CN202736078U - Feature extraction module used for digital image processing - Google Patents
Feature extraction module used for digital image processing Download PDFInfo
- Publication number
- CN202736078U CN202736078U CN 201220369660 CN201220369660U CN202736078U CN 202736078 U CN202736078 U CN 202736078U CN 201220369660 CN201220369660 CN 201220369660 CN 201220369660 U CN201220369660 U CN 201220369660U CN 202736078 U CN202736078 U CN 202736078U
- Authority
- CN
- China
- Prior art keywords
- register
- buffer zone
- feature extraction
- array
- data
- 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
Images
Landscapes
- Image Processing (AREA)
Abstract
The utility model discloses a feature extraction module used for digital image processing. The feature extraction module comprises a feature extraction register array, an upper buffer area, a lower buffer area and a right butter area, wherein the upper butter area and the lower buffer area are formed by the register array of X rows and N+Y columns. A traversal method comprises one or more traversal flows comprising a plurality of shifting down operations, one shifting right operation, a plurality of shifting up operations and one shifting right operation. The feature extraction module, after each basic operation, quickly determines the basic operation needed in the next time according to circumstances of a state machine. According to the feature extraction module, through the technical scheme, hardware resources can be effectively reduced, and furthermore, the higher the resolution of an image is, the more effectively the required resources can be reduced.
Description
Technical field
The utility model relates to the image pattern recognition field, and particularly a kind of characteristic extracting module for Digital Image Processing is used for feature and searches and travel through.
Background technology
Along with the increase to the man-machine interaction demand, people have proposed more and more higher requirement to man-machine interactive system.One of them very important index is exactly the resolution of system.But its inherent characteristics of man-machine interaction is that data processing amount is large, and requirement of real-time is high.Here it is hinders the large bottleneck that it moves towards the high resolution system application.
Field of human-computer interaction has been used a large amount of mode identification methods.But the difference of both maximums is---the given width of cloth figure of pattern-recognition (be 20*20 such as size), go to identify whether comprise required target of looking for (such as people's face) in this image again; And the given width of cloth figure of man-machine interaction (be 640*480 such as size), whether go to identify where going up among this width of cloth figure has required target of looking for (be the people face of 20*20 such as size) again.Therefore, compare with pattern-recognition, man-machine interaction has had more a process of searching in the entire image traversal.
Specific objective (such as people's face) searched to use feature.Feature is generally done by a plurality of pixels and is formed, and its pixel number that comprises of different features is different.Therefore calculate number of calculations that each feature needs and also be operation time different, this is unfavorable for the realization of hardware.So generally all adopt now the method based on the integrogram calculated characteristics.The value of each point is its upper left gray-scale value sum of having a few in the integrogram.Therefore when calculating each feature, only need to carry out 2 sub-addition computings and 1 subtraction gets final product with the integrated value of its 4 end points.So not only reduce computational complexity but also guaranteed each operation time.
Man-machine interaction is large because of its data processing amount, and the high characteristics of requirement of real-time become the large bottleneck that it is applied to common treatment.Recent years, also begun some man-machine interactions all over the world and processed the trial of accomplishing on FPGA or the ASIC.Consider the factors such as computational complexity, arithmetic speed, power consumption, the way of main flow all adopts based on the integration Graph Character and searches and travel through now.Wherein, 3 kinds of typical methods are arranged.First method stores the integrogram of entire image in the register (register) into.Second method stores integrogram among the RAM into by row or by row.The third method, the capable buffer(such as the image that add specific quantity between RAM and register are 640*480, the size that detects target is 20*20, then adding (20+ stepping) individual length is 640 capable buffer), data are first from RAM to buffer, arrive at last again register, no matter adopt above any design, in the middle of practical application, all can have restriction.First method, its required register number of using is a lot, can only be used for the smaller situation of entire image.Second method, the speed of extracting feature will be dragged the speed of slow whole system too slowly.Process although method three has been carried out compromise to the two kinds of methods in front, additive decrementation a lot of buffer resources.Present stage searches and travels through maximum difficult point based on the integration Graph Character and is: how to carry out fast as far as possible feature extraction with the least possible hardware resource.
The utility model content
Shortcoming and deficiency that the utility model exists in order to overcome prior art provide a kind of characteristic extracting module for Digital Image Processing.
The utility model adopts following technical scheme:
A kind of characteristic extracting module for Digital Image Processing, comprise feature extraction register array, upper buffer zone, lower buffer zone, right buffer zone, described upper buffer zone is positioned at the top of feature extraction register array, lower buffer zone is positioned at the below of feature extraction register array, and right buffer zone is positioned at the right side of feature extraction register array;
Described feature extraction register array is made of the register array of the capable N row of M, and described M is the natural number greater than 2, and N is the natural number greater than 1;
In the described feature extraction register array, the Y column register on, right side capable except the X X capable, the below of top, other register is connected for the register of X with its distance up, and the below is connected for the register of X with its distance, and the right side is connected for the register of Y with its distance;
Described upper and lower buffer zone consists of by the register array of the capable N+Y row of X;
Each register in the capable N of X in the described upper buffer zone row and this register of feature extraction register array middle distance are that the register of X is connected, and the register during the capable Y of other X is listed as is connected with the register that right this register of buffer zone middle distance is X;
Each register in the capable N of X in the described lower buffer zone row and this register of feature extraction register array middle distance are that the register of X links to each other, and the register during the capable Y of other X is listed as is connected with the register that right this register of buffer zone middle distance is X;
Described right buffer zone is made of the register array of the capable Y of M row, and each register in the right buffer zone and this register of feature extraction register array middle distance are that the register of Y is connected;
Described X is the row stepping, and X is the natural number less than M/2, and described Y is the row stepping, and Y is the natural number less than N.
Described feature extraction register array is for being used for many input registers array of storage integrogram data.
Described upper and lower buffer zone is for being used for single shift register array of inputting many outputs of storage integrogram data.
Described right buffer zone is for being used for single shift register array of inputting many outputs of storage integrogram data.
A kind of traversal method of the characteristic extracting module for Digital Image Processing, described traversal method comprise and repeat once abovely by repeatedly moving down operation that a right-shift operation is moved operation on repeatedly, the traversal flow process that right-shift operation consists of.
Described moving down is operating as:
Except the capable register of top X, it is in the register of X apart from this register that other register is write into the top with the data of storing in the described feature extraction register array;
Register in the described lower buffer zone is write into the data of storing in the register of the register of connected feature extraction register array and right buffer zone;
Except the capable register of top X, it is in the register of X apart from this register that other register is write into the top with the data of storing in the described right buffer zone;
Described right-shift operation is:
Except left Y column register, it is in the register of Y apart from this register that other register is write into left with the data of storing in the described feature extraction register array;
Register in the described right buffer zone is write into register in the connected feature extraction register array with the data of storing;
Move on described and be operating as:
Described feature extraction register array is except the capable register of below X, and it is in the register of X apart from this register that other register is write into the below with the data of storing;
The register of described upper buffer zone is write into the data of storing in the connected feature extraction register array;
Except the capable register of below X, it is in the register of X apart from this register that the data of other registers storage are write into the below in the described right buffer zone;
Register in the described upper buffer zone is write into connected right buffer zone with the data of storing.
Described traversal method specifically comprises the steps:
When (1) traversal began, the data in the feature extraction register array were invalid, carry out feature extraction and must extract register array to the most upper left data input feature vector of image first; Because feature extraction register array and periphery do not have interface, so first by lower buffer zone input data;
(2) then the lower buffer zone of data inputs once moves down operation, and the X of the below of data shift-in feature extraction register array and right buffer zone is capable;
(3) repeating step (2) is until the whole shift-in feature extraction of the most upper left data of image register array;
(4) carry out an image characteristics extraction, because operation next time so the image view data that down X is capable again deposits lower buffer zone in simultaneously, once moves down operation also for moving down operation after this feature extraction is finished;
(5) repeating step (4) is not until image has data toward the below again; At this moment, carry out right-shift operation one time, because when repeatedly moving down operation, right buffer zone is also being constantly updated data, so this time, right buffer zone was deposited active data, directly carried out right-shift operation, need not to wait pending data to write right buffer zone;
(6) carry out right-shift operation after, being operating as next time moves, upper buffer zone is started working, in feature extraction, image deposits buffer zone in toward the capable view data of top X again;
(7) carry out again image characteristics extraction after moving operation on carrying out once; Simultaneously, because operation next time moves operation on also being, image deposits buffer zone in toward the capable data of top X again;
(8) repeating step (7) is not until image has data toward the top again; At this moment, carry out right-shift operation one time, because when repeatedly moving operation, right buffer zone is also being constantly updated data, so this time, right buffer zone was deposited active data, directly carried out right-shift operation, need not to wait pending data to write right buffer zone;
(9) step (4) before repeating is to step (8), until that whole target image is traversed is complete.
Compared to existing technology, the utlity model has following advantage:
When (1) being applied to higher resolution image, needed hardware resource seldom.Take image as 640*480, the size that detects target is 20*20, and stepping is 2 as an illustration.Adopt the used scheme of the utility model to need 528 (20*20+2*20+2*22*2) registers.If adopt the scheme of whole integrogram storage to need 307200 (640*480) registers.If adopt adding the scheme of buffer, to need 400 (20*20) registers and 22 degree of depth be 640 block RAM.This shows, adopt scheme described in the utility model can effectively reduce hardware resource, and for the higher image of resolution, more can effectively reduce resource requirement.
(2) when effectively reducing hardware resource, on not impact of processing speed.The utility model can be made rapidly next time required basic operation according to the situation of state machine after each time basic operation.When the feature extraction register array carries out the data displacement, data are write in the corresponding buffer zone.Because the two carries out simultaneously, and in most of situation, stepping is smaller, the time spent of feature extraction is longer than data and writes, and can't bring extra time-delay like this.And the situation that stepping is large and the very fast end of feature extraction is so relatively seldom, and in this case, although can bring certain time-delay, this also is complete acceptable compared with the saving of resource.In addition, because when repeatedly moving operation, moving down operation, right buffer zone is also being constantly updated data, therefore when needs carry out right-shift operation, right buffer zone has been deposited active data, can directly carry out right-shift operation, need not to wait pending data to write right buffer zone, reach the purpose of speed-raising.
Description of drawings
Fig. 1 is the structured flowchart of characteristic extracting module of the present utility model, and reg represents register among the figure.
Fig. 2 be characteristic extracting module on move operation, scheming medium and small square frame is register reg.
Fig. 3 is the operation that moves down of characteristic extracting module, and scheming medium and small square frame is register reg.
Fig. 4 is the right-shift operation of characteristic extracting module, and scheming medium and small square frame is register reg.
Fig. 5 is the characteristic extracting module among the embodiment, and scheming medium and small square frame is register reg.
Fig. 6 is the procedure chart of whole image traversal method among the embodiment.
Embodiment
Below in conjunction with embodiment and accompanying drawing, the utility model is described in further detail, but embodiment of the present utility model is not limited to this.
Embodiment
Be illustrated in figure 1 as the structured flowchart of characteristic extracting module of the present utility model, comprise feature extraction register array, upper buffer zone, lower buffer zone, right buffer zone, described upper buffer zone is positioned at the top of feature extraction register array, lower buffer zone is positioned at the below of feature extraction register array, and right buffer zone is positioned at the right side of feature extraction register array;
Described feature extraction register array is made of the register array of the capable N row of M, and described M is the natural number greater than 2, and N is the natural number greater than 1;
In the described feature extraction register array, the Y column register on, right side capable except the X X capable, the below of top, other register is connected for the register of X with its distance up, and the below is connected for the register of X with its distance, and the right side is connected for the register of Y with its distance;
Described upper and lower buffer zone consists of by the register array of the capable N+Y row of X;
Each register in the capable N of X in the described upper buffer zone row and this register of feature extraction register array middle distance are that the register of X is connected, and the register during the capable Y of other X is listed as is connected with the register that right this register of buffer zone middle distance is X;
Each register in the capable N of X in the described lower buffer zone row and this register of feature extraction register array middle distance are that the register of X links to each other, and the register during the capable Y of other X is listed as is connected with the register that right this register of buffer zone middle distance is X;
Described right buffer zone is made of the register array of the capable Y of M row, and each register in the right buffer zone and this register of feature extraction register array middle distance are that the register of Y is connected;
Described X is the row stepping, and X is the natural number less than M/2, and described Y is the row stepping, and Y is the natural number less than N.
Described feature extraction register array is for being used for many input registers array of storage integrogram data.
Described upper and lower buffer zone is for being used for single shift register array of inputting many outputs of storage integrogram data.
Described right buffer zone is for being used for single shift register array of inputting many outputs of storage integrogram data.
A kind of traversal method of the characteristic extracting module for Digital Image Processing, described traversal method comprise and repeat once abovely by repeatedly moving down operation that a right-shift operation is moved operation on repeatedly, the traversal flow process that right-shift operation consists of.
Be illustrated in figure 2 as characteristic extracting module on move operation,
Move on described and be operating as:
Described feature extraction register array is except the capable register of below X, and it is in the register of X apart from this register that other register is write into the below with the data of storing;
The register of described upper buffer zone is write into the data of storing in the connected feature extraction register array;
Except the capable register of below X, it is in the register of X apart from this register that the data of other registers storage are write into the below in the described right buffer zone;
Register in the described upper buffer zone is write into connected right buffer zone with the data of storing.
Be illustrated in figure 3 as the operation that moves down of characteristic extracting module,
Except the capable register of top X, it is in the register of X apart from this register that other register is write into the top with the data of storing in the described feature extraction register array;
Register in the described lower buffer zone is write into the data of storing in the register of the register of connected feature extraction register array and right buffer zone;
Except the capable register of top X, it is in the register of X apart from this register that other register is write into the top with the data of storing in the described right buffer zone.
Be illustrated in figure 4 as the right-shift operation of characteristic extracting module:
Except left Y column register, it is in the register of Y apart from this register that other register is write into left with the data of storing in the described feature extraction register array;
Register in the described right buffer zone is write into register in the connected feature extraction register array with the data of storing.
Be illustrated in figure 5 as the characteristic extracting module among the embodiment: its feature extraction register array size is 10 * 10; In the above and below of feature extraction register array, be respectively a row * classify 2 * 12 buffer zone as, be called buffer zone and lower buffer zone; Right-hand at the feature extraction register array is a row * classify 10 * 2 buffer zone as, is referred to as right buffer zone.
Be illustrated in figure 6 as the procedure chart of the whole image traversal method of embodiment,
Described traversal method specifically comprises the steps:
When (1) traversal began, the data in the feature extraction register array were invalid, carry out feature extraction and must extract register array to the most upper left data input feature vector of image first; Because feature extraction register array and periphery do not have interface, so first by lower buffer zone input data;
(2) then the lower buffer zone of data inputs once moves down operation, two row of the below of data shift-in feature extraction register array and right buffer zone;
(3) repeating step (2), 5 times altogether, until the whole shift-in feature extraction of the most upper left data of image register array;
(4) carry out an image characteristics extraction because next time operation is also for moving down operation, so simultaneously with image more down the view data of two row deposit lower buffer zone in, after this feature extraction is finished, once move down operation;
(5) repeating step (4) is not until image has data toward the below again; At this moment, carry out right-shift operation one time, because when repeatedly moving down operation, right buffer zone is also being constantly updated data, so this time, right buffer zone was deposited active data, directly carried out right-shift operation, need not to wait pending data to write right buffer zone;
(6) carry out right-shift operation after, being operating as next time moves, upper buffer zone is started working, in feature extraction, image deposits buffer zone in toward the view data of top two row again;
(7) carry out again image characteristics extraction after moving operation on carrying out once; Simultaneously, because operation next time moves operation on also being, image deposits buffer zone in toward the data of top two row again;
(8) repeating step (7) is not until image has data toward the top again; At this moment, carry out right-shift operation one time, because when repeatedly moving operation, right buffer zone is also being constantly updated data, so this time, right buffer zone was deposited active data, directly carried out right-shift operation, need not to wait pending data to write right buffer zone;
(9) step (4) before repeating is to step (8), as shown in Figure 6, until that whole target image is traversed is complete.
When the utility model is applied to higher resolution image, needed hardware resource seldom, take image as 640*480, the size that detects target is 20*20, stepping is 2 as an illustration.Adopt the used scheme of the utility model to need 528 (20*20+2*20+2*22*2) registers.If adopt the scheme of whole integrogram storage to need 307200 (640*480) registers.If adopt adding the scheme of buffer, to need 400 (20*20) registers and 22 degree of depth be 640 block RAM.This shows, adopt scheme described in the utility model can effectively reduce hardware resource, and for the higher image of resolution, more can effectively reduce resource requirement.
Above-described embodiment is the better embodiment of the utility model; but embodiment of the present utility model is not limited by the examples; other any do not deviate from change, the modification done under Spirit Essence of the present utility model and the principle, substitutes, combination, simplify; all should be the substitute mode of equivalence, be included within the protection domain of the present utility model.
Claims (4)
1. characteristic extracting module that is used for Digital Image Processing, it is characterized in that, comprise feature extraction register array, upper buffer zone, lower buffer zone, right buffer zone, described upper buffer zone is positioned at the top of feature extraction register array, lower buffer zone is positioned at the below of feature extraction register array, and right buffer zone is positioned at the right side of feature extraction register array;
Described feature extraction register array is made of the register array of the capable N row of M, and described M is the natural number greater than 2, and N is the natural number greater than 1;
In the described feature extraction register array, the Y column register on, right side capable except the X X capable, the below of top, other register is connected for the register of X with its distance up, and the below is connected for the register of X with its distance, and the right side is connected for the register of Y with its distance;
Described upper and lower buffer zone consists of by the register array of the capable N+Y row of X;
Each register in the capable N of X in the described upper buffer zone row and this register of feature extraction register array middle distance are that the register of X is connected, and the register during the capable Y of other X is listed as is connected with the register that right this register of buffer zone middle distance is X;
Each register in the capable N of X in the described lower buffer zone row and this register of feature extraction register array middle distance are that the register of X links to each other, and the register during the capable Y of other X is listed as is connected with the register that right this register of buffer zone middle distance is X;
Described right buffer zone is made of the register array of the capable Y of M row, and each register in the right buffer zone and this register of feature extraction register array middle distance are that the register of Y is connected;
Described X is the row stepping, and X is the natural number less than M/2, and described Y is the row stepping, and Y is the natural number less than N.
2. described a kind of characteristic extracting module for Digital Image Processing according to claim 1 is characterized in that, described feature extraction register array is for being used for many input registers array of storage integrogram data.
3. described a kind of characteristic extracting module for Digital Image Processing according to claim 1 is characterized in that, described upper and lower buffer zone is the shift register arrays of the many outputs of single input that are used for storage integrogram data.
4. described a kind of characteristic extracting module for Digital Image Processing according to claim 1 is characterized in that, described right buffer zone is the shift register arrays of the many outputs of single input that are used for storage integrogram data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201220369660 CN202736078U (en) | 2012-07-27 | 2012-07-27 | Feature extraction module used for digital image processing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201220369660 CN202736078U (en) | 2012-07-27 | 2012-07-27 | Feature extraction module used for digital image processing |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN202736078U true CN202736078U (en) | 2013-02-13 |
Family
ID=47661756
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN 201220369660 Expired - Fee Related CN202736078U (en) | 2012-07-27 | 2012-07-27 | Feature extraction module used for digital image processing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN202736078U (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102867181A (en) * | 2012-07-27 | 2013-01-09 | 华南理工大学 | Characteristic extraction module for digital image processing and traversing method |
| CN107438860A (en) * | 2015-04-23 | 2017-12-05 | 谷歌公司 | Framework for the efficient programmable graphics processing of high performance power |
-
2012
- 2012-07-27 CN CN 201220369660 patent/CN202736078U/en not_active Expired - Fee Related
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102867181A (en) * | 2012-07-27 | 2013-01-09 | 华南理工大学 | Characteristic extraction module for digital image processing and traversing method |
| CN107438860A (en) * | 2015-04-23 | 2017-12-05 | 谷歌公司 | Framework for the efficient programmable graphics processing of high performance power |
| US10719905B2 (en) | 2015-04-23 | 2020-07-21 | Google Llc | Architecture for high performance, power efficient, programmable image processing |
| CN107438860B (en) * | 2015-04-23 | 2021-03-23 | 谷歌有限责任公司 | Architecture for high performance power efficient programmable image processing |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106157319B (en) | The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks | |
| CN113874883A (en) | Hand pose estimation | |
| CN107808131A (en) | Dynamic gesture identification method based on binary channel depth convolutional neural networks | |
| CN106203617A (en) | A kind of acceleration processing unit based on convolutional neural networks and array structure | |
| CN109949214A (en) | A kind of image Style Transfer method and system | |
| CN108364281A (en) | A kind of ribbon edge hair defect defect inspection method based on convolutional neural networks | |
| CN107679522A (en) | Action identification method based on multithread LSTM | |
| CN106022392A (en) | Deep neural network sample automatic accepting and rejecting training method | |
| CN111046917B (en) | Object-based enhanced target detection method based on deep neural network | |
| CN103020890A (en) | Visual processing device based on multi-layer parallel processing | |
| CN109766949A (en) | Convolutional neural networks light weight method, device and electronic equipment | |
| CN105976382A (en) | TFT-LCD Mura defect detection algorithm based on defect area prognosis and level set | |
| CN104143096A (en) | Binary connected domain labeling method based on address event representation | |
| CN109165733A (en) | Multi-input multi-output matrix maximum pooling vectorization implementation method | |
| CN104680531A (en) | Connection flux statistical information extraction method and VLSI structure | |
| CN202736078U (en) | Feature extraction module used for digital image processing | |
| CN102867181A (en) | Characteristic extraction module for digital image processing and traversing method | |
| CN113920516A (en) | Calligraphy character skeleton matching method and system based on twin neural network | |
| Jiang et al. | Joint attention-guided feature fusion network for saliency detection of surface defects | |
| CN114612709A (en) | Multi-scale target detection method guided by image pyramid characteristics | |
| CN102074003A (en) | Mean shift-based embedded type image tracking system | |
| CN102117409B (en) | Hardware feature box for searching and traversing feathers based on integrogram | |
| CN102117326A (en) | Traversal method used for searching image features | |
| Yang et al. | Scinet: Semantic cue infusion network for lane detection | |
| Kumawat et al. | ShuffleBlock: Shuffle to regularize deep convolutional neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20130213 Termination date: 20170727 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |