CN111091495A - High-resolution compressive sensing reconstruction method for laser image based on residual error network - Google Patents
High-resolution compressive sensing reconstruction method for laser image based on residual error network Download PDFInfo
- Publication number
- CN111091495A CN111091495A CN201910952609.1A CN201910952609A CN111091495A CN 111091495 A CN111091495 A CN 111091495A CN 201910952609 A CN201910952609 A CN 201910952609A CN 111091495 A CN111091495 A CN 111091495A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- residual
- image block
- laser
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4053—Super resolution, i.e. output image resolution higher than sensor resolution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4038—Scaling the whole image or part thereof for image mosaicing, i.e. plane images composed of plane sub-images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4046—Scaling the whole image or part thereof using neural networks
-
- H—ELECTRICITY
- H03—ELECTRONIC CIRCUITRY
- H03M—CODING; DECODING; CODE CONVERSION IN GENERAL
- H03M7/00—Conversion of a code where information is represented by a given sequence or number of digits to a code where the same, similar or subset of information is represented by a different sequence or number of digits
- H03M7/30—Compression; Expansion; Suppression of unnecessary data, e.g. redundancy reduction
- H03M7/3059—Digital compression and data reduction techniques where the original information is represented by a subset or similar information, e.g. lossy compression
- H03M7/3062—Compressive sampling or sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2200/00—Indexing scheme for image data processing or generation, in general
- G06T2200/32—Indexing scheme for image data processing or generation, in general involving image mosaicing
Abstract
The invention discloses a high-resolution compressed sensing reconstruction method for a laser image based on a residual error network, which comprises the steps of measuring a matrix W and a real laser image block xiMultiplying to obtain a compressed sensing measurement yi(ii) a The compressed sensing measure yiBy convex optimization based linear mapping network FfOutputting initial estimated image blocks of a fully connected layer output reconstructed imageThe initial estimated image blockBy means of a residual network Fr(. output prediction residual image blockFor the initial estimation image blockAnd pre-estimating residual image blockAnd reconstructing, and splicing all reconstructed image blocks to obtain a desired high-quality laser image reconstruction result. The invention properly increases the depth of the network through the jump connection, so that the network model is easier to optimize, the convergence of the training process is faster and more stable, and the reconstruction quality is better.
Description
Technical Field
The invention belongs to the field of compressed sensing laser image reconstruction, and particularly relates to a high-resolution compressed sensing reconstruction method for a laser image based on a residual error network.
Background
The compressed sensing theory utilizes the characteristic that signals have sparsity, and finally recovers the signals through discrete samples obtained by random sampling under the condition that the sampling rate is far less than the Nyquist sampling rate, so that the compressed sensing theory draws wide attention once put forward, and has great application value in the fields of communication, mode recognition, biomedicine, optical imaging, image processing and the like. The traditional reconstruction method based on compressed sensing comprises a greedy algorithm, a convex optimization algorithm, an iterative algorithm and the like, the method recovers images by solving an optimization problem, and the algorithm is serious in time consumption, low in reconstruction speed and low in reconstruction precision, so that the problems of low frame frequency, poor real-time performance and the like in practical application are caused.
Aiming at the problems of time consumption, low reconstruction accuracy and the like of the algorithm, with the continuous development of deep learning in recent years, in order to reduce the computational complexity of reconstruction, a Convolutional Neural Network (CNN) based method is also introduced into compressed sensing reconstruction, and the CNN based reconstruction method is initially inspired by a CNN super-resolution algorithm, and finally obtains a high-quality reconstruction result by obtaining initial rough estimation and performing subsequent operations such as super-resolution optimization and the like. However, in the current CNN-based compressed sensing reconstruction method, nonlinear mapping is performed through convolution operation when a one-dimensional sparse signal is mapped into a two-dimensional image, and the principle of compressed sensing is not considered; and the algorithm network structure and model parameters for super-resolution reconstruction are complex, the training process takes a long time, the reconstruction result is not ideal, and further improvement space exists.
Disclosure of Invention
In view of the above, the main object of the present invention is to provide a high-resolution compressive sensing reconstruction method for a laser image based on a residual error network.
In order to achieve the purpose, the technical scheme of the invention is realized as follows:
the embodiment of the invention provides a high-resolution compressive sensing reconstruction method for a laser image based on a residual error network, which comprises the following steps:
by measuring matrix W and real laser image block xiMultiplying to obtain a compressed sensing measurement yi;
The compressed sensing measure yiBy based on convex advantagesLinearized mapping network FfOutputting initial estimated image blocks of a fully connected layer output reconstructed image
The initial estimated image blockBy means of a residual network Fr(. output prediction residual image block
For the initial estimation image blockAnd pre-estimating residual image blockAnd reconstructing, and splicing all reconstructed image blocks to obtain a desired high-quality laser image reconstruction result.
In the above scheme, the real laser image x is processed in blocks to obtain a plurality of real laser image blocks xiMultiplying the real laser image blocks by a measurement matrix W to obtain a compressed sensing measurement yi。
In the above scheme, the compressed sensing measurement yiBy convex optimization based linear mapping network FfOutputting an initial estimated image block of a fully connected layer output reconstructed image blockThe method specifically comprises the following steps:
the mapping process to recover the real laser image x from the compressed sensing measurement value y is denoted as x-Wy, where W ∈ Rn×mIs a measurement matrix;
converting the mapping matrix into a solution mapping matrix WfTo the optimization problem ofWherein, yi∈Rn×1And xi∈Rn×1Respectively representing the compressed perceptual measurement and the corresponding original signal, { (y)1,x1),(y2,x2),…,(yN,xN) Denotes training data, X ═ X1,x2,…,xN],Y=[y1,y2,…,yN];
Obtaining a mapping matrix W using a loss functionf:Wherein, Ff(yi,Wf) Is yiInputting the output after passing through the linear mapping network;
compressed sensing measurement vector y corresponding to any original image blockiBy the mapping matrix WfThen all the image blocks can be obtainedThe mapping process is represented as:
in the above scheme, the method further comprises mapping the mapping matrix WfOptimizing, specifically: training an optimization network by a random gradient descent method until the network converges and obtaining an optimal mapping matrix Wf。
In the above scheme, the image blocks are initially estimatedAs a residual network FrObtaining estimated residual image block by inputtingWherein, WrAre parameters of the residual network.
In the above scheme, the initial estimation image block is processedAnd pre-estimating residual image blockReconstructing, and splicing all reconstructed image blocks to obtain a desired high-quality laser image reconstruction result, specifically: the initial estimation image blockAnd pre-estimating residual image blockAdding the two to obtain a corresponding laser image block reconstruction result:all reconstructed image blocksSplicing to obtain a complete laser reconstruction image x*。
In the above scheme, the network F is mapped through the linear mappingfThe parameters of (a) initialize the residual network, representing the final desired laser image reconstruction as:wherein the content of the first and second substances,
compared with the prior art, the method comprises the steps of firstly outputting an initial estimation result of a reconstructed image through a linear mapping network by utilizing convex optimization constraint, inputting the initial estimation result into a residual error reconstruction network, gradually reducing the difference value between the initial estimation image and a real laser image by utilizing the residual error network, and adding the output results of the initial estimation image and the real laser image to obtain a final expected laser reconstructed image with higher quality; because the invention introduces the residual error network, the depth of the network is properly increased through jump connection, so that the network model is easier to optimize, the convergence of the training process is faster and more stable, and the reconstruction quality is better.
Drawings
FIG. 1 is a high-resolution compressive sensing reconstruction network model diagram of a laser image based on a residual error network according to the present invention;
FIG. 2 is a network model diagram of a single reconstruction unit of the residual error reconstruction network according to the present invention;
FIG. 3 is a simulation result of the method of the present invention when the measurement rate is 0.01, (3a), (3b) represent the output results of two different sets of laser images, respectively;
FIG. 4 is a simulation result of the method of the present invention when the measurement rate is 0.10, where (4a) and (4b) represent the output results of two different laser images, respectively;
fig. 5 shows the simulation results of the method of the present invention when the measurement rate is 0.25, and (5a) and (5b) represent the output results of two different laser images, respectively.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The embodiment of the invention provides a high-resolution compressive sensing reconstruction method for a laser image based on a residual error network, which is realized by the following steps:
step 1: obtaining compressed sensing measurement y by multiplying measurement matrix W with real laser image blocki;
Specifically, the real laser image x is subjected to blocking processing to obtain a plurality of real laser image blocks xiMultiplying the real laser image blocks by a measurement matrix W to obtain a compressed sensing measurement yi∈Rn×1;
Step 2: the compressed sensing measure yiBy convex optimization based linear mapping network FfFull connected layer output of (a) initial estimated image blocks of a reconstructed image
Specifically, in the compressive sensing theory, the mapping process for recovering the real laser image x from the compressive sensing measurement value y can be represented as x ═ Wy, where W ∈ Rn×mIs a measurement matrix.
This equation is a sick equation, with no unique solution, and therefore translates it into solving the mapping matrix WfBy estimatingThe minimum, namely:
wherein { (y)1,x1),(y2,x2),…,(yN,xN) Denotes training data, X ═ X1,x2,…,xN],Y=[y1,y2,…,yN]。
The mapping network adopts a loss function shown as a formula (2) to obtain an optimal mapping matrix Wf:
Wherein, Ff(yi,Wf) Is yiThe output after the input passes through the linear mapping network.
Mapping network Ff(. to) is a full connection layer, all samples are trained by a random gradient descent (SGD) training method to obtain a corresponding optimal mapping matrix Wf。
Compressed sensing measurement vector y of any image block through linear mapping networkiAll can obtain better initial estimation image blockThe mapping process is represented as:
and step 3: the initial estimated image blockBy means of a residual network Fr(. output prediction residual image block(3a) Solving a residual error:
specifically, because the difficulty in solving the optimal solution of the formula (3) is high, the linear mapping is generally adopted for approximate solution, and because the quality of the approximate result is poor, further optimization is required to further reduce the size of the solutionAnd xiThe difference between the two signals is estimated by introducing a residual error network, and the residual error diCan be expressed as:
the residual error reconstruction network of the present invention comprises a plurality of residual error reconstruction units, each unit comprising three convolutional layers having convolutional kernel sizes of 11 × 11, 1 × 1 and 7 × 7, respectively, each convolutional layer being followed by a linear activation (ReLU) layer except for the last convolutional layer.
By initially estimating image blocksTo input, residual error networkGenerating pre-estimated residual image blocksWherein WrIs a parameter of the residual network, the process can be expressed as:
further, training the residual error reconstruction network by a random gradient descent (SGD) training method to obtain an optimal parameter WfAnd Wr:
In the training process, in order to avoid the overfitting condition, the invention adds the attenuation weight coefficient to the error function to restrain a larger weight so as to lead the error function to be stably converged.
The method adopts a fixed learning strategy in the initial stage of a training stage, obtains a loss function and an accuracy curve chart after the training is finished, adjusts the learning rate according to the convergence condition, modifies the learning strategy into a stepping type after a plurality of times of training, and obtains corresponding parameters according to the formula (7).
lr=base_lr*gammafloor(iter/stepsize)(7)
Wherein the gamma parameter is set to 0.5, and the base _ lr parameter is set to 10-2. And when a better result is obtained by training, storing the current model, and performing high-resolution reconstruction on the target laser image by using the model parameters.
And 4, step 4: for the initial estimation image blockAnd pre-estimating residual image blockAnd carrying out reconstruction to obtain a desired high-quality laser image reconstruction result.
Specifically, output results of the linear mapping network are mappedAnd residual network outputAdd to obtainTo the final desired high quality laser reconstructed image blockThis process can be expressed as:
after all the original image blocks are subjected to the steps, the obtained reconstructed image blocks are spliced to obtain a final complete laser reconstructed image
The method of the invention fully considers the characteristics of the reconstruction solution of the compressed sensing laser image, adopts convex optimization constraint, outputs a better initial estimation image through a linear mapping network, inputs the result into a residual error reconstruction network for further optimization, obtains a residual error image between the initial estimation and a real laser image, and adds the initial estimation and the real laser image to obtain the expected high-resolution laser image reconstruction result.
In order to prove the effectiveness of the method, three values of the compression measurement rate of 0.01, 0.10 and 0.25 are respectively selected for verification, fig. 3, fig. 4 and fig. 5 clearly show the reconstruction results of the method when the compression measurement rate is 0.01, 0.10 and 0.25, and table 1 shows the peak signal-to-noise ratio (PSNR) values of the two groups of laser image output results when the compression measurement rate is 0.01, 0.10 and 0.25 respectively.
TABLE 1 PSNR values of two groups of laser image output results at different measurement rates
The above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention.
Claims (7)
1. A high-resolution compressive sensing reconstruction method for a laser image based on a residual error network is characterized by comprising the following steps:
by measuring matrix W and real laser image block xiMultiplying to obtain a compressed sensing measurement yi;
The compressed sensing measure yiBy convex optimization based linear mapping network FfOutputting initial estimated image blocks of a fully connected layer output reconstructed image
The initial estimated image blockBy means of a residual network Fr(. output prediction residual image block
2. The residual network-based high-resolution compressive sensing reconstruction method for laser images as claimed in claim 1, wherein the real laser image x is partitioned to obtain a plurality of real laser image blocks xiMultiplying the real laser image blocks by a measurement matrix W to obtain a compressed sensing measurement yi。
3. The residual error network-based laser image high-resolution compressed sensing reconstruction method according to claim 1 or 2, wherein the compressed sensing measurement y isiBy convex optimization based linear mapping network FfOutputting an initial estimated image block of a fully connected layer output reconstructed image blockThe method specifically comprises the following steps:
the mapping process to recover the real laser image x from the compressed sensing measurement value y is denoted as x-Wy, where W ∈ Rn×mIs a measurement matrix;
converting the mapping matrix into a solution mapping matrix WfTo the optimization problem ofWherein, yi∈Rn×1And xi∈Rn ×1Respectively representing the compressed perceptual measurement and the corresponding original signal, { (y)1,x1),(y2,x2),…,(yN,xN) Denotes training data, X ═ X1,x2,…,xN],Y=[y1,y2,…,yN];
Obtaining a mapping matrix W using a loss functionf:Wherein, Ff(yi,Wf) Is yiInputting the output after passing through the linear mapping network;
4. the residual error network-based laser image high-resolution compressed sensing reconstruction method according to claim 3, wherein the method is characterized in thatThen, the method further comprises mapping the matrix WfOptimizing, specifically: training an optimization network by a random gradient descent method until the network converges and obtaining an optimal mapping matrix Wf。
5. The residual network-based laser image high-resolution compressed sensing reconstruction method according to claim 4, wherein the initial estimation image blocks are used as the initial estimation image blocksAs a residual network FrObtaining estimated residual image block by inputting Wherein, WrAre parameters of the residual network.
6. The residual network-based high-resolution compressed sensing reconstruction method for laser images according to claim 5, wherein the initial estimation image blocks are subjected to image reconstructionAnd pre-estimating residual image blockReconstructing, and splicing all reconstructed image blocks to obtain a desired high-quality laser image reconstruction result, specifically: the initial estimation image blockAnd pre-estimating residual image blockAdding the two to obtain a corresponding laser image block reconstruction result:all reconstructed image blocksSplicing to obtain a complete laser reconstruction image x*。
7. The residual error network-based laser image high-resolution compressed sensing reconstruction method according to claim 6, characterized in that the linear mapping network F is usedfThe parameters of (a) initialize the residual network, representing the final desired laser image reconstruction as:wherein the content of the first and second substances,
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910952609.1A CN111091495A (en) | 2019-10-09 | 2019-10-09 | High-resolution compressive sensing reconstruction method for laser image based on residual error network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910952609.1A CN111091495A (en) | 2019-10-09 | 2019-10-09 | High-resolution compressive sensing reconstruction method for laser image based on residual error network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111091495A true CN111091495A (en) | 2020-05-01 |
Family
ID=70393020
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910952609.1A Pending CN111091495A (en) | 2019-10-09 | 2019-10-09 | High-resolution compressive sensing reconstruction method for laser image based on residual error network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111091495A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20220156884A1 (en) * | 2019-05-06 | 2022-05-19 | Sony Group Corporation | Electronic device, method and computer program |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107730451A (en) * | 2017-09-20 | 2018-02-23 | 中国科学院计算技术研究所 | A kind of compressed sensing method for reconstructing and system based on depth residual error network |
CN107784676A (en) * | 2017-09-20 | 2018-03-09 | 中国科学院计算技术研究所 | Compressed sensing calculation matrix optimization method and system based on autocoder network |
KR20190040586A (en) * | 2017-10-11 | 2019-04-19 | 인하대학교 산학협력단 | Method and apparatus for reconstructing single image super-resolution based on artificial neural network |
-
2019
- 2019-10-09 CN CN201910952609.1A patent/CN111091495A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107730451A (en) * | 2017-09-20 | 2018-02-23 | 中国科学院计算技术研究所 | A kind of compressed sensing method for reconstructing and system based on depth residual error network |
CN107784676A (en) * | 2017-09-20 | 2018-03-09 | 中国科学院计算技术研究所 | Compressed sensing calculation matrix optimization method and system based on autocoder network |
KR20190040586A (en) * | 2017-10-11 | 2019-04-19 | 인하대학교 산학협력단 | Method and apparatus for reconstructing single image super-resolution based on artificial neural network |
Non-Patent Citations (2)
Title |
---|
应自炉等: "基于特征补偿的深度神经网络重建超分辨率图像", 《五邑大学学报(自然科学版)》 * |
秦翰林等: "基于改进的分块压缩感知红外图像重建", 《强激光与粒子束 》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20220156884A1 (en) * | 2019-05-06 | 2022-05-19 | Sony Group Corporation | Electronic device, method and computer program |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11153566B1 (en) | Variable bit rate generative compression method based on adversarial learning | |
CN106991646B (en) | Image super-resolution method based on dense connection network | |
CN107730451B (en) | Compressed sensing reconstruction method and system based on depth residual error network | |
CN109087273B (en) | Image restoration method, storage medium and system based on enhanced neural network | |
CN109191382B (en) | Image processing method, device, electronic equipment and computer readable storage medium | |
WO2022267641A1 (en) | Image defogging method and system based on cyclic generative adversarial network | |
CN110675321A (en) | Super-resolution image reconstruction method based on progressive depth residual error network | |
CN109003229B (en) | Magnetic resonance super-resolution reconstruction method based on three-dimensional enhanced depth residual error network | |
CN112508125A (en) | Efficient full-integer quantization method of image detection model | |
CN109102461B (en) | Image reconstruction method, device, equipment and medium for low-sampling block compressed sensing | |
TW202141358A (en) | Method and apparatus for image restoration, storage medium and terminal | |
CN107341776A (en) | Single frames super resolution ratio reconstruction method based on sparse coding and combinatorial mapping | |
CN112116601A (en) | Compressive sensing sampling reconstruction method and system based on linear sampling network and generation countermeasure residual error network | |
CN111047543A (en) | Image enhancement method, device and storage medium | |
CN112912758A (en) | Method and system for adaptive beamforming of ultrasound signals | |
CN109949217A (en) | Video super-resolution method for reconstructing based on residual error study and implicit motion compensation | |
KR20190139781A (en) | CNN-based high resolution image generating apparatus for minimizing data acquisition time and method therefor | |
CN111145102A (en) | Synthetic aperture radar image denoising method based on convolutional neural network | |
CN105184742B (en) | A kind of image de-noising method of the sparse coding based on Laplce's figure characteristic vector | |
CN112950480A (en) | Super-resolution reconstruction method integrating multiple receptive fields and dense residual attention | |
WO2020062074A1 (en) | Reconstructing distorted images using convolutional neural network | |
CN109819256B (en) | Video compression sensing method based on feature sensing | |
CN111091495A (en) | High-resolution compressive sensing reconstruction method for laser image based on residual error network | |
CN110070541B (en) | Image quality evaluation method suitable for small sample data | |
CN116634162A (en) | Post-training quantization method for rate-distortion optimized image compression neural 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 |
Application publication date: 20200501 |
|
RJ01 | Rejection of invention patent application after publication |