CN106023272B - Three-dimensional Self-organizing Maps image encoding method based on new learning function - Google Patents
Three-dimensional Self-organizing Maps image encoding method based on new learning function Download PDFInfo
- Publication number
- CN106023272B CN106023272B CN201610316567.9A CN201610316567A CN106023272B CN 106023272 B CN106023272 B CN 106023272B CN 201610316567 A CN201610316567 A CN 201610316567A CN 106023272 B CN106023272 B CN 106023272B
- Authority
- CN
- China
- Prior art keywords
- vector
- pattern
- trained
- image
- encoded
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/001—Model-based coding, e.g. wire frame
Abstract
The present invention discloses a kind of three-dimensional Self-organizing Maps image encoding method based on new learning function, firstly, carrying out continuous learning training to image pattern, obtains best match pattern base;Then, piecemeal processing is carried out to image to be encoded, and the pattern vector in these image blocks to be encoded and best match pattern base is subjected to image model matching, encode the index of each piece of match pattern in pattern base.There is the present invention pattern base to regulate the speed fastly, the high feature of pattern base training effectiveness.
Description
Technical field
The present invention relates to image coding technology fields, and in particular to a kind of three-dimensional self-organizing based on new learning function is reflected
Penetrate image encoding method.
Background technique
Image coding is also referred to as compression of images, refers to and is meeting certain mass (requirement or subjective assessment score of signal-to-noise ratio)
Under the conditions of, the technology of information included in image or image is indicated with less bit.Image neural network based encodes
The new image encoding method of the one kind occurred in recent years.Application of the artificial neural network in compression of images is to pass through Preliminary Simulation
The local initial function of nerve system of human body, the characteristics of according to image itself, autonomous carries out compressed encoding to it.Self-organizing mind
It is widely used through network in image coding neighborhood.
Self-organizing Maps (Self-Organizing Map, abbreviation SOM) algorithm is group of the Kohonen from human brain neuron
Knit a kind of artificial neural network algorithm with Self-organization for being inspired and being proposed in principle.Algorithm simulation human brain
Nervous system considers the condition that can manually realize to the feature of a certain figure or the specific excitement of a certain frequency, in number
In-depth study and application have been obtained with fields such as image steganalysis according to excavating, has been a kind of efficient data clustering method.SOM
Network uses double-layer network structure, is divided into input layer and mapping layer, input layer is for perceiving input pattern, each input neuron
It is connect by weight with each mapping layer neuron, mapping layer exports as a result, the neuron of mapping layer interconnects.Traditional SOM
Network generally uses one-dimensional input layer and two-dimensional map layer, it can effectively handle a peacekeeping 2D signal.In recent years, three-dimensional figure
The processing of the three dimensional signals such as picture and video increasingly causes concern, however tradition SOM algorithm can not be directly used in three dimensional signal
Processing.Three-dimensional S OM algorithm has well solved this problem, two dimension input effectively can be mapped as three-dimensional output, realized
The Nonlinear Mapping of three dimensional signal.
Learning function is the important component of SOM algorithm, in order to guarantee to restrain, need to meet and be gradually reduced with iteration time
Principle.Learning function usually takes linear function or exponential function.However, there are some problems, such as mode for these learning functions
Library is regulated the speed slowly, and pattern base training effectiveness is low etc., affects the performance of three-dimensional S OM algorithm.
Summary of the invention
Regulating the speed the technical problem to be solved by the present invention is to existing SOM algorithm pattern library, slow and training effectiveness is low etc. to ask
Topic, provides a kind of three-dimensional Self-organizing Maps image encoding method based on new learning function.
To solve the above problems, the present invention is achieved by the following technical solutions:
Three-dimensional Self-organizing Maps image encoding method based on new learning function, includes the following steps:
Step 1 carries out continuous learning training to image pattern, obtains best match pattern base;
Step 1.1 carries out piecemeal to image pattern, and each image block is a trained vector, is obtained containing L training arrow
The trained vector collection of amount;
Step 1.2 selects N number of trained vector to constitute initialization pattern library from trained vector concentration, the initial pattern base
In trained vector be known as pattern vector, and the pattern vector in initial pattern base is arranged in three-dimensional structure;Above-mentioned N < < L;
Step 1.3, setting pace of learning function;
Wherein, B0Maximum pace of learning when starting for training, C1To learn attenuation constant;T=0,1 ..., L-1, L are instruction
Practice the number of trained vector in vector set;
One step 1.4, input trained vector, and calculate separately each mode in the trained vector and initial pattern base
The distortion of vector is therefrom selected and is distorted the smallest pattern vector as triumph pattern vector with trained vector;
Pattern vector within the scope of step 1.5, according to the following formula adjustment triumph pattern vector and its three dimensional neighborhood;
Wherein, WjIt (t+1) is the t+1 times pattern vector when trained, WjIt (t) is the t times pattern vector when trained, X
It (t) is trained vector, j*For triumph pattern vector,For triumph pattern vector j*Neighborhood function when the one t times trained, α
It (t) is the t times pace of learning function when trained;J=0,1 ..., N-1, N are the big of the three-dimensional self-organized mapping network of setting
It is small;T=0,1 ..., L-1, L are the number of trained vector concentration training vector;
Step 1.6, return step 1.4 reselect a trained vector and obtain until having inputted all trained vectors
Best match pattern base;
Step 2 carries out piecemeal processing to image to be encoded, and by these image blocks to be encoded and best match pattern base
In pattern vector carry out image model matching, encode the index of each piece of match pattern in pattern base.
In above-mentioned steps 1.2, is concentrated using randomized from trained vector and N number of trained vector is selected to constitute initial mode
Library.
In above-mentioned steps 1.4, the distortion d of trained vector and pattern vector is calculated using following formulaj(t);
dj(t)=| | X (t)-Wj(t)||2
Wherein, WjIt (t) is pattern vector, X (t) is trained vector;J=0,1 ..., N-1, N are the three-dimensional self-organizing of setting
The size of mapping network;T=0,1 ..., L-1, L are the number of trained vector concentration training vector.
In above-mentioned steps 1.5, triumph pattern vector j*Neighborhood function are as follows:
Wherein, A0For triumph pattern vector j*Minimum neighborhood, A1For triumph pattern vector j*Maximum neighborhood, T1For neighborhood
Attenuation constant, t=0,1 ..., L-1, L are the number of trained vector concentration training vector.
Above-mentioned steps 2 specifically:
Step 2.1, by image block to be encoded, wherein the piecemeal of image block size and image pattern to be encoded is big
It is small consistent;
Step 2.2, the distortion for calculating separately each pattern vector in each image block to be encoded and best match pattern base,
And encode the index that there is the pattern vector of minimum distortion in pattern base;
Step 2.3, the processing that all images to be encoded are carried out with step 2.2, obtain the code stream of image to be encoded.
Compared with prior art, the present invention has a characteristic that pattern base is regulated the speed fastly, and pattern base training effectiveness is high.
Detailed description of the invention
Fig. 1 is the training process of pattern base.
Fig. 2 is three-dimensional Self-organizing Maps image encoding process.
Specific embodiment
Below with reference to embodiment, the content of present invention is described in further detail, but embodiments of the present invention are unlimited
In this.
A kind of three-dimensional Self-organizing Maps image encoding method based on new learning function, includes the following steps:
Step 1) constantly train to image pattern, obtains the pattern base of best match.Referring to Fig. 1.
Step 1.1) sets 3DSOM network size as (N, M), and wherein N, M are respectively the size of pattern base, mould in pattern base
The size of formula vector.
Step 1.2) generally takes 8 × 8 to image pattern piecemeal, the size of block.Each image block is a trained vector, altogether
Obtain the trained vector collection { X (t) containing L trained vector;T=0,1 ..., L-1 }.Selection N is concentrated from trained vector with randomized
(N < < L) a trained vector constitutes initialization pattern library { Wj(0);J=0,1 ..., N-1 }, and will be in initial pattern base
Pattern vector is arranged in three-dimensional structure.
Step 1.3) is set as N for neighborhood is initializedj(0), j=0,1 ..., N-1.
Step 1.4) inputs the trained vector X=(x that a trained vector is concentrated1,x2,…,xM)T。
Step 1.5) selects mean square error distortion criterion, mean square error dj(t)=| | X (t)-Wj(t)||2, calculate separately
The distortion d of each pattern vector in the trained vector and pattern basej(t);And the pattern vector with minimum distortion is selected to win
Pattern vector j*。
Step 1.6) adjusts triumph pattern vector j as the following formula*And its three dimensional neighborhood Nj* the pattern vector in (t) range,
Wherein,For neighborhood function, exponential function is selectedA0、A1Respectively triumph mode
Vector j*Minimum neighborhood and maximum neighborhood, T1For neighborhood attenuation constant.In the trained initial stage, the radius of neighbourhood is larger, with
The increase of frequency of training, network gradually tend towards stability, and subtleer weighed value adjusting, thus neighborhood half need to be only carried out to winning node
Diameter constantly reduces.α (t) is pace of learning function, it reflects the amplitude size of pattern vector adjustment, uses function
Wherein, B0Maximum pace of learning when starting for training, C1To learn attenuation constant.
Step 1.7) return step step 1.4) concentrates all trained vectors until having inputted trained vector to get arriving
Best match pattern base.
Step 2) treats coded image and carries out pattern-recognition, encodes the index of each piece of match pattern in pattern base.
Referring to fig. 2.
Step 2.1, by image block to be encoded, wherein the piecemeal size one of image block size to be encoded and image pattern
It causes;
Step 2.2, the distortion for calculating separately each pattern vector in each image block to be encoded and best match pattern base,
And match pattern of the pattern vector with minimum distortion as the image block is selected, it is only encoded when to the image block coding
Index with mode in pattern base;
Each pattern vector in image block to be encoded and best match pattern base is calculated using mean-square error criteria, that is, following formula
Distortion bj(t′);
bj(t ')=| | Y (t ')-Wj(t′)||2
Wherein, Wj(t ') is pattern vector, and Y (t ') is image block to be encoded;J=0,1 ..., N-1, N are the three-dimensional of setting
The size of self-organized mapping network;T '=0,1 ..., K-1, K are the number of image block to be encoded.
Step 2.3 treats the processing that each subimage block of coded image carries out step 2.2, completes the volume to image
Code.
Specifically, in order to illustrate the performance of this method, done a large amount of emulation experiment, experimental image select resolution ratio for
Normal brightness test image Lena, Peppers and Boat of 512 × 512 × 8bit, for pattern base training and image coding.
Reconstruction image quality is objectively evaluated using Y-PSNRE in formulaMSEFor original image and rebuild
Mean square error between image.Image compression rate isM is the dimension of pattern vector, B in formulaOFor original image
Every pixel bit number, BCFor pattern vector number of address bits.
Experiment is respectively with the three-dimensional S OM algorithm based on index learning function and the three-dimensional based on the new learning function of this patent
SOM algorithm designed image pattern-recognition pattern base, and compared by the quality of reconstruction image after coding based on both study
The performance of the three-dimensional S OM algorithm of function, wherein trained vector number is 40960, and pattern base size is 1024.Table 1 is to choose difference
When experimental image, the three-dimensional S OM algorithm based on index learning function and the three-dimensional S OM algorithm based on the new learning function of this patent
The Y-PSNR (PSNR) of reconstruction image for image steganalysis coding.
The PSNR of 1 reconstruction image of table
As can be seen from Table 1, the performance ratio of the three-dimensional S OM algorithm based on the new learning function of this patent is learnt based on index
The three-dimensional S OM algorithm of function improves 0.181-0.3464dB.
Claims (5)
1. the three-dimensional Self-organizing Maps image encoding method based on new learning function, characterized in that include the following steps:
Step 1 carries out continuous learning training to image pattern, obtains best match pattern base;
Step 1.1 carries out piecemeal to image pattern, and each image block is a trained vector, is obtained containing L trained vector
Trained vector collection;
Step 1.2 selects N number of trained vector to constitute initial pattern base from trained vector concentration, the instruction in the initial pattern base
Practice vector and be known as pattern vector, and the pattern vector in initial pattern base is arranged in three-dimensional structure;Above-mentioned N < < L;
Step 1.3, setting pace of learning function;
Wherein, B0Maximum pace of learning when starting for training, C1To learn attenuation constant;T=0,1 ..., L-1, L are training arrow
The number of trained vector in quantity set;
One step 1.4, input trained vector, and calculate separately each pattern vector in the trained vector and initial pattern base
Distortion, therefrom select and be distorted the smallest pattern vector as triumph pattern vector with trained vector;
Pattern vector within the scope of step 1.5, according to the following formula adjustment triumph pattern vector and its three dimensional neighborhood;
Wherein, WjIt (t+1) is the t+1 times pattern vector when trained, WjIt (t) is the t times pattern vector when trained, X (t) is
Trained vector, j*For triumph pattern vector,For triumph pattern vector j*Neighborhood function when the one t times trained, α (t) are
The t times it is trained when pace of learning function;J=0,1 ..., N-1, N are the size of the three-dimensional self-organized mapping network of setting;T=
0,1 ..., L-1, L are the number of trained vector concentration training vector;
Step 1.6, return step 1.4 reselect a trained vector, until having inputted all trained vectors, obtain best
Match pattern library;
Step 2 carries out piecemeal processing to image to be encoded, and will be in these image blocks to be encoded and best match pattern base
Pattern vector carries out image model matching, encodes the index of each piece of match pattern in pattern base.
2. the three-dimensional Self-organizing Maps image encoding method based on new learning function according to claim 1, characterized in that
In step 1.2, is concentrated using randomized from trained vector and N number of trained vector is selected to constitute initial pattern base.
3. the three-dimensional Self-organizing Maps image encoding method based on new learning function according to claim 1, characterized in that
In step 1.4, the distortion d of trained vector and pattern vector is calculated using following formulaj(t);
dj(t)=| | X (t)-Wj(t)||2
Wherein, WjIt (t) is the t times pattern vector when trained, X (t) is trained vector;J=0,1 ..., N-1, N are the three of setting
Tie up the size of self-organized mapping network;T=0,1 ..., L-1, L are the number of trained vector concentration training vector.
4. the three-dimensional Self-organizing Maps image encoding method based on new learning function according to claim 1, characterized in that
In step 1.5, triumph pattern vector j*Neighborhood function when the one t times trainedAre as follows:
Wherein, A0For triumph pattern vector j*Minimum neighborhood, A1For triumph pattern vector j*Maximum neighborhood, T1For neighborhood decaying
Constant, t=0,1 ..., L-1, L are the number of trained vector concentration training vector.
5. the three-dimensional Self-organizing Maps image encoding method based on new learning function according to claim 1, characterized in that
Step 2 specifically:
Step 2.1, by image block to be encoded, wherein the piecemeal size phase of image block size to be encoded and image pattern
Unanimously;
Step 2.2, the distortion for calculating separately each pattern vector in each image block to be encoded and best match pattern base, and compile
Code has index of the pattern vector of minimum distortion in pattern base;
Step 2.3, the processing that all images to be encoded are carried out with step 2.2, obtain the code stream of image to be encoded.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610316567.9A CN106023272B (en) | 2016-05-13 | 2016-05-13 | Three-dimensional Self-organizing Maps image encoding method based on new learning function |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610316567.9A CN106023272B (en) | 2016-05-13 | 2016-05-13 | Three-dimensional Self-organizing Maps image encoding method based on new learning function |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106023272A CN106023272A (en) | 2016-10-12 |
CN106023272B true CN106023272B (en) | 2019-03-19 |
Family
ID=57100734
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610316567.9A Active CN106023272B (en) | 2016-05-13 | 2016-05-13 | Three-dimensional Self-organizing Maps image encoding method based on new learning function |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106023272B (en) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101221662A (en) * | 2008-01-31 | 2008-07-16 | 复旦大学 | Remote sensing image mixed image element decomposition method based on self-organizing mapping neural network |
CN103763565A (en) * | 2014-01-24 | 2014-04-30 | 桂林电子科技大学 | Anaglyph coding method based on three-dimensional self-organizing mapping |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
AU2011253255B2 (en) * | 2010-05-08 | 2014-08-14 | The Regents Of The University Of California | Method, system, and apparatus for pressure image registration |
-
2016
- 2016-05-13 CN CN201610316567.9A patent/CN106023272B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101221662A (en) * | 2008-01-31 | 2008-07-16 | 复旦大学 | Remote sensing image mixed image element decomposition method based on self-organizing mapping neural network |
CN103763565A (en) * | 2014-01-24 | 2014-04-30 | 桂林电子科技大学 | Anaglyph coding method based on three-dimensional self-organizing mapping |
Non-Patent Citations (4)
Title |
---|
《一种新的基于自学习神经网络的静止图像编码方案》;黎洪松等;《北京师范大学学报》;20061031;第42卷(第5期);全文 |
《一种新的基于自组织神经网络的运动估计算法》;黎洪松等;《湖南大学学报》;20060831;第33卷(第4期);全文 |
《一种新的自学习特征映射算法研究》;黎洪松等;《北京师范大学学报》;20060630;第42卷(第3期);全文 |
《新的学习矢量量化初始码书算法》;黎洪松等;《北京邮电大学学报》;20060831;第29卷(第4期);全文 |
Also Published As
Publication number | Publication date |
---|---|
CN106023272A (en) | 2016-10-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109543745A (en) | Feature learning method and image-recognizing method based on condition confrontation autoencoder network | |
CN110290387B (en) | Image compression method based on generative model | |
CN107680077A (en) | A kind of non-reference picture quality appraisement method based on multistage Gradient Features | |
CN110191299A (en) | A kind of multiplex frame interpolation method based on convolutional neural networks | |
CN110163815A (en) | Low-light (level) restoring method based on multistage variation self-encoding encoder | |
CN103971329B (en) | A kind of multisource image anastomosing method based on genetic optimization cell neural network | |
CN107172428B (en) | The transmission method of image, device and system | |
CN111861945B (en) | Text-guided image restoration method and system | |
CN109996073B (en) | Image compression method, system, readable storage medium and computer equipment | |
CN106815876B (en) | Image sparse characterizes the combined optimization training method of more dictionary learnings | |
CN109949222A (en) | Image super-resolution rebuilding method based on grapheme | |
CN111797891A (en) | Unpaired heterogeneous face image generation method and device based on generation countermeasure network | |
CN110647891B (en) | CNN (convolutional neural network) -based automatic extraction method and system for time sequence data characteristics of self-encoder | |
CN114897694A (en) | Image super-resolution reconstruction method based on mixed attention and double-layer supervision | |
Yoo et al. | Real-time semantic communications with a vision transformer | |
Muruganandham et al. | Adaptive fractal image compression using PSO | |
CN109889848A (en) | Based on the multiple description coded of convolution self-encoding encoder, coding/decoding method and system | |
CN106028043B (en) | Three-dimensional Self-organizing Maps image encoding method based on new neighborhood function | |
CN116600119B (en) | Video encoding method, video decoding method, video encoding device, video decoding device, computer equipment and storage medium | |
He et al. | End-to-end facial image compression with integrated semantic distortion metric | |
Li et al. | Steganography of steganographic networks | |
CN106023272B (en) | Three-dimensional Self-organizing Maps image encoding method based on new learning function | |
CN108259914B (en) | Cloud image encoding method based on object library | |
US8285053B2 (en) | Codebook generating method | |
CN116137043A (en) | Infrared image colorization method based on convolution and transfomer |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |