_stage1_iter_1000000.caffemodel tables
Show that reconstructed network DR2 full articulamentum is trained on the basis of model 1.
3rd step, image observation and reconstruct are carried out using the model trained.
3.1) pictures are picked out in the test set picture in step 1.1) as artwork, as shown in Figure 3；
3.2) matlab softwares are opened under the caffe_dr2-master/DR2/ of path, road is opened in matlab softwares
Test_everything.m code files under the test/ of footpath, the code file contain the random Gaussian observation process to artwork
With the restructuring procedure by reconstructed network DR2, run the code file and complete test, export the reconstructed image and reconstruct image of artwork
The Y-PSNR PSNR of picture, as shown in Fig. 4 (b)；
It will be compared, can seen with reconstruction result Fig. 4 (b) that existing method obtains with reconstruction result Fig. 4 (a) of the present invention
Go out：
Reconstruction result of the reconstruction result of adaptive observation network proposed by the present invention substantially than random Gaussian observation is clear；
The characteristic that the present invention is adaptive to data set can preferably extract useful information in data, improve reconstructed image
Y-PSNR PSNR, and then improve reconstructed image quality, while also have on the basis of higher reconstruction quality is ensured
The characteristics of real-time is good.
Claims (4)
1. the adaptive observation compressed sensing image reconstructing method based on deep learning, including：
1) adaptive observation network is built：
Reconstructed network DR2 network structure file, parameter setting file, generation training set 1a) are carried above and below github websites
Code, test code and the test set picture of code, generation checking collection, training set picture is downloaded from SRCNN websites and is tested
Card collection picture；
Caffe platforms 1b) are built under Linux system, and matlab business softwares are installed；
Reconstructed network structured file 1c) is changed on the basis of reconstructed network DR2, i.e., before reconstructed network DR2 full articulamentum
Addition output is the second full articulamentum of low-dimensional, obtains adaptive observation network；
2) change data form starts to train：
2a) by 1a) in training set picture and checking collection picture be put into respectively Training_Data/Train files and
Under Training_Data/Test files, change and operation code generates the hdf5 forms of training set and the hdf5 lattice of checking collection
Formula file；
2b) use 1c) amended network structure file and parameter setting file and 2a) in training set hdf5 forms and
Verify the hdf5 formatted files of collection
2c) adaptive observation network is trained on caffe platforms：
2c1) only train reconstructed network DR2 full articulamentum and the second full articulamentum newly added；
2c2) in 2c1) on the basis of the second full articulamentum for training reconstructed network DR2 all layers and newly adding, that is, train oneself
Observation grid, the model trained are adapted to, the model trained includes the second full articulamentum and reconstructed network DR2；
3) image observation and reconstruct are carried out using the model trained：
3a) from 1a) in test set picture in choose an image as artwork, be divided into fixed-size several image blocks,
Function in caffe is called by matlab business softwares, these image blocks of artwork are input in the model trained, often
Individual image block obtains a group observations after the second full articulamentum；
Reconstructed image block 3b) is obtained after reconstructed network DR2 per group observations, these reconstructed image blocks are entered by artwork order
Row merges, and obtains reconstructed image.
2. according to the method for claim 1, wherein step 1c) in add second on the basis of DR2 reconstructed network structures
Full articulamentum, adaptive observation network is obtained, carried out as follows：
Network structure file 1c1) is opened, this document includes full articulamentum fc1, reshape layer, 4 identical residual errors
Block is practised, first residual error study block includes the first convolutional layer conv1, the second convolutional layer conv2, the 3rd convolutional layer conv3, first
BN layers, the first scale layers, the first relu layers, second residual error study block include Volume Four lamination conv4, the 5th convolutional layer
Conv5, the 6th convolutional layer conv6, the 2nd BN layers, the 2nd scale layers, the 2nd relu layers, the 3rd residual error study block include the
Seven convolutional layer conv7, the 8th convolutional layer conv8, the 9th convolutional layer conv9, the 3rd BN layers, the 3rd scale layers, the 3rd relu
Layer, the 4th residual error study block include the tenth convolutional layer conv10, the 11st convolutional layer conv11, the 12nd convolutional layer
Conv12, one layer of output dimension is added before full articulamentum fc1 less than the second full articulamentum of input dimension, the network structure
Each layer all title name comprising the layer and title bottom of this layer of preceding layer in file；
1c2) name of the second full articulamentum is changed to " fc1 ", the name of full articulamentum is changed to " fc2 ", and bottom is changed to
" fc1 ", the bottom of reshape layers are changed to " fc2 ", i.e., full articulamentum.
3. according to the method for claim 2, wherein step 1c1) in the second full articulamentum output dimension, by input picture
The dimension and observation rate of block determine that the dimension of fixed input picture block is 1089, observation rate of the selection less than 50%；Again by this two
Person is multiplied, and obtains the output dimension of the second full articulamentum.
4. according to the method for claim 1, wherein step 2b) in use 1c) amended network structure file and parameter
Set file to be trained on caffe platforms to adaptive observation network, carry out as follows：
2b1) open Linux system terminal；
Catalogue 2b2) is switched to network structure file and parameter setting file position from current location in the terminal；
2b3) input starts training of the instruction completion of training to adaptive observation network in the terminal.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710923137.8A CN107610192B (en) | 2017-09-30 | 2017-09-30 | Self-adaptive observation compressed sensing image reconstruction method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710923137.8A CN107610192B (en) | 2017-09-30 | 2017-09-30 | Self-adaptive observation compressed sensing image reconstruction method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107610192A true CN107610192A (en) | 2018-01-19 |
| CN107610192B CN107610192B (en) | 2021-02-12 |
Family
ID=61067960
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710923137.8A Active CN107610192B (en) | 2017-09-30 | 2017-09-30 | Self-adaptive observation compressed sensing image reconstruction method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107610192B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108510464A (en) * | 2018-01-30 | 2018-09-07 | 西安电子科技大学 | Compressed sensing network and full figure reconstructing method based on piecemeal observation |
| CN108537104A (en) * | 2018-01-30 | 2018-09-14 | 西安电子科技大学 | Compressed sensing network based on full figure observation and perception loss reconstructing method |
| CN108810651A (en) * | 2018-05-09 | 2018-11-13 | 太原科技大学 | Wireless video method of multicasting based on depth-compression sensing network |
| CN109086819A (en) * | 2018-07-26 | 2018-12-25 | 北京京东尚科信息技术有限公司 | Caffemodel model compression method, system, equipment and medium |
| CN111192334A (en) * | 2020-01-02 | 2020-05-22 | 苏州大学 | Trainable compressed sensing module and image segmentation method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107123089A (en) * | 2017-04-24 | 2017-09-01 | 中国科学院遥感与数字地球研究所 | Remote sensing images super-resolution reconstruction method and system based on depth convolutional network |
| CN107123091A (en) * | 2017-04-26 | 2017-09-01 | 福建帝视信息科技有限公司 | A kind of near-infrared face image super-resolution reconstruction method based on deep learning |
| CN107154064A (en) * | 2017-05-04 | 2017-09-12 | 西安电子科技大学 | Natural image compressed sensing method for reconstructing based on depth sparse coding |
-
2017
- 2017-09-30 CN CN201710923137.8A patent/CN107610192B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107123089A (en) * | 2017-04-24 | 2017-09-01 | 中国科学院遥感与数字地球研究所 | Remote sensing images super-resolution reconstruction method and system based on depth convolutional network |
| CN107123091A (en) * | 2017-04-26 | 2017-09-01 | 福建帝视信息科技有限公司 | A kind of near-infrared face image super-resolution reconstruction method based on deep learning |
| CN107154064A (en) * | 2017-05-04 | 2017-09-12 | 西安电子科技大学 | Natural image compressed sensing method for reconstructing based on depth sparse coding |
Non-Patent Citations (2)
| Title |
|---|
| HANTAO YAO等: "DR2-Net：Deep Residual Reconstruction Network for Image Compressive Sensing", 《HTTPS://ARXIV.ORG/PDF/1702.05743V3.PDF》 * |
| XUEMEI XIE等: "Adaptive Measurement Network for CS Image Reconstruction", 《HTTPS://ARXIV.ORG/PDF/1710.01244V1.PDF》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108510464A (en) * | 2018-01-30 | 2018-09-07 | 西安电子科技大学 | Compressed sensing network and full figure reconstructing method based on piecemeal observation |
| CN108537104A (en) * | 2018-01-30 | 2018-09-14 | 西安电子科技大学 | Compressed sensing network based on full figure observation and perception loss reconstructing method |
| CN108510464B (en) * | 2018-01-30 | 2021-11-30 | 西安电子科技大学 | Compressed sensing network based on block observation and full-image reconstruction method |
| CN108810651A (en) * | 2018-05-09 | 2018-11-13 | 太原科技大学 | Wireless video method of multicasting based on depth-compression sensing network |
| CN108810651B (en) * | 2018-05-09 | 2020-11-03 | 太原科技大学 | Wireless video multicast method based on deep compression sensing network |
| CN109086819A (en) * | 2018-07-26 | 2018-12-25 | 北京京东尚科信息技术有限公司 | Caffemodel model compression method, system, equipment and medium |
| CN109086819B (en) * | 2018-07-26 | 2023-12-05 | 北京京东尚科信息技术有限公司 | Method, system, equipment and medium for compressing caffemul model |
| CN111192334A (en) * | 2020-01-02 | 2020-05-22 | 苏州大学 | Trainable compressed sensing module and image segmentation method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107610192B (en) | 2021-02-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107610192A (en) | Adaptive observation compressed sensing image reconstructing method based on deep learning | |
| CN109145730B (en) | Automatic semantic segmentation method for mining area in remote sensing image | |
| CN109410261B (en) | Monocular image depth estimation method based on pyramid pooling module | |
| CN107392019A (en) | A kind of training of malicious code family and detection method and device | |
| CN109508717A (en) | A kind of licence plate recognition method, identification device, identification equipment and readable storage medium storing program for executing | |
| CN110321874A (en) | A kind of light-weighted convolutional neural networks pedestrian recognition method | |
| CN112184554B (en) | Remote sensing image fusion method based on residual mixed expansion convolution | |
| CN109146788A (en) | Super-resolution image reconstruction method and device based on deep learning | |
| CN107562918A (en) | A kind of mathematical problem knowledge point discovery and batch label acquisition method | |
| DE112015007134T5 (en) | Method and device for synchronous image display | |
| CN108961220B (en) | Image collaborative saliency detection method based on multilayer convolution feature fusion | |
| CN107316054A (en) | Non-standard character recognition methods based on convolutional neural networks and SVMs | |
| CN110689012A (en) | End-to-end natural scene text recognition method and system | |
| CN107240136A (en) | A kind of Still Image Compression Methods based on deep learning model | |
| CN112017116A (en) | Image super-resolution reconstruction network based on asymmetric convolution and construction method thereof | |
| CN115100039A (en) | Lightweight image super-resolution reconstruction method based on deep learning | |
| CN111666114A (en) | Plug-in type well logging data conversion method | |
| CN103903240B (en) | Image super-resolution method based on multi output Least square support vector regression | |
| CN110399344A (en) | Choose the method and device of multiimage | |
| CN113839838A (en) | Business type identification method for federal learning based on cloud edge cooperation | |
| Fu et al. | Screen content image quality assessment using Euclidean distance | |
| CN115984949A (en) | Low-quality face image recognition method and device with attention mechanism | |
| CN115272667B (en) | Farmland image segmentation model training method and device, electronic equipment and medium | |
| CN105469116A (en) | Picture recognition and data extension method for infants based on man-machine interaction | |
| CN114973271A (en) | Text information extraction method, extraction system, electronic device and storage medium |
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 |