CN113657521B - Method for separating two mutually exclusive components in image - Google Patents
Method for separating two mutually exclusive components in image Download PDFInfo
- Publication number
- CN113657521B CN113657521B CN202110966494.9A CN202110966494A CN113657521B CN 113657521 B CN113657521 B CN 113657521B CN 202110966494 A CN202110966494 A CN 202110966494A CN 113657521 B CN113657521 B CN 113657521B
- Authority
- CN
- China
- Prior art keywords
- image
- feature
- activated
- relu
- activation function
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2134—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on separation criteria, e.g. independent component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a method for separating two mutually exclusive components in an image, which is a neural network based on a double-flow information interaction module and is characterized in that: the double-flow information interaction module adopts the following steps to realize separation of two mutually exclusive components in the image: s1, outputting image features of acquired images by two parallel convolution layers in a neural network; s2, the decomposition unit adopts two branches to activate the image features to obtain a first activated image feature and a second activated image feature; s3, combining the first activated image feature and the second activated image feature by the fusion image unit to generate a first combined feature and a second combined feature; compared with the traditional dual-branch network which is activated by using the ReLU activation function, the invention introduces the negative ReLU activation function to reserve the characteristics discarded by the ReLU activation function and transmit the characteristics to the other branch, thereby maintaining the total information, avoiding the information loss and improving the information transmission efficiency.
Description
Technical Field
The invention belongs to an image processing method, and particularly relates to a method for separating two mutually exclusive components in an image.
Background
The existing general double-branch image restoration framework is a general image restoration network structure and can be used for image compression artifact removal, image denoising and image superdivision, the double-branch strategy in the work decomposes image restoration into two branches of textures and structures, and a convolution layer is used as an interaction mode of the two branches. In reality, however, the work involves only one-way interactions (Lateral Connections), and the features of the interactions are not filtered, affine transformations are performed using only convolution, and the information utilization efficiency is poor.
Another dual-branch image restoration structure is used for coping with various composite image degradation problems. This work, while adding Gate modules (Gate modules) and multi-channel attention selection modules (MSAM) for feature selection, is only applied at the end of the network structure, and does not involve any form of information interaction in the feed-forward process of the respective branches, so the flexibility of information flow in the network is poor. In fact, the mechanism proposed by this work is in principle also applicable to the fusion of the features of the two branches, with no way of information interaction between the two branches. Meanwhile, the prior art also proposes to use two branches with different depths to process image restoration tasks, specifically including two tasks of image denoising and image superclassification, and needs to pay attention to that the two tasks are solved by using respective models respectively, while in the prior art, the basic network structure part is shared. This work suggests that branches of different depths can be used to extract features of different semantic hierarchies, thus connecting features of two branches and screening using an attention mechanism. Obviously, the characteristics between the two branches of the work lack effective interaction in the feedforward process, the characteristic information flows singly, and the utilization rate of the characteristic information is lower.
Existing task-specific image restoration frameworks: in the image reflection decomposition task, CEILNet is a single-branch network, edges are estimated first, then clear images are estimated, and no reflection component is estimated; zhang et al, is a single-branch network, and finally splits the output 6 channels to obtain two component estimates, so that no explicit information interaction operation between the components exists, the network capacity is weak, and complex reflection scenes are difficult to process; the BDN estimates two components in a cascading way, but no interaction exists among the characteristics of the components, and the cascading optimization mode in a natural image domain is easy to accumulate errors, even the reflection components in a transmission diagram are enhanced, and the visual effect of the reflection component estimation is relatively poor; ERRNet introduces attention and multiscale modules, but still lacks an estimate of the reflectogram, and naturally there is no interaction between component features, thus being unfavorable for obtaining better recovery results; IBCLN attempts to use a dual-branch iterative reflectometry model, but only interacts in the natural image domain, similar to BDN, and is also prone to cause accumulation of erroneous estimates, which is worse on most samples of the test set, and the two branches in the feed-forward process are not interacted in any form, and after each round of connection of the original image, the estimation of the reflectogram and the transmission map are not mutually exclusive, and deviate from the original linear assumption.
In the image denoising task, dnCNN uses a single-branch network to model noise in an image, and uses an additive model to subtract predicted noise from the noise image to obtain a sharp image. Obviously, the model is only regular in the search range of the noise domain for image degradation, lacks the constraint of clear images, and therefore has poor generalization performance. After the noise level is improved, the denoising effect of the method can be obviously reduced. The other methods, such as CBDNet, SADNet and the like, mostly enable the model to learn noise residual errors in a long-distance and short-distance residual error connection mode, and do not apply a noise model in a feature domain or have feature interaction between two components.
The main problem of the existing DPRN network is that the two components assumed by the method are strongly correlated depending on t, and the mutual exclusion between the two components cannot be guaranteed by decomposing the two components, so that the characteristics of one branch which are not needed are not necessarily exactly the characteristics of the other branch which are needed. The residual connection between the two branches also lacks an assessment of the usefulness of the feature, simply by summing. In addition, this mode cannot be generalized to other tasks, such as the addition mode in the image reflection decomposition task, which can cause the feature content of two branches to be almost the same, and is unfavorable for learning mutually exclusive component features.
Disclosure of Invention
In order to solve the problems existing in the prior art, the invention provides a method for separating two mutually exclusive components in an image, which comprises the steps of establishing a double-flow information interaction module comprising two branches of two pairs of ReLU and negative ReLU activation functions, and performing activation processing on the characteristics of the image by using the pair of ReLU and negative ReLU activation functions during the image processing, so that the evaluation on the characteristic usefulness is introduced; meanwhile, the two branches can exchange the characteristic information activated by the negative ReLU and useless for the current branch in any module, and can be reused in the other branch; by the aid of the method, the utilization efficiency of information can be obviously improved, meanwhile, information interaction of features between two branches is enhanced, and more reasonable hierarchical decomposition results are obtained.
The invention adopts the following technical scheme:
the method is a neural network based on a double-flow information interaction module, and the double-flow information interaction module adopts the following steps to separate two mutually exclusive components in an image:
s1, outputting image features of acquired images by two parallel convolution layers in a neural network;
s2, the decomposition unit adopts two branches to activate the image features to obtain a first activated image feature and a second activated image feature;
s3, combining the first activated image feature and the second activated image feature by the fusion image unit to generate a first combined feature and a second combined feature; wherein:
the two-way branch is adopted by the decomposition unit and is composed of a ReLU and a negative ReLU activation function, and the first activated image feature is composed of a feature activated by the first ReLU activation function and a feature activated by the first negative ReLU activation function; the second activated image feature is a feature activated by a second ReLU activation function and a feature activated by a second negative ReLU activation function;
the first combination feature is obtained by fusing the feature activated by the first ReLU activation function with the feature activated by the second negative ReLU activation function;
the second binding characteristic is obtained by fusing the characteristic activated by the second ReLU activation function with the characteristic activated by the first negative ReLU activation function.
Further, the decomposition unit adopts two branches to activate the image feature:
s201, the negative ReLU activation function is established in the two-way branch as follows:
ReLU - (x)：＝x-ReLU(x)＝min(x,0)#(1)
wherein the ReLU function is defined asRetrieving features discarded by a ReLU function through a negative ReLU activation function, where x is taken to be 0 Input denoted as first layer of neural network, < >>The feature of the ith branch after passing through the l superimposed network modules is also denoted +.> Where i ε {1,2};
s202, inputting the (i+1) th module requires double-flow information interaction processing as shown in the following formula:
wherein the method comprises the steps ofThe method further comprises the step that the fusion image unit adopts channel connection or point-to-point addition to fuse the first activated image feature and the second activated image feature.
Advantageous effects
1. Compared with the traditional dual-branch network which is activated by using the ReLU activation function, the invention introduces the negative ReLU activation function to reserve the characteristics discarded by the ReLU activation function and transmit the characteristics to the other branch, thereby maintaining the total information, avoiding the information loss and improving the information transmission efficiency; the invention solves the problems of information loss and DeadReLU of the ReLU activation function commonly used in the existing neural network.
2. Compared with a double-branch network without information exchange, the method introduces information interaction between two branches, and is beneficial to gradual decoupling of two layers of characteristics of complex coupling by the neural network at each layer.
3. The invention improves the information utilization efficiency of the existing neural network model in a double-flow information interaction mode, can obtain a faster convergence speed during training, and can show higher performance than most methods during testing.
4. The invention has compatibility with the prior two-decomposition task, such as image reflection decomposition, image denoising, image mole pattern removal, eigenvector decomposition task and the like, and has wider application prospect.
Drawings
FIG. 1 is a neural network model of a dual-stream information interaction module of a method of separating two mutually exclusive components in an image according to the present invention;
FIG. 2 is a schematic diagram of a dual-stream information interaction process in a method for separating two mutually exclusive components in an image according to the present invention;
FIG. 3 is a graph showing the result of an output image of a reflectometry using a method of separating two mutually exclusive components of an image according to the present invention;
FIG. 4 is a graph showing the output of image noise decomposition using a method of separating two mutually exclusive components in an image according to the present invention;
FIG. 5 is an output of image moire decomposition using a method of the present invention for separating two mutually exclusive components in an image;
FIG. 6 is an output image result of an image eigen-map decomposition using a method of the present invention for separating two mutually exclusive components in an image;
FIG. 7 is a comparison of the visual effects of reflectometry decomposition of different network models on a real test dataset of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the following detailed description of the present invention will be made with reference to the accompanying drawings and examples, which are given by way of illustration only, and not by way of limitation, and thus should not be construed as limiting the scope of the present invention.
As shown in fig. 1, the dual-stream information interaction network provided by the invention separates two mutually exclusive components in an image by establishing a dual-branch information interaction module, wherein the dual-stream information interaction network is characterized in that features to be activated, which are output by two parallel convolution layers in two branches of a neural network, are activated by using a ReLU and a negative ReLU activation function respectively, features 1B and 2B activated by the negative ReLU activation function in the two branches are exchanged, and then a channel connection or a point-to-point addition and other feature combination methods are used for obtaining a combination feature 1 and a combination feature 2, and the combination feature 2 is transmitted into a subsequent dual-stream information interaction module, and a main frame structure of the whole dual-stream information interaction network is formed by the dual-stream information interaction modules.
The invention adopts a double-flow information interactive image decomposition strategy, and the strategy finally recovers the decomposition results of two components in the input image by carrying out characteristic interaction of the input image layer by layer in two branches of the neural network. The double-flow information interaction module is a basic module for forming the neural network used by the invention. In particular, the captured image I with a hierarchical superposition phenomenon may be generally expressed in the form of a linear combination of hierarchical components, such as i=t+r for image reflection decomposition, where the hierarchical component T represents a transmission layer and R represents a reflection layer; for image denoising there is i=b+n, where the hierarchical component B represents the background layer and N represents the noise layer. Consider first, an estimate for any set of decompositionAnd->If they add to the input image I, there is always one residual term Q that satisfies: />The object of the decomposition task is to obtain the residual term and to follow this residual term from +.>Minus from (a) and add->And (3) upper part. Thus, in this view, no useless information exists in the network, but only misplaced information. Then, re-write Q as Q: =q T -Q R Then there is +.>And +.>The invention is to->And->The information of no value is respectively marked as Q T And Q R . In other words, two components of the network prediction +.>And->Instead of directly discarding the garbage, it is possible to obtain information that is not useful to the counterpart but may be useful to itself by exchanging information between each other. The invention is applied to the interaction of the middle characteristics of the dual-branch neural network, and provides a general dual-flow information interaction strategy for designing the neural network.
In the design of the neural network, the activation function can well evaluate the usefulness of the feature information, that is, the feature activated by the activation function is the feature useful for the branch and needs to be reserved; while features suppressed by the activation function are those that are not useful for that branch, but may be valuable for another branch, so need to be passed to another branch (insteadRather than being directly discarded as in a typical network design). Corresponding to the formula described above, forInformation Q of no value T 、Q R Corresponds to the portion of the information that is suppressed by the activation function. Through the ReLU activation function used in the invention (expressed as +.> ) Intermediate features F of two branches in a network T And F R Is equivalent to subtracting the respective suppressed information, i.e., sigma (F T )＝F T -Q T ，σ(F R )＝F R -Q R Where σ (·) represents the ReLU activation function. Thus, only the information of each suppressed in the two branches is added to the other branch, namely F' T ＝σ(F T )+Q R ＝σ(F T )+F R -σ(F R )＝σ(F T )+ρ(F R ) And F' R ＝σ(F R )+Q T ＝σ(F R )+F T -σ(F T )＝σ(F R )+ρ(F T ) Where ρ (·) is referred to as the negative ReLU activation function, which acts exactly opposite to the ReLU activation function, denoted ReLU - (x) The method comprises the following steps =min (x, 0) =x-ReLU (x). Thus, in the feed-forward process of the neural network, sigma (·) and ρ (·) functions may be used to extract useful and useless information for the current branch, and the interaction of the feature information may be performed by transferring the extracted useless information to another branch (here, the information transfer may be performed by directly adding the useful information of the current branch and the useless information of the other branch, or by channel connection re-fusion, etc., and the manner of transferring or fusing is not fixed).
As shown in fig. 2, the present invention proposes a dual-stream information interactive module for use in an image decomposition strategy, S201, in which a negative ReLU activation function is established in a dual-path branch as follows:
ReLU - (x)：＝x-ReLU(x)＝min(x,0)#(2)
wherein the ReLU function is defined asThe feature discarded by the ReLU function is obtained by the negative ReLU activation function, and a schematic diagram of the two functions in a planar rectangular coordinate system is shown in fig. 2 (a).
The invention uses x 0 Denoted as input to the first layer of the neural network,the feature of the ith branch after passing through the l superimposed network modules is also denoted +.>Where i e {1,2}.
S202, inputting the (i+1) th module requires double-flow information interaction processing as shown in the following formula:
wherein the method comprises the steps ofThe fusion manner of features activated by the ReLU and negative ReLU functions, such as channel connection operation and feature addition operation, respectively, is represented, and two different dual-flow information interaction modules as shown in fig. 2 (b) and (c) are designed according to different fusion operations.
X in the figure 1 And x 2 For inputting image features, activating the ReLU and the negative ReLU activating functions respectively after convolution layer processing, exchanging activation of the negative ReLU in two branches, wherein the YTMT Block-C module shown in the 2 (b) diagram also needs to use 1x1 convolution to fuse the features of channel connection after exchange, and then the two modules need to use the channel attention and pixel attention module to further screen the featuresSelecting. The characteristics after double-flow information interaction can be obviously obtained by observing the formula (3)And->Is +.>And->Without any loss. This feature ensures that no information will flow away from the activation function after activation, thereby improving the information utilization efficiency between layers.
The double-flow information exchange module provided by the invention is a task of processing image reflection decomposition, image denoising, image moire removal, eigenvector decomposition, image defogging and the like through a decomposition strategy, and can be directly or simply regarded as an additive model through transformation (such as logarithmic transformation). For images with interference captured under certain scenes, the invention can realize separation of two parts by respectively carrying out supervised reconstruction on clear and interference components of the images.
After the network is trained and deployed on the data sets of different tasks, the network can be applied to corresponding decomposition tasks. The application of the dual stream information interaction strategy in some of its tasks is briefly described below.
When a photograph is taken through a transparent medium such as glass, a superimposed image I of an image T formed by light transmitted through the glass and an image R formed by light reflected by the glass is often taken, and it is generally assumed that the linear addition principle, i.e., i=t+r, is satisfied therebetween. Such superimposed images may seriously affect the overall effect of the photograph and the recognition and segmentation of objects behind the glass. Related expressions are widely available for shooting from inside the window to outside and other scenes that need to be shot through various transparent media. In fact, there is also often a tendency for reflections to existIn some valuable information, the method has potential utilization value in the criminal investigation field and the like. Researchers often utilize computer algorithms to separate the two layers. This task has a high degree of discomfort on a single image, that is to say, when an input image I is given, the principle of linear addition is satisfiedAnd->There are numerous combinations. While none of the previous methods effectively uses the linear relationship between the features of the two components, the neural network designed by the method for separating the two mutually exclusive components in the image of the present invention can effectively use the linear relationship, and the output image result of the reflection decomposition performed by the present invention is shown in fig. 3, so that the present invention can give a reflection decomposition result with good visual effect even for the image with two layers being seriously overlapped, which illustrates the effectiveness of the method for separating the two mutually exclusive components in the image of the present invention.
Image denoising is a common image restoration task, and usually generates obvious noise when the ISO of a digital camera is too high, and generates noisy images when the ISO of the digital camera is too high and electronic interference is generated in a complex environment. When noise is present in the captured image, the high semantic tasks downstream can be severely impacted. The prior art often models only noise information and ignores the information of the clear image, while another part of the method mainly focuses on the reconstruction of the clear image, ignoring the residual information of noise as a constraint in supervision, so it is obvious that a strategy using double branches is quite necessary here. For additive noise, the reconstruction constraint of the invention is i=b+n, where I is the input noisy image, B is the background clear image, and N is the noise image, so it is reasonable to use a dual stream information interaction strategy for such additive model. The output of the gaussian noise decomposition with standard deviation 85 using the strategy of the present invention is shown in fig. 4.
Moire patterns in imagesIt often happens that two patterns with similar spatial frequencies are superimposed, and irregular color interference fringes are produced at this time. In daily life, moire phenomenon occurs when the spatial frequency of pixels of a photographic apparatus used is close to the spatial frequency of stripes in a photographed picture. This phenomenon is likely to occur when photographing objects with regular textures such as electronic screens, clothes, and buildings. The moire phenomenon not only can obviously influence the visual quality of the image, but also can influence the performance of other downstream computer vision tasks. However, the research of removing mole marks by using a depth model is new, and no research is currently being conducted to provide a proper physical model to deal with the problem. For the task, the residual errors M of the clear images B and the moire image I are respectively regarded as double-flow information, so that the image moire removing task can still be formally expressed as an additive model I=B+M, and a double-flow information interaction network can be trained to remove the moire in the image. As shown in FIG. 5, which shows the output result of the moire decomposition of an image by a method for separating two mutually exclusive components in an image according to the present invention, it can be seen that the moire has been substantially obtained from a clear image despite imperfect residual assumptionsIs separated into residual pictures->Is a kind of medium.
The eigenvector decomposition task is to decompose an image into a reflection chart and an illumination chart, the reflection chart shows the color of an object, the reflection chart of the same object is unchanged for the angle and illumination condition of a camera, and the illumination chart changes according to the different angles of view and illumination conditions, wherein the shadows, the brightness change, the mutual reflection and the like are included. Since visual tasks are in most cases more focused on extracting information of the object itself, and changes in illumination can cause serious disturbances to the task itself, eigenmap decomposition is significant for high semantic tasks (e.g. detection and segmentation, etc.). Intrinsic images are also widely used in the field of computational photography, such as material re-coloring, re-illumination, texture re-generation, and stylization. The eigen-map decomposition is formally expressed in the logarithmic domain as logi=loga+logs, where I is the input image, a is the reflection map, and S is the illumination map. It is apparent that the model is additive in logarithmic form, and that the strategy of the present invention can be well applied, and the output of the eigenmap decomposition using this strategy is shown in fig. 6.
The double-flow information interaction module can be used as the design basis of double-branch image recovery networks with various structures. The neural network according to the present invention can effectively process multiple kinds of image decomposition tasks, wherein the present invention obtains almost all the best numerical comparison results on the test data set of the reflection decomposition of a single image (see table 1, here, YTMT-UCT network structure is used, left table is the predicted numerical index evaluation of T, right table is the predicted numerical index evaluation of R, here, two evaluation indexes of PSNR and SSIM are adopted to evaluate the decomposition result of the model, and the best/next best results are respectively indicated by red/blue fonts).
TABLE 1 comparison of reflectance decomposition values of different network models on real test data sets
Fig. 7 shows a comparison of visual effects of different network models on a real test data set, and in combination with table 1, it can be explained that the strategy proposed by the present invention can effectively separate two different components in an image, and can obtain more prominent visual and numerical expressions than other methods, and effectively explain the effectiveness of the method of the present invention.
The invention is not limited to the embodiments described above. The above description of specific embodiments is intended to describe and illustrate the technical aspects of the present invention, and is intended to be illustrative only and not limiting. Numerous specific modifications can be made by those skilled in the art without departing from the spirit of the invention and scope of the claims, which are within the scope of the invention.
Claims (2)
1. A method for separating two mutually exclusive components in an image is a neural network based on a double-flow information interaction module, and is characterized in that: the double-flow information interaction module adopts the following steps to realize separation of two mutually exclusive components in the image:
s1, outputting image features of acquired images by two parallel convolution layers in a neural network;
s2, the decomposition unit adopts two branches to activate the image features to obtain a first activated image feature and a second activated image feature; the two-way branch image feature activation process is adopted by the decomposition unit:
s201, the negative ReLU activation function is established in the two-way branch as follows:
ReLU - (x)：＝x-ReLU(x)＝min(x,0)
wherein the ReLU function is defined as ReLU (x): =max (x, 0), the feature discarded by the ReLU function is taken by the negative ReLU activation function, where x is taken as 0 Denoted as input to the first layer of the neural network,the feature of the ith branch after passing through the l superimposed network modules is also denoted +.>Where i ε {1,2};
s202, inputting the (i+1) th module requires double-flow information interaction processing as shown in the following formula:
wherein the method comprises the steps ofA fusion approach representing features activated by the ReLU and negative ReLU functions, respectively;
s3, combining the first activated image feature and the second activated image feature by the fusion image unit to generate a first combined feature and a second combined feature; wherein:
the two-way branch is adopted by the decomposition unit and is composed of a ReLU and a negative ReLU activation function, and the first activated image feature is composed of a feature activated by the first ReLU activation function and a feature activated by the first negative ReLU activation function; the second activated image feature is a feature activated by a second ReLU activation function and a feature activated by a second negative ReLU activation function;
the first combination feature is obtained by fusing the feature activated by the first ReLU activation function with the feature activated by the second negative ReLU activation function;
the second binding characteristic is a characteristic activated by a second ReLU activation function and a first negative
Feature fusion of ReLU activation function activation is obtained.
2. A method of separating two mutually exclusive components in an image according to claim 1, wherein: the fusion image unit fuses the first activation image feature and the second activation image feature by adopting channel connection or point-to-point addition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110966494.9A CN113657521B (en) | 2021-08-23 | 2021-08-23 | Method for separating two mutually exclusive components in image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110966494.9A CN113657521B (en) | 2021-08-23 | 2021-08-23 | Method for separating two mutually exclusive components in image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113657521A CN113657521A (en) | 2021-11-16 |
| CN113657521B true CN113657521B (en) | 2023-09-19 |
Family
ID=78480656
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110966494.9A Active CN113657521B (en) | 2021-08-23 | 2021-08-23 | Method for separating two mutually exclusive components in image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113657521B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108446729A (en) * | 2018-03-13 | 2018-08-24 | 天津工业大学 | Egg embryo classification method based on convolutional neural networks |
| CN110574051A (en) * | 2017-05-31 | 2019-12-13 | 英特尔公司 | Computationally efficient quaternion-based machine learning system |
| WO2020155614A1 (en) * | 2019-01-31 | 2020-08-06 | 北京市商汤科技开发有限公司 | Image processing method and device |
| WO2021088300A1 (en) * | 2019-11-09 | 2021-05-14 | 北京工业大学 | Rgb-d multi-mode fusion personnel detection method based on asymmetric double-stream network |
| CN113034353A (en) * | 2021-04-09 | 2021-06-25 | 西安建筑科技大学 | Essential image decomposition method and system based on cross convolution neural network |
-
2021
- 2021-08-23 CN CN202110966494.9A patent/CN113657521B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110574051A (en) * | 2017-05-31 | 2019-12-13 | 英特尔公司 | Computationally efficient quaternion-based machine learning system |
| CN108446729A (en) * | 2018-03-13 | 2018-08-24 | 天津工业大学 | Egg embryo classification method based on convolutional neural networks |
| WO2020155614A1 (en) * | 2019-01-31 | 2020-08-06 | 北京市商汤科技开发有限公司 | Image processing method and device |
| WO2021088300A1 (en) * | 2019-11-09 | 2021-05-14 | 北京工业大学 | Rgb-d multi-mode fusion personnel detection method based on asymmetric double-stream network |
| CN113034353A (en) * | 2021-04-09 | 2021-06-25 | 西安建筑科技大学 | Essential image decomposition method and system based on cross convolution neural network |
Non-Patent Citations (3)
| Title |
|---|
| 叶绿 ; 段婷 ; 朱家懿 ; Nwobodo Samuel Chuwkuebuka * |
| 基于多层特征融合的单目深度估计模型;叶绿;段婷;朱家懿;Nwobodo Samuel Chuwkuebuka;nnor Arnold Antwi;;浙江科技学院学报(第04期);全文 * |
| 基于心脏电影磁共振图像的左心肌分割新进展;王慧;王丽嘉;;中国生物医学工程学报(第02期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113657521A (en) | 2021-11-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tang et al. | Investigating haze-relevant features in a learning framework for image dehazing | |
| Yang et al. | Joint rain detection and removal via iterative region dependent multi-task learning | |
| CN110223242A (en) | A kind of video turbulent flow removing method based on time-space domain Residual Generation confrontation network | |
| Chi et al. | Single image reflection removal using deep encoder-decoder network | |
| CN110838092B (en) | Underwater image restoration method based on convolutional neural network | |
| CN111260560A (en) | Multi-frame video super-resolution method fused with attention mechanism | |
| Zhou et al. | Underwater image restoration via feature priors to estimate background light and optimized transmission map | |
| US20210287345A1 (en) | A Priori Constraint and Outlier Suppression Based Image Deblurring Method | |
| Sun et al. | Deep maximum a posterior estimator for video denoising | |
| CN115357944A (en) | Image tampering detection method based on shallow feature enhancement and attention mechanism | |
| Yang et al. | Sir-former: Stereo image restoration using transformer | |
| Wei et al. | DA-DRN: A degradation-aware deep Retinex network for low-light image enhancement | |
| CN113657521B (en) | Method for separating two mutually exclusive components in image | |
| CN114140366A (en) | Infrared image and visible light image fusion method, system, equipment and terminal | |
| Zhang et al. | Single-image dehazing using extreme reflectance channel prior | |
| CN116137023B (en) | Low-illumination image enhancement method based on background modeling and detail enhancement | |
| Ponomaryov et al. | Fuzzy color video filtering technique for sequences corrupted by additive Gaussian noise | |
| CN112132757A (en) | General image restoration method based on neural network | |
| CN116823662A (en) | Image denoising and deblurring method fused with original features | |
| Guan et al. | NODE: Extreme low light raw image denoising using a noise decomposition network | |
| CN112767261B (en) | Non-local denoising frame for color image and video based on generalized non-convex tensor robust principal component analysis model | |
| Wang et al. | Removing image artifacts from scratched lens protectors | |
| Hsu et al. | Structure-transferring edge-enhanced grid dehazing network | |
| Kong et al. | A brief review of real-world color image denoising | |
| WO2020237366A1 (en) | System and method for reflection removal using dual-pixel sensor |
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 |