CN114596228A - Multispectral image denoising algorithm based on depth 3D convolution sparse coding - Google Patents
Multispectral image denoising algorithm based on depth 3D convolution sparse coding Download PDFInfo
- Publication number
- CN114596228A CN114596228A CN202210212512.9A CN202210212512A CN114596228A CN 114596228 A CN114596228 A CN 114596228A CN 202210212512 A CN202210212512 A CN 202210212512A CN 114596228 A CN114596228 A CN 114596228A
- Authority
- CN
- China
- Prior art keywords
- network
- depth
- convolution
- sparse coding
- multispectral image
- 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
- 238000012549 training Methods 0.000 claims abstract description 20
- 238000012360 testing method Methods 0.000 claims abstract description 9
- 238000012545 processing Methods 0.000 claims abstract description 5
- 238000013178 mathematical model Methods 0.000 claims abstract description 4
- 238000000034 method Methods 0.000 claims description 25
- 238000013528 artificial neural network Methods 0.000 claims description 13
- 238000005457 optimization Methods 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 2
- 230000002194 synthesizing effect Effects 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000003384 imaging method Methods 0.000 description 3
- 241001134453 Lista Species 0.000 description 2
- 206010028980 Neoplasm Diseases 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000005856 abnormality Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012733 comparative method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000002689 soil Substances 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a multispectral image denoising algorithm based on depth 3D convolution sparse coding, which comprises the following steps: step 1, according to the inter-spectral information of the multispectral image, a convolution sparse coding model is expanded into a 3D form, and a 3D-CSC mathematical model is built; step 2, performing iterative solution on the built 3D-CSC model by adopting an iterative shrinkage soft threshold algorithm; step 3, building a deep network: expanding the iterative solution form of the 3D-CSC in the step 2 into a corresponding depth network according to the idea of depth expansion; step 4, training a network after constructing a data set; and 5, testing the performance of the network. The invention applies 3D convolution and considers the inter-spectral characteristics of the image, so that the invention obtains better denoising effect on the aspect of processing the multispectral image.
Description
Technical Field
The invention relates to the technical field of image processing, in particular to a multispectral image denoising algorithm based on depth 3D convolution sparse coding.
Background
The multispectral image is an image containing a plurality of wave bands, and is formed by dividing an incident full-wave band or wide-wave band optical signal into a plurality of narrow-wave band light beams by a specific imaging technology and then imaging the light beams by a plurality of wave band sensors to obtain an image of the multispectral wave bands. Compared with gray level images and common color images, the multispectral image has more wave bands and contains richer spectral information, and is widely applied to multiple scenes, such as crop detection, medical tumor positioning, industrial detection and the like. However, in the multispectral image imaging process, interference from equipment, external environmental noise and the like often occurs, so that the acquired image contains noise, thereby affecting accuracy of subsequent research, such as soil component detection error, tumor positioning inaccuracy, component inspection abnormality and the like. Therefore, multispectral image denoising has important practical application value.
The task of image denoising is to reconstruct a clean image from a noisy image. The existing image denoising methods are mainly divided into a model-based method and a deep learning-based method. The main idea of the model-based method is to transform the image denoising problem into an optimal solution problem by introducing different prior information. The common prior information includes total variation constraint, sparse prior constraint and the like. The model-based method has strong interpretability, but the performance of the model-based method depends on manual parameter design to a great extent, and the flexibility is poor. In recent years, with the rise of deep learning, a deep neural network-based denoising model is gaining wide attention, such as DnCNN, CNLNet, and the like, and the core idea of these networks is to train a neural network through an end-to-end strategy and map a noise image to a clean image. Compared with a model-based method, the image denoising algorithm based on deep learning achieves great performance improvement. However, the existing deep denoising network is mainly constructed manually, and the interpretability is poor.
In order to improve the interpretability of deep neural networks, researchers have proposed deep neural networks based on deep expansion, originating from a type of iterative systolic soft threshold algorithm (litap) called learnable. The LISTA expands the iterative solution of the soft threshold algorithm to be solved for the sparse model into a deep network, wherein each layer in the deep network corresponds to each iteration. Based on the LISTA concept, depth expansion is also applied to image denoising, such as CSCNet, GroupSC, etc. The existing experimental results show that the interpretable networks have the similar performance with the classical deep denoising networks, and meanwhile, the interpretable networks have fewer parameters and stronger interpretability. However, the interpretable networks do not well depict the inter-spectral characteristics of the image, so the invention designs a multispectral image denoising network based on a depth 3D convolution sparse coding model.
Disclosure of Invention
In order to solve the above problems, the present invention provides a multispectral image denoising algorithm based on depth 3D convolution sparse coding, which can improve the interpretability of a depth neural network.
In order to achieve the purpose, the invention is realized by the following technical scheme:
the invention relates to a multispectral image denoising algorithm based on depth 3D convolution sparse coding, which comprises the following steps:
step 2, performing iterative solution on the built 3D-CSC model by adopting an iterative shrinkage soft threshold algorithm;
step 3, building a deep network: according to the idea of depth expansion, expanding the iterative solution form of the 3D-CSC in the step 2 into a corresponding depth network 3D-CSCNet;
step 4, training a network after constructing a data set;
and 5, testing the performance of the network.
The invention is further improved in that: the specific operation of the step 1 is as follows:
step 1.1, the convolution sparse coding model represents X into a linear combination form of m convolution characteristics, and the expansion of the convolution sparse coding model to a 3D form according to the spatial structure of a multispectral image comprises the following steps:
where X represents a clean multi-spectral image,denotes m filters of size h x h, denotes a three-dimensional convolution,denotes the ith filter diCorresponding convolution characteristics;
step (ii) of1.2, convolutional sparse coding assumptionIs sparse, i.e. contains a small amount of non-zero elements, and is obtained by the following optimization problem:
wherein 1 represents l1Norm, λ, is a regularization parameter.
The invention is further improved in that: the specific process of iterative solution in step 2 is as follows: the iterative update formula of formula (2) is:
whereinA is the corresponding sparse code, the superscript T denotes the T (T0, 1, 2.., K-1) th iteration value, T is the transposed symbol, L is the laplace constant, λtFor the regularization parameter corresponding to the t-th iteration,is a soft threshold operator defined as:
the invention is further improved in that: the step 3 comprises the following specific steps: by introducing variablesB ═ D converts formula (3) to:
expanding the formula (5) into a deep neural network, wherein the t-th layer of the neural network corresponds to the t-th iteration of the formula,the method plays a role of an activation function in the network, and C and B are realized through convolution and continuously learn optimization in the network training process.
The invention is further improved in that: the step 4 comprises the following steps:
step 4.1, constructing a data set;
step 4.2, initializing parameters;
step 4.3, training the 3D-CSCNet;
and 4.4, reconstructing an output image.
The beneficial effects of the invention are: according to the method, 3D convolution is applied, the inter-spectral characteristics of the multispectral image are considered, the original convolution sparse coding is expanded to a 3D form, and a corresponding 3D denoising depth network is built, so that the built network has certain interpretability, and meanwhile, a good denoising effect is achieved on the aspect of processing the multispectral image.
Drawings
FIG. 1 is a network structure diagram of a 3D-CSCNet.
Detailed Description
In order to more clearly illustrate the technical solution of the present invention, the following detailed description is made with reference to the accompanying drawings:
the invention discloses a multispectral image denoising algorithm based on depth 3D convolution sparse coding, and provides a convolution sparse coding model based on a 3D convolution dictionary and solves the multispectral image denoising algorithm by utilizing an alternative iteration optimization algorithm in consideration of certain spatial structure of a multispectral image. The iterative optimization solution is then translated into a 3D-CSCNet using a depth-unwrapping strategy, and the corresponding network graph is shown in FIG. 1.
A common noisy image can be represented as:
Y=X+N (1)
wherein,representing a clean multi-spectral image of the image,for noise, the present invention considers only gaussian noise,representing a noisy image, M × N representing the spatial size of the image, and C representing the number of bands of the image. Image denoising is how to recover X from Y. The denoising method specifically comprises the following steps:
step 2, iterative solution is carried out on the built 3D-CSC model by adopting an iterative shrinkage soft threshold algorithm (ISTA);
step 3, building a deep network: according to the idea of depth expansion, expanding the iterative solution form of the 3D-CSC in the step 2 into a corresponding depth network;
step 4, training a network after constructing a data set;
and 5, testing the performance of the network.
In step 1, the convolution sparse coding model represents X as a linear combination form of m convolution characteristics, and the expansion of the convolution sparse coding model to a 3D form according to the spatial structure of a multispectral image comprises the following steps:
whereinRepresenting m filters of size h x h, representing a three-dimensional convolution,denotes the ith filter diCorresponding convolutionAnd (5) characterizing. Convolution sparse coding hypothesisIs sparse, i.e. contains a small number of non-zero elements, and can be obtained by the following optimization problem:
wherein | · | purple1Represents l1Norm, λ, is the regularization parameter.
In the step 2, iterative solution of the original problem in the step 1 is carried out by applying an ISTA algorithm, and a sparse representation vector of each layer is obtained;
the solving process is as follows:
equation (3) the iterative update equation for ISTA is:
whereinA is the corresponding sparse code, t denotes the t (t ═ 0, 1, 2.., K-1) th iteration value,is a soft threshold operator, T is a transposed symbol, L is a Laplace constant, λtThe regularization parameter corresponding to the t-th iteration is defined as:
in step 3, a depth expansion strategy is adopted to expand the 3D-CSC model into a 3D-CSCNet, and the specific method comprises the following steps: and (3) expanding the formula (4) into a deep neural network by adopting a deep expansion idea. First, by introducing variablesConverting formula (4) to:
then, the formula (6) is expanded into a deep neural network, wherein the t-th layer of the neural network corresponds to the t-th iteration of the formula, i.e., each iteration solves each layer in the neural network,the method plays a role of an activation function in the network, and C and B are realized through convolution and continuously learn optimization in the network training process.
The step 4 specifically comprises the following steps: constructing a data set, initializing parameters, training the 3D-CSCNet, and reconstructing an output image.
Constructing a data set: the network training of the invention adopts a data set CAVE, and has 32 scenes in total, wherein each scene has 31 gray images with the size of 512 multiplied by 512, and the 31 wave bands correspond to the scene. Therefore, the 31 gray images are sequentially synthesized into a three-dimensional tensor with the size of 512 × 512 × 31. Before reading data, the three-dimensional tensors are cut according to the size of 32 x 32, the step length of a cutting interval is 31 pixel values, the synthesized tensors are stored into a mat file for importing data in the following process, and finally the cut data are normalized according to the ratio of 6: 4, dividing the data into a training set and a testing set, wherein the gray scale range of the data after normalization processing is [0, 1 ].
Initializing parameters: inputting a multispectral image Y, randomly initializing a 3D dictionary D, and regularizing a parameter lambda;
training 3D-CSCNet: the invention simulates a noisy image by manually adding gaussian noise. Network parameters of the whole 3D-CSCNet are Θ ═ C, B, D, λ, and these parameters can be obtained by training the network in a form of end-to-end supervised learning. The loss function employed in the present invention is a Mean Square Error (MSE) loss function defined as
Where M is the number of training samples, fΘ3D-CSCNet, Y with the parameter thetajRepresents the jth training sample, fΘ(Yj) Is an estimate of 3D-CSCNet. The method realizes 3D-CSCNet on PyTorch, trains and tests on RTX2070Super GPU, sets the number of samples (Batch Size) and the number of training rounds (epoch) captured at one time of training to be 8 and 200, sets the Size of a convolution kernel to be 7 multiplied by 7, and adopts an Adam optimizer. When training the 3D-CSCNet, the stack expansion number K is set, and when solving the sparse code corresponding to the t (t ═ 0, 1, 2.., K) layer, equation (6) is applied, and it is known that the set iteration is ended.
Step 5, the concrete method for testing the model comprises the following steps: the network performance is evaluated under three different Noise levels of [ 0.1 ], 0.2 and 0.3 ], and the evaluation index adopts average Peak Signal-to-Noise Ratio (MPSNR). Comparative methods include LRMR, LRTV, LLRGTV and CSCNet, and the results are shown in table 1:
TABLE 1 MPSNR results of different methods on CAVE test set
Algorithm | σ=25 | σ=50 | σ=75 |
LRMR | 33.48 | 28.44 | 25.00 |
LRTV | 36.21 | 32.03 | 29.41 |
LLRGTV | 34.40 | 29.71 | 26.82 |
CSCNet | 35.36 | 32.50 | 30.21 |
3D-CSCNet | 37.40 | 33.53 | 25.69 |
As can be seen from Table 1, the 3D-CSCNet of the invention has better denoising performance. Notably, 3D-CSCNet performed better on CAVE, a dataset with sigma values of 25 and 50, over the common noise, compared to CSCNet, which also demonstrates the effectiveness of the model.
Claims (5)
1. The multispectral image denoising algorithm based on the depth 3D convolution sparse coding is characterized by comprising the following steps: the method comprises the following steps:
step 1, according to the inter-spectral information of the multispectral image, a convolution sparse coding model is expanded into a 3D form, and a 3D-CSC mathematical model is built;
step 2, performing iterative solution on the built 3D-CSC model by adopting an iterative shrinkage soft threshold algorithm;
step 3, building a deep network: according to the idea of depth expansion, expanding the iterative solution form of the 3D-CSC in the step 2 into a corresponding depth network 3D-CSCNet;
step 4, training a network after constructing a data set;
and 5, testing the performance of the network.
2. The depth 3D convolutional sparse coding based multispectral image denoising algorithm of claim 1, wherein: the specific operation of the step 1 is as follows:
step 1.1, the convolution sparse coding model represents an input signal X into a linear combination form of m convolution characteristics, and the expansion of the input signal X into a 3D form according to the spatial structure of a multispectral image comprises the following steps:
where X represents a clean multi-spectral image,representing m filters of size h x h, representing a three-dimensional convolution,denotes the ith filter diCorresponding convolution characteristics;
step 1.2, convolutional sparse coding assumptionIs sparse, i.e. contains a small amount of non-zero elements, and is obtained by the following optimization problem:
wherein | · | charging1Represents l1Norm, λ, is a regularization parameter.
3. The multispectral image denoising algorithm based on depth 3D convolutional sparse coding of claim 2, wherein: the specific process of iterative solution in step 2 is as follows: the iterative update formula of formula (2) is:
whereinA is the corresponding sparse code, the superscript T denotes the T (T0, 1, 2.., K-1) th iteration value, T is the transposed symbol, L is the laplace constant, λtFor the regularization parameter corresponding to the t-th iteration,is a soft threshold operator defined as:
4. the depth 3D convolutional sparse coding based multispectral image denoising algorithm of claim 3, wherein: the step 3 comprises the following specific steps: by introducing variablesB ═ D converts formula (3) to:
expand equation (5) intoA deep neural network, wherein the t-th level of the neural network corresponds to the t-th iteration of the formula,the method plays a role of an activation function in the network, and C and B are realized through convolution and continuously learn optimization in the network training process.
5. The depth 3D convolutional sparse coding based multispectral image denoising algorithm of claim 1, wherein: the step 4 comprises the following steps:
step 4.1, constructing a data set: and synthesizing images in each scene into a three-dimensional tensor according to the adopted data set CAVE, cutting the three-dimensional tensor according to requirements, and performing normalization processing according to the following steps of 6: 4 into a training set and a test set;
step 4.2, initializing parameters;
step 4.3, training the 3D-CSCNet;
and 4.4, reconstructing an output image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210212512.9A CN114596228A (en) | 2022-03-04 | 2022-03-04 | Multispectral image denoising algorithm based on depth 3D convolution sparse coding |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210212512.9A CN114596228A (en) | 2022-03-04 | 2022-03-04 | Multispectral image denoising algorithm based on depth 3D convolution sparse coding |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114596228A true CN114596228A (en) | 2022-06-07 |
Family
ID=81808250
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210212512.9A Pending CN114596228A (en) | 2022-03-04 | 2022-03-04 | Multispectral image denoising algorithm based on depth 3D convolution sparse coding |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114596228A (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190096049A1 (en) * | 2017-09-27 | 2019-03-28 | Korea Advanced Institute Of Science And Technology | Method and Apparatus for Reconstructing Hyperspectral Image Using Artificial Intelligence |
AU2020100460A4 (en) * | 2020-03-26 | 2020-04-30 | Huang, Shuying DR | Single image deraining algorithm based on multi-scale dictionary |
CN112907449A (en) * | 2021-02-22 | 2021-06-04 | 西南大学 | Image super-resolution reconstruction method based on deep convolution sparse coding |
CN113344811A (en) * | 2021-05-31 | 2021-09-03 | 西南大学 | Multilayer convolution sparse coding weighted recursive denoising deep neural network and method |
-
2022
- 2022-03-04 CN CN202210212512.9A patent/CN114596228A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190096049A1 (en) * | 2017-09-27 | 2019-03-28 | Korea Advanced Institute Of Science And Technology | Method and Apparatus for Reconstructing Hyperspectral Image Using Artificial Intelligence |
AU2020100460A4 (en) * | 2020-03-26 | 2020-04-30 | Huang, Shuying DR | Single image deraining algorithm based on multi-scale dictionary |
CN112907449A (en) * | 2021-02-22 | 2021-06-04 | 西南大学 | Image super-resolution reconstruction method based on deep convolution sparse coding |
CN113344811A (en) * | 2021-05-31 | 2021-09-03 | 西南大学 | Multilayer convolution sparse coding weighted recursive denoising deep neural network and method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107633486B (en) | Structural magnetic resonance image denoising method based on three-dimensional full-convolution neural network | |
CN112233026A (en) | SAR image denoising method based on multi-scale residual attention network | |
CN111127354B (en) | Single-image rain removing method based on multi-scale dictionary learning | |
CN106408530A (en) | Sparse and low-rank matrix approximation-based hyperspectral image restoration method | |
CN110490219B (en) | Method for reconstructing seismic data by U-net network based on texture constraint | |
CN111598786B (en) | Hyperspectral image unmixing method based on depth denoising self-coding network | |
CN111738954B (en) | Single-frame turbulence degradation image distortion removal method based on double-layer cavity U-Net model | |
CN111080555B (en) | Hyperspectral image noise reduction method based on three-dimensional quasi-recurrent neural network | |
CN114723631B (en) | Image denoising method, system and device based on depth context prior and multi-scale reconstruction sub-network | |
CN102542542A (en) | Image denoising method based on non-local sparse model | |
CN112767271B (en) | Method for deep noise reduction of hyperspectral image based on three-dimensional variational network | |
Zhang et al. | VST-Net: Variance-stabilizing transformation inspired network for Poisson denoising | |
CN110533591B (en) | Super-resolution image reconstruction method based on codec structure | |
CN114419392A (en) | Hyperspectral snapshot image recovery method, device, equipment and medium | |
CN106067165B (en) | High spectrum image denoising method based on clustering sparse random field | |
CN109887050A (en) | A kind of code aperture spectrum imaging method based on self-adapting dictionary study | |
CN112967210A (en) | Unmanned aerial vehicle image denoising method based on full convolution twin network | |
CN113327205B (en) | Phase denoising method based on convolutional neural network | |
CN115861083A (en) | Hyperspectral and multispectral remote sensing fusion method for multi-scale and global features | |
CN106296583B (en) | Based on image block group sparse coding and the noisy high spectrum image ultra-resolution ratio reconstructing method that in pairs maps | |
CN115205136A (en) | Image rain removing method based on Fourier prior | |
CN111091580B (en) | Stumpage image segmentation method based on improved ResNet-UNet network | |
CN111986114B (en) | Double-scale image blind denoising method and system based on self-supervision constraint | |
CN114596228A (en) | Multispectral image denoising algorithm based on depth 3D convolution sparse coding | |
CN116167939A (en) | Hyperspectral image denoising method and system based on spatial spectrum combined self-attention mechanism |
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 |