CN106651975B - A kind of Census adaptive transformation method based on odd encoder - Google Patents
A kind of Census adaptive transformation method based on odd encoder Download PDFInfo
- Publication number
- CN106651975B CN106651975B CN201611085880.2A CN201611085880A CN106651975B CN 106651975 B CN106651975 B CN 106651975B CN 201611085880 A CN201611085880 A CN 201611085880A CN 106651975 B CN106651975 B CN 106651975B
- Authority
- CN
- China
- Prior art keywords
- group
- queue
- census
- noise
- window
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/007—Transform coding, e.g. discrete cosine transform
Abstract
The invention belongs to technique of binocular stereoscopic vision fields, are related to a kind of Census adaptive transformation method based on odd encoder.The adaptive transformation method mainly improves to convert to traditional Census at two coding mode, noise control aspects, has used new coding strategy.Luminance information is more fully utilized while in view of local message in window, and decrease processing is carried out to noise, the present invention is made to have more robustness.
Description
Technical field
The invention belongs to technique of binocular stereoscopic vision fields, are related to a kind of adaptive transformation side Census based on odd encoder
Method.
Background technique
Stereo matching is always the important technical links in binocular stereo vision, and Stereo matching is found in two images
Conjugate point, that is, space same point are projected in the position of the different pixels point in the two images of left and right, finally generate disparity map.
But Stereo matching always exists the problem of precision and efficiency.Since the parallel ability of computer in recent years is higher and higher, because
This scientific research personnel has been devoted to find one kind can parallel high-precision solid matching method.
Usual Stereo matching process can be divided into four parts: matching cost calculates, matching cost polymerize, initial parallax meter
It calculates, parallax optimization.Research now for Stereo matching is also carried out all around this four steps, and Census transformation is three-dimensional
Matching cost calculates a kind of step commonly operation in matching process.
Census converts the local message fully considered center pixel in support window and weakens brightness of image letter
Breath.Have the characteristics that algorithm it is simple, can parallel, robustness height and effect it is good.The concrete operations of Census transformation are as described below:
It is w for a sizepThe window of (m × n), center pixel p, Census transformation can be by image in window
Brightness value relevance map is to a length in the binary string of m × n-1 length.Assuming that q is the non-central point in window
Pixel.So for the transformation of q point position are as follows:
Wherein i (x) is the gray scale of pixel x.The feature binary string final for p point:
Indicate step-by-step connection.Each of window traversal image using this calculation pixel.Such one
Image luminance information is just reduced, singly remains the local message in window.It is finally for the p point parallax in original image
D corresponds to the resulting cost of pixel p ' calculating in reference picture are as follows:
Indicate Hamming distance from calculating.
Although traditional Census transformation reduces luminance information, it is thick that it, which processes the decrease of luminance information,
Rough, there is no adequately utilize the luminance information in image.Referring to attached drawing 1 it can be seen that luminance information loss excessively causes
It is unobvious to similar area difference, and also will lead to it to noise-sensitive.
Summary of the invention
The present invention is to use new coding strategy on the basis of traditional Census transformation.In view of locally believing in window
Luminance information is more fully utilized while breath, and decrease processing is carried out to noise, and the present invention is made to have more robustness.
A kind of Census adaptive transformation method based on odd encoder, including following content:
1) it encodes
Traditional Census, which is converted, is divided into large and small two groups according to center pixel brightness value for luminance information.The present invention exists
On the basis of traditional Census converts large and small two groups, big group is further divided respectively with group, organize greatly with it is small
Group marks off k subgroup respectively, and k is positive integer, then marks off 2k group altogether.
The present invention uses a kind of odd encoder strategy, and each group carries out coded representation using multidigit binary digit.It marks off altogether
2k group, then each group of corresponding code length is 2k-1 position.It is w for a sizepThe window of (m × n);In window
Imago element is p, and non-central pixel is q in window;Brightness value is mapped to the position one (2 × k-1) × (m × n-1) in window
On the binary string of length.
The transformation of q point position are as follows:
Wherein, bin (x) is the binary representation of x;Non-central pixel grey scale i (q) in preceding k grouping is all
Greatly than center gray scale i (p), belong to big group, non-central pixel grey scale i (q) in rear k grouping is all than center gray scale i (p)
It is small, belong to group;εiIt is the threshold value for stepping, wherein [1,2...2k-2] i ∈;Threshold value is ε there are size relation1> ε2>
ε3... > εk-1, ε2k-2> ε2k-3... > εk。
Brightness of image is mapped to 2k group using the above strategy, and is differed after carrying out xor operation between adjacent gear positions
It is 1, difference is 2 after being mutually divided into 1 gear xor operation, and difference is 2k-1 after two farthest gear xor operations;Pass through
This mode present invention is by the significantly more efficient utilization of the correlation of brightness of image.
Above-mentioned threshold value selection uses a kind of adaptive threshold selection mode;Big group is done as follows respectively with group:
A) pixel in forms is greater than i (p) according to brightness and is divided into two groups of queue with i (p) is less than or equal to1And queue2。
B) to queue1With queue2It carries out ascending sort and generates sequence queue_asc1With queue_asc2And it calculates
queue1Length l1, queue2Length l2。
C) threshold epsilon is calculatediWherein the formula of i ∈ [1,2...k-1] is as follows:
D) threshold epsilon is calculatedjWherein the formula of j ∈ [k, k+1 ... 2k-2] is as follows:
The present invention realizes adaptive threshold value selection using this threshold strategies, can better choice threshold using the strategy
Value, uniformly assigns to each gear for image brightness information, and the adaptive strategy can make coding robustness more as far as possible
Height, and eliminate the interference of some noises.
Preferred embodiment is that luminance information in window is divided into k=2 group, marks off 4 groups altogether.Coding using three carrys out table
Show this four groups.It is w for a sizepThe window of (m × n), center pixel p, the present invention is by the correlation of brightness value in window
Property be mapped on the binary string of one 3 × (m × n-1) length, q be window in non-central point pixel.The transformation of q point position
Are as follows:
In the case where dividing four groups, there are ε1With ε2Two threshold values, sampling process are as follows:
1) pixel in forms is greater than i (p) according to brightness value and is divided into two groups of queue with i (p) is less than or equal to1With queue2。
2) to queue1With queue2It carries out ascending order arrangement and generates ascending sequence queue_asc1With queue_asc2, calculate
queue1Length is l1、queue2Length is l2。
3) it calculates
4) it calculates
By the group where non-central pixel-map to oneself in forms, non-central pixel coder step-by-step is connected and is generated
Central point p coding.
2) noise is handled
The present invention weakens the influence of noise using a kind of redundancy encoding mode, finally adds n in each pixel coder
Position redundancy encoding.The redundant code of original image is all set to 0, and the redundant code of reference picture is all set to 1, when existing in forms
Certain point brightness | i (p)-i (q) | > σ is it is considered that there are noises at one.σ is the threshold value being manually entered for judging noise.Whenever
It was found that the redundant code of image where noise is wherein negated for one at one when noise.N redundant codes at most can detect n noise
Point.
Original image pixels p is identical as reference picture pixel p ' coding and there is no its Hamming distances in the case where noise
From for n, when a certain pixel is by influence of noise, redundancy encoding changes, so that the redundant code of original image and reference picture
Hamming distance reduces, and the present invention is exactly by reducing the Hamming distance of redundant code from weakening the influence that noise generates it.
It is preferred that using a redundancy encoding, it is added 0 after original image pixels feature coding, in reference picture feature coding
1 is added below.The redundancy encoding after image where noise is negated when noise is detected, can only be examined using a redundancy encoding
Noise is measured with the presence or absence of existing amount of noise cannot be detected.
3) distance is calculated
The present invention continues to use Hamming distance as shown in formula (3) apart from calculation using traditional Census
From measuring similarity.
The present invention is on the basis of traditional Census transformation using new coding strategy, in view of locally believing in window
Luminance information is more fully utilized while breath, and decrease processing is carried out to noise, and the present invention is made to have more robustness.
Detailed description of the invention
Attached drawing 1 converts comparison diagram for the similitude region present invention and Census.
2 coding mode specific flow chart of attached drawing.
3 present invention of attached drawing converts final effect comparison diagram with tradition Census.
Specific embodiment
Embodiment 1
With reference to summary of the invention and attached drawing 2.The present invention specifically grasps in the case where adding a redundancy encoding at four groups of selection point
It is as follows to make process:
1) window size 5 × 5 is manually entered, input σ=100 judge for noise.
2) according to the invention vary one's tactics encodes each pixel.
Referring to attached drawing 2. for the coding mode of image point:
3) brightness data that size is 5 × 5 is read respectively to original image p point and reference picture p '.If reading out
Matrix is wpWith wp′
4) data read out are grouped according to p point brightness value i (p) and p ' brightness value i (p '), if each window
All there are two groupings for body: grouping 1 and grouping 2.
5) threshold epsilon is found out to the data after grouping1With ε2.It can acquire with reference to threshold calculations trifle of the invention for wpWindow
First group of 8 data of body, we select the 4 to 6th data to carry out threshold calculations.
(112-103+115-103+123-103) ÷ 3=13.7.Such as operate wpSecond group of data threshold of forms be
17.3。wp′First group of data threshold value be 9.5, wp′Second group of data threshold value be 19.
6) referring to coding mode chapters and sections of the invention, stepping operation is carried out to pixel in two forms respectively.After stepping
Forms are wcpAnd wcp′.Then to wcpAnd wcp′Carry out step-by-step connection output cen (p) and cen (p ').
7) redundancy check bit is added below in Bit String cen (p) and cen (p ') referring to noise processed chapters and sections of the invention,
Due to do not have noise we cen (p) add 0, add 1 afterwards in cen (p ').Form last coding.
Cen (p)=00,111,100,000,111,100,000,000,101,101,101,101,100,111,101,100,101,111,111 1111
1110111111110
Cen (p ')=00,011,100,000,111,100,000,001,111,111,101,101,100,101,101,101,100,101,111 111
11110111111111
After encoding operation using Hamming distance from come find out the Hamming distance of original image point p and reference picture point p ' from.
If the Hamming of figure is 9.
Referring to attached drawing 3, the present invention is substantially better than traditional approach in the case where purifying without using cost polymeric and parallax,
In the case where using perpendicular quadrature polymerization, the disparity map of generation will also be substantially better than traditional Census transformation.
Claims (3)
1. a kind of Census adaptive transformation method based on odd encoder, which is characterized in that
(1) it encodes
Census transformation point big group and group, then k subgroup is marked off respectively with group to big group, k is positive integer, is marked off altogether
2k group;Each group of corresponding code length is 2k-1 position, is w for a sizepThe window of (m × n);The center of window
Pixel is p, and non-central pixel is q in window;Brightness value in window is mapped to one (2 × k-1) × (m × n-1) bit length
On the binary string of degree;
The point position q varies one's tactics are as follows:
Wherein, bin (x) is the binary representation of x;During non-central pixel grey scale i (q) in preceding k grouping all compares
Heart gray scale i (p) greatly, belongs to big group, and non-central pixel grey scale i (q) in rear k grouping is all smaller than center gray scale i (p),
Belong to group;εiIt is the threshold value for stepping, wherein [1,2...2k-2] i ∈;Threshold value is ε there are size relation1> ε2> ε3…
> εk-1, ε2k-2> ε2k-3... > εk;
Brightness of image is mapped to 2k group using the above strategy, and difference is 1 after progress xor operation between adjacent gear positions,
Difference is 2 after being mutually divided into 1 gear xor operation, and difference is 2k-1 after two farthest gear xor operations;
Above-mentioned threshold value selection uses a kind of adaptive threshold selection mode, i.e., is done as follows respectively to big group with group:
A. pixel in forms is greater than i (p) according to brightness and is divided into two groups of queue with i (p) is less than or equal to1And queue2;
B. to queue1With queue2It carries out ascending sort and generates sequence queue_asc1With queue_asc2And calculate queue1It is long
Spend l1, queue2Length l2;
C. threshold epsilon is calculatedi, wherein the formula of i ∈ [1,2...k-1] is as follows:
D. threshold epsilon is calculatedj, wherein the formula of j ∈ [k, k+1 ... 2k-2] is as follows:
(2) noise is handled
N redundancy encodings are finally added in each pixel coder;The redundant code of original image is all set to 0, reference picture it is superfluous
Complementary is all set to 1;When there are certain to put brightness in forms | i (p)-i (q) | then there is noise at one in > σ, wherein σ is artificial
Input the threshold value for judging noise;When finding noise at one, the redundant code of image where noise is wherein negated for one;n
Position redundant code can detect n noise spot;
(3) distance is calculated
Using Census apart from calculation, and using Hamming distance from measurement similarity.
2. a kind of Census adaptive transformation method based on odd encoder according to claim 1, which is characterized in that described to big group
The k value for marking off k subgroup respectively with group is 2, marks off 4 groups altogether.
3. a kind of Census adaptive transformation method based on odd encoder according to claim 1, which is characterized in that redundancy encoding position
The value of number n is 1.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611085880.2A CN106651975B (en) | 2016-12-01 | 2016-12-01 | A kind of Census adaptive transformation method based on odd encoder |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611085880.2A CN106651975B (en) | 2016-12-01 | 2016-12-01 | A kind of Census adaptive transformation method based on odd encoder |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106651975A CN106651975A (en) | 2017-05-10 |
CN106651975B true CN106651975B (en) | 2019-08-13 |
Family
ID=58814729
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611085880.2A Expired - Fee Related CN106651975B (en) | 2016-12-01 | 2016-12-01 | A kind of Census adaptive transformation method based on odd encoder |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106651975B (en) |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101072366A (en) * | 2007-05-24 | 2007-11-14 | 上海大学 | Free stereo display system and method based on light field and binocular vision technology |
CN102136136A (en) * | 2011-03-17 | 2011-07-27 | 南京航空航天大学 | Luminosity insensitivity stereo matching method based on self-adapting Census conversion |
CN102970545A (en) * | 2012-12-11 | 2013-03-13 | 东南大学 | Static image compression method based on two-dimensional discrete wavelet transform algorithm |
WO2014193418A1 (en) * | 2013-05-31 | 2014-12-04 | Hewlett-Packard Development Company, L.P. | Three dimensional data visualization |
CN105894499A (en) * | 2016-03-25 | 2016-08-24 | 华南理工大学 | Binocular-vision-based rapid detection method for three-dimensional information of space object |
-
2016
- 2016-12-01 CN CN201611085880.2A patent/CN106651975B/en not_active Expired - Fee Related
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101072366A (en) * | 2007-05-24 | 2007-11-14 | 上海大学 | Free stereo display system and method based on light field and binocular vision technology |
CN102136136A (en) * | 2011-03-17 | 2011-07-27 | 南京航空航天大学 | Luminosity insensitivity stereo matching method based on self-adapting Census conversion |
CN102970545A (en) * | 2012-12-11 | 2013-03-13 | 东南大学 | Static image compression method based on two-dimensional discrete wavelet transform algorithm |
WO2014193418A1 (en) * | 2013-05-31 | 2014-12-04 | Hewlett-Packard Development Company, L.P. | Three dimensional data visualization |
CN105894499A (en) * | 2016-03-25 | 2016-08-24 | 华南理工大学 | Binocular-vision-based rapid detection method for three-dimensional information of space object |
Also Published As
Publication number | Publication date |
---|---|
CN106651975A (en) | 2017-05-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105634499B (en) | Data conversion method based on new short floating point type data | |
KR940010959B1 (en) | Coding device and method | |
CN103400174B (en) | The coded method of a kind of Quick Response Code, coding/decoding method and system | |
JP7127137B2 (en) | Encoding method, decoding method and apparatus | |
CN1133329C (en) | Method and apparatus for encoding contour of object based on contour motion estimation technique | |
CN108846873A (en) | A kind of Medical Image Lossless Compression method based on gray probability | |
CN113763269B (en) | Stereo matching method for binocular images | |
WO2010135082A1 (en) | Localized weak bit assignment | |
CN106560010A (en) | The efficient huffman coding apparatus and method of VLSI | |
CN112070821A (en) | Low-power-consumption stereo matching system and method for acquiring depth information | |
CN103632182A (en) | Three-dimensional (3D) image code generating and analyzing methods and device thereof | |
CN102095439A (en) | Single-code-channel absolute-position encoding method, decoding method and measuring device | |
Paul et al. | Design of absolute encoder disk coding based on affine n digit N-ary gray code | |
CN109495211A (en) | A kind of channel coding and coding/decoding method | |
CN106651975B (en) | A kind of Census adaptive transformation method based on odd encoder | |
EP4199520A1 (en) | Point cloud attribute encoding and decoding method and device | |
CN111695276B (en) | Finite element model proxy modeling method based on generation countermeasure learning | |
CN1841939A (en) | Encoder and decoder using run-length-limited code | |
CN1941910B (en) | Method and apparatus for encoding coordinate built in symbol and decoding bit data flow | |
CN115866251B (en) | Image information rapid transmission method based on semantic segmentation | |
CN1261475A (en) | Method and device for source-controlled channel decoding using a kalman filter | |
CN1189061A (en) | Method and apparatus for encoding contour of object by using vertex inserting technique | |
CN114136351B (en) | Single code channel absolute coding and decoding method with error detection and correction function | |
CN107609059B (en) | Chinese domain name similarity measurement method based on J-W distance | |
CN102589442B (en) | Single-code coding and decoding method based on absolute position determination |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190813 Termination date: 20211201 |