CN110503612B - Bit plane based data separation and recombination enhancement method - Google Patents
Bit plane based data separation and recombination enhancement method Download PDFInfo
- Publication number
- CN110503612B CN110503612B CN201910700184.5A CN201910700184A CN110503612B CN 110503612 B CN110503612 B CN 110503612B CN 201910700184 A CN201910700184 A CN 201910700184A CN 110503612 B CN110503612 B CN 110503612B
- Authority
- CN
- China
- Prior art keywords
- bit
- image
- pixel
- plane
- bit plane
- 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
- 238000005215 recombination Methods 0.000 title claims abstract description 28
- 230000006798 recombination Effects 0.000 title claims abstract description 28
- 238000000926 separation method Methods 0.000 title claims abstract description 19
- 238000000034 method Methods 0.000 title claims abstract description 16
- 230000000007 visual effect Effects 0.000 claims abstract description 14
- 239000011159 matrix material Substances 0.000 claims description 21
- 238000004458 analytical method Methods 0.000 claims description 7
- 238000010586 diagram Methods 0.000 claims description 3
- 229910002056 binary alloy Inorganic materials 0.000 claims 1
- 230000003321 amplification Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 238000003199 nucleic acid amplification method Methods 0.000 description 3
- 238000010835 comparative analysis Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
Images
Classifications
-
- 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
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/60—Rotation of a whole image or part thereof
-
- G06T5/90—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/90—Determination of colour characteristics
Abstract
The invention provides a bit plane based data enhancement method for separating and recombining, which solves the problems of the unsolvable scenes in the data enhancement in the field of the traditional computer vision. The method utilizes a bit-level 0-1 binary image of a bit plane to carry out bit plane separation and a certain proportion recombination, the order of the bit plane is disturbed in the recombination, new pixel points of a gray channel or an RGB channel are generated, the original basic visual effect of the image is kept through some specific modes, and new enhanced image data are generated so as to expand an image data set. The invention is also compatible with the traditional image turning, rotating, zooming, cutting, translating, amplifying and other modes, greatly increases the number of effective image data, improves the robustness of a data set, generates more universal results and solves the problem of serious data shortage in computer vision to a great extent.
Description
Technical Field
The invention relates to a bit plane separation and recombination computer vision image data enhancement technology. In particular to a bit plane scrambling and bit plane pixel area acquisition technology, namely a bit plane separation and recombination based data enhancement method.
Background
Computer vision has various machine learning and deep learning tasks, generally, these machine learning tasks need a huge image data set as learning training data, however, the situation that the number of image data sets is insufficient usually occurs, which causes a serious overfitting phenomenon of a deep learning network, in order to solve the overfitting phenomenon, the common practice in the industry is data enhancement, that is, a series of images are transformed by operating the images of the training data set, so as to create more data, and thus a network with stronger generalization capability can be obtained.
In addition, the conventional computer vision uses the conventional data enhancement mode that only a pixel matrix is moved in an image, and does not generate new pixel points or change the pixel sequence of the image, namely, does not generate a new image. The situation of overfitting cannot be effectively solved, the bit plane separation and recombination adopted by the invention separates the image into the binary matrix of the bit plane, then carries out certain processing on the image at the bit level, and then strengthens the data by certain scrambling sequence combination under the condition of keeping certain image visual effect, thus being better applied to the field of computer vision.
Existing data enhancement techniques for pictures are typically by flipping, rotating, scaling and cropping, translating, etc. the picture. And different special data enhancement modes are available for different computer vision tasks, for example, a noise adding mode is adopted for some tasks in image recognition to obtain more characteristics through network learning and abandon invalid characteristics, or a super-resolution image amplification task in the super-resolution task needs to amplify an enhanced data set image matrix. The modes effectively increase the robustness of the data set, but have the defects that the data enhancement modes are all image-level changes, the pixel matrix of the image is only moved, new pixel points are not generated, and the pixel sequence of the image is not changed, namely, a new image is not generated. Sometimes, too many similar images generated by the conventional data enhancement technology can be trapped in a situation that not only the overfitting cannot be solved, but also more overfitting is generated.
Disclosure of Invention
The invention provides a bit plane based data enhancement method for separating and recombining, which solves the problems of the unsolvable scenes in the data enhancement in the field of the traditional computer vision. The method utilizes a bit-level 0-1 binary image of a bit plane to carry out bit plane separation and a certain proportion recombination, the order of the bit plane is disturbed in the recombination, new pixel points of a gray channel or an RGB channel are generated, the original basic visual effect of the image is kept through some specific modes, and new enhanced image data are generated so as to expand an image data set.
The technical solution adopted by the invention is as follows:
a bit-plane-based method for separating and recombining data enhancement comprises the following steps:
(1) and separating color channels of an RGB three-channel color image to obtain three single-channel images of R, G and B. Since the pixel values for RGB are in the range of 0-255, there are at most two octaves of binary values, with each bit being given the formula bm-12m-1+bm-22m-2…+ b020(m-8) are developed separately, and 8 bit planes can be obtained; the three graphs are subjected to bit plane separation, and each graph is divided into 8 bit planes, namely 8 binary image matrixes. Where b represents the binary pixel of each bit plane, 2 represents each bin, and m represents 8 bit planes.
(2) Analyzing the image bit-plane diagram to determine the importance of the visual effect of the source image in each bit-plane, which is arranged from small to large in bits of 1 to 8 bits, and the eighth bit-plane is 2 pixels in binary7～28When the eighth bit is 1, the largest pixel value is 11111111(255), and the smallest pixel value is 10000000 (128); when 0 is taken as the eighth bit, the maximum pixel value is 01111111(127), the minimum pixel value is 00000000(0), the decimal range of the source image corresponding to 0/1 of the eighth pixel is (0-127)/(128-255), and thus, the decimal range of the source image corresponding to 0/1 of the seventh pixel is (0-191)/(64-255). It can be seen that the aliased pixel area of the eighth bit is 0, and the aliased pixel area of the seventh bit is (64 ~ 191). The sixth bit is analyzed to obtain the region of the confused pixel (32-223).
It can be seen that, every time the bit plane with the next bit smaller is reached, the image confusion pixel area is larger, in order to ensure the visual effect of the image, but not change the image into a noise pixel matrix after separation and recombination, the invention adopts recombination 678 bit planes or carries out image recombination analysis by extracting the largest pixel range ratio through code analysis pixel ratio, so as to obtain data enhancement which ensures a certain visual effect.
First, an image to be read is in a matrix format, and then different bits of the image are stored in a newly created binary _ matrix by cyclically reading the different bits of the image. Thus, the binary _ matrix stores the bit plane 0-1 matrix of eight channels, as shown in fig. 1, fig. 2, and fig. 3, the bit planes of the images of three channels are separated to obtain 24 bit plane images, and then there are two replacement order modes, the first mode is to adopt a default scrambling order, that is, the aforementioned minimum aliasing pixel region, and adopt three minimum aliasing pixel regions for scrambling, that is, 678 bit planes, the combination can be six recombination modes of 12345678, 12345687, 12345768, 12345786, 12345867, and 12345876, and the second mode is to extract the largest pixel range for image recombination analysis by observing the visual structure of the image itself, and analyzing the pixel ratio through code analysis, so as to ensure that the recombination visual effect can maintain the former source image structure.
And recombining after image bit plane separation to obtain a new RGB image which is put into a data set to be used as new training data.
The bit plane based data enhancement method of the invention enhances the image data set by using the separation and recombination of the bit plane, can increase the new pixel image matrix of different channels of gray scale or RGB while maintaining the visual effect, avoids the over-fitting phenomenon when the image data set is used for machine learning or deep learning, is compatible with the traditional image turning, rotating, zooming, clipping, translating, amplifying and other modes, greatly increases the quantity of effective image data, improves the robustness of the data set, generates more universal results, and solves the problem of serious data shortage in computer vision to a great extent.
Drawings
The 8 bit planes (high to low) of the R channel image of FIG. 1.
Fig. 2G channel image 8 bit planes (high to low).
The 8 bit planes (high to low) of the channel image of fig. 3B.
FIG. 4 is a flow diagram illustrating a default shuffle order image bit-plane separation and recombination process.
FIG. 5 is a flow chart illustrating the process of separating and recombining the bit planes of the self-defined scrambled sequential image.
Detailed Description
The present invention will be described in further detail below with reference to the accompanying drawings.
Step 1 image bit-plane separation
Reading a color image matrix of an RGB channel, performing bit plane separation operation on the image to obtain 3 single-channel images RGB, or reading a YCrCb color image matrix, performing bit plane separation on a Y-channel image only, and only recombining the Y channel during recombination to ensure algorithm speed when new data is created. Applying binary bit-plane separation formula bm- 12m-1+bm-22m-2…+b020(m-8), a bit-plane partition of 3 single-channel images can be obtained, resulting in a 24-bit-plane image binary matrix.
Step 2, enhancing general data in the computer vision field
Each different computer vision task has different data enhancement requirements, but basically has several basic requirements: turning, rotating, zooming, cutting, translating and super-resolution. According to the requirements, the operations can be carried out on the bit plane so as to meet the structural change of the image data on the bit plane level, and the data enhancement requirement can also be carried out after the image recombination. For the 8 bit-plane image matrices, image [:]the flip in the weight direction can be completed, and the image [ < in-1 >:]the turning in the height direction can be completed, and the replacement of the height and weight channels of the matrix can complete the turning effect of rot90 degrees. For the task of amplification, pixel arrangement combination similar to pixelschuffle is required, and the task of integral multiple amplification can be completed. If the shape of the bit-plane image matrix is h x w, the task is to enlarge the matrix by r times, and the same bit-plane matrix r needs to be copied2The same pixel is filled in the enlarged position (p × r ) of each pixel p.
Step 3, sequential scrambling
The first method adopts a default scrambling sequence, namely the minimum confusion pixel area mentioned before, the confusion pixel area of the eighth bit is 0, the confusion pixel area of the seventh bit is (64-191), the confusion pixel area of the sixth bit is (32-223) and adopts three minimum confusion pixel areas for scrambling, the combination can be six recombination methods of 12345678, 12345687, 12345768, 12345786, 12345867 and 12345876, and the second method obtains the first three or four pixel points with the maximum ratio by observing the visual structure of the image and analyzing the pixel ratio, wherein the first three pixel points p are taken1，p2，p3. According to 2n-1≤pm≤2nObtaining the region n of the pixel range bit plane1，n2，n3Then by the formula
And performing image recombination analysis to ensure that the original image structure can be maintained by the recombination visual effect as much as possible.
Step 4 bit plane image recombination
B is formed bym-12m-1+bm-22m-2…+b020(m-8) and the previously obtained default scrambling order bm-1，bm-2，bm-3Or a custom scrambling sequence n1，n2，n3B of1，b2，b3And combining the data into new image matrix data to be used as new training data to be put into a data set.
2. Bit-plane separation and recombination data enhancement and common data enhancement are used for carrying out contrast analysis
The bit plane separation and recombination data enhancement and the common data enhancement are compared and analyzed in the super-resolution field of computer vision, wherein the super-resolution adopts two different data sets Set5, DIV2 k:
TABLE 1 comparative analysis
As can be seen from the comparative analysis, the PSNR index of the super-resolution task improves the index. The bit plane data enhancement has certain promotion on the diversity of the data, and the richness of the data is promoted.
Claims (1)
1. A bit-plane-based method for separating and recombining data enhancement is characterized by comprising the following steps:
(1) separating color channels of an RGB three-channel color image to obtain R, G, B single-channel images; each binary bit is according to formula bm-12m-1+bm-22m-2…+b020Respectively unfolding m-8 to obtain 8 bit planes; separating bit planes of the three images, and dividing each image into 8 bit planes, namely 8 binary image matrixes; wherein b represents a binary pixel of each bit plane, 2 represents each bin, and m represents 8 bit planes;
(2) analyzing the importance degree of each bit plane occupying the visual effect of the source image through an image bit plane diagram, wherein the bit planes are arranged from small to large in a bit number of 1 to 8 bits, and the bit plane of the eighth bit is 2 of a pixel in a binary system7～28When the eighth bit takes 1, the largest pixel value is 11111111111 (255), and the smallest pixel value is 10000000 (128); when the eighth bit is 0, the maximum pixel value is 01111111(127), the minimum pixel value is 00000000(0), the decimal range of a source image corresponding to 0/1 of the eighth bit pixel is 0-127/128-255, and the decimal range of a source image corresponding to 0/1 of the seventh bit pixel is 0-191/64-255; the confusion pixel area of the eighth bit is 0, and the confusion pixel area of the seventh bit is 64-191; the sixth bit obtains a mixed pixel area of 32-223;
firstly, reading an image into a matrix format, and then circularly reading different bits of the image and storing the different bits of the image into a newly-built binary _ matrix; thus, the binary _ matrix stores the bit-plane 0-1 matrices of the eight channels; separating out the bit planes of the images of the three channels to obtain 24 bit plane images, then carrying out two replacement sequence modes, wherein the first mode adopts a default scrambling sequence, adopts three smallest confusion pixel regions for scrambling, namely 678 bit planes, and combines the confusion pixel regions into six recombination modes of 12345678, 12345687, 12345768, 12345786, 12345867 and 12345876, and the second mode adopts a mode of carrying out image recombination analysis by observing the visual structure of the image, analyzing the pixel proportion through codes, extracting the largest pixel range proportion and ensuring that the recombination visual effect can keep the original image structure;
and recombining after image bit plane separation to obtain a new RGB image, and putting the new RGB image into a data set as new training data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910700184.5A CN110503612B (en) | 2019-07-31 | 2019-07-31 | Bit plane based data separation and recombination enhancement method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910700184.5A CN110503612B (en) | 2019-07-31 | 2019-07-31 | Bit plane based data separation and recombination enhancement method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110503612A CN110503612A (en) | 2019-11-26 |
| CN110503612B true CN110503612B (en) | 2022-01-14 |
Family
ID=68586812
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910700184.5A Active CN110503612B (en) | 2019-07-31 | 2019-07-31 | Bit plane based data separation and recombination enhancement method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110503612B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021016923A1 (en) * | 2019-07-31 | 2021-02-04 | 东北大学 | Data enhancement method employing bit plane separation and recombination |
| CN112257580A (en) * | 2020-10-21 | 2021-01-22 | 中国石油大学(华东) | Human body key point positioning detection method based on deep learning |
| CN112614052B (en) * | 2020-12-24 | 2023-07-28 | 东北大学 | Bit-level data enhancement method for super resolution of image |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8547382B2 (en) * | 2008-05-30 | 2013-10-01 | Advanced Micro Devices, Inc. | Video graphics system and method of pixel data compression |
| CN106530207A (en) * | 2016-12-06 | 2017-03-22 | 南京理工大学 | Logistic chaotic mapping-based digital image encryption method |
| US10185880B2 (en) * | 2017-03-31 | 2019-01-22 | Here Global B.V. | Method and apparatus for augmenting a training data set |
| CN108492343B (en) * | 2018-03-28 | 2021-09-21 | 东北大学 | Image synthesis method for training data for expanding target recognition |
| CN109829881A (en) * | 2018-12-17 | 2019-05-31 | 广东电网有限责任公司 | Bird's Nest detection method and system based on deep learning |
-
2019
- 2019-07-31 CN CN201910700184.5A patent/CN110503612B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN110503612A (en) | 2019-11-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110503612B (en) | Bit plane based data separation and recombination enhancement method | |
| JP2968582B2 (en) | Method and apparatus for processing digital data | |
| CN101919254B (en) | Prediction-based image processing | |
| US6115496A (en) | Method and apparatus for accelerating image data compression | |
| US5768481A (en) | Method and apparatus for compression of digitized image data using a dynamic band recompression scheme | |
| JP4531749B2 (en) | Color image compression using spectral decorrelation and removal of spatial redundancy | |
| CN1012302B (en) | Color image dispaly system | |
| DE102015114978A1 (en) | data compression | |
| JPH03119486A (en) | Method of compressing information included in entered form for storage or transfer | |
| US6956965B1 (en) | Compressing and decompressing an image | |
| CN110971904B (en) | Control method for image compression | |
| CN112949754B (en) | Text recognition data synthesis method based on image fusion | |
| CN112118449B (en) | Method and device for compressing and decompressing image | |
| JPH07322074A (en) | Equipment and method for data processing to process 2 level image file that a dither ring happened | |
| CN104809747B (en) | The statistical method and its system of image histogram | |
| Nagarajan et al. | An enhanced approach in run length encoding scheme (EARLE) | |
| Rajesh et al. | FastSS: Fast and smooth segmentation of JPEG compressed printed text documents using DC and AC signal analysis | |
| Larabi et al. | A fast color quantization using a matrix of local pallets | |
| CN108200433B (en) | Image compression and decompression method | |
| WO1996039682A1 (en) | Block classification for accelerating image data compression | |
| WO2021016923A1 (en) | Data enhancement method employing bit plane separation and recombination | |
| WO2016184485A1 (en) | Image compression | |
| Meyyappan et al. | Lossless Digital Image Compression Method For Bitmap Images | |
| JP2000049618A (en) | Hvq compression method combined with 90° rotation | |
| WO2006115384A1 (en) | Apparatus and method of conversing data |
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 |