CN102722723B - Multi-scale-based Adaboost detection method and system - Google Patents
Multi-scale-based Adaboost detection method and system Download PDFInfo
- Publication number
- CN102722723B CN102722723B CN201210169536.7A CN201210169536A CN102722723B CN 102722723 B CN102722723 B CN 102722723B CN 201210169536 A CN201210169536 A CN 201210169536A CN 102722723 B CN102722723 B CN 102722723B
- Authority
- CN
- China
- Prior art keywords
- yardstick
- checked
- length
- measuring point
- array
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a multi-scale-based Adaboost detection method. The method comprises the following steps of: acquiring the information of a point to be detected from a queue of points to be detected; judging whether step length which is acquired from a step length array and corresponds to the current detection scale is 0, and judging whether the point to be detected is at an initial position; if the step length is 0 or the step length is not 0 and the point to be detected is at the initial position, setting an image local area which is loaded from an internal memory to a cache and corresponds to the point to be detected or an image local area which is stored in the cache and corresponds to the point to be detected as an image which is detected by a weak classifier; and if the step length is not 0 and the point to be detected is not at the initial position, updating the step length which is stored in the step length array and corresponds to the current detection scale. The invention also provides a multi-scale-based Adaboost detection system. By adoption of the method and the system, Adaboost detection efficiency can be improved, and the processing performance of the system is improved.
Description
Technical field
The present invention relates to computer vision and image and process, particularly a kind of Adaboost detection method and system based on multiple dimensioned.
Background technology
Object detection is the binary classification problems in pattern-recognition, can be widely used in a plurality of fields, as the pedestrian detection in intelligent monitoring, the detection of people's face etc., retrieves similar image etc. in video.Object detecting system, a pattern classification system, generally comprises sensor reception image or video data, feature generation, feature selecting, classifier design and detection system and assesses this several parts.
At present, conventionally utilize Adaboost algorithm to carry out detection system assessment.When utilizing Adaboost algorithm to detect testing image, conventionally same testing image is carried out to N detection that detects yardstick, same detection yardstick for same testing image, X the check point that conventionally need to comprise this testing image carries out traverse scanning, if carry out N detection that detects yardstick, need X the check point that this testing image is comprised to carry out traverse scanning N time, and when each pointwise traverse scanning, computing machine can be written into the temporary storage between CPU and external memory by the image local area relevant to check point, be temporary in Cache buffer memory, in the testing process of some detection yardsticks of the X that testing image an is comprised check point, the image local area relevant to check point that is written in advance Cache buffer memory can constantly be replaced, and in the testing process of follow-up difference detection yardstick need the above-mentioned image local area relevant to check point to be again written into Cache buffer memory, there is the cache miss (Cache Miss) of surveyed area.
The existing Adaboost of utilization algorithm carries out in the detection method of N detection yardstick to testing image, if in system will there is Cache Miss one time in average every M check point, the numerical value that the number of times that the existing Adaboost of utilization algorithm carries out occurring in N detection method that detects yardstick Cache Miss to testing image rounds up for the product of N and X and the business of M.The M of now only take is greater than 1 for example as 4, X is greater than 4, N, and the cache miss of the surveyed area occurring in existing detection method is described, specific as follows:
Can inclusion test point P2, check point P3 and image local area corresponding to check point P4 in image local area corresponding to check point P1, because Cache Miss can occur one time every 4 check points, suppose that check point P1 and regional area corresponding to check point Pa are not stored in Cache buffer memory; Wherein, a is 4j+1; J is more than or equal to 1 integer.
At the 1st, detect under yardstick, the image local area that check point P1 is corresponding is not stored in Cache buffer memory, there is the 1st Cache Miss under this detection yardstick, system loads the 1st image local area that detects check point P 1 correspondence under yardstick to Cache buffer memory from internal memory, utilize the image local area that check point P1 is corresponding, to check point P1, check point P2, check point P3 and check point P4 detect, when check point Pa is detected, the regional area that check point Pa is corresponding is not stored in Cache buffer memory, system loads the 1st and detects image local area that under yardstick, check point Pa is corresponding to Cache buffer memory from internal memory, image local area corresponding to check point P1 originally leaving in Cache buffer memory can be replaced, there is the 2nd Cache Miss under this detection yardstick, by that analogy, at the 1st, detect under yardstick, the number of times that X check point is carried out to the Cache Miss that occurs in testing process is the numerical value after the business of X and M rounds up.
At the 2nd, detect under yardstick, image local area corresponding to check point P1 is due to by the replacement of image local area corresponding to the follow-up check point being written into, be not stored in Cache buffer memory, there is the 1st Cache Miss under this detection yardstick, system loads the 2nd and detects image local area that under yardstick, check point P1 is corresponding to Cache buffer memory from internal memory, utilize the image local area that check point P1 is corresponding, to check point P1, check point P2, check point P3 and check point P4 detect, when check point Pa is detected, the regional area that check point Pa is corresponding is not stored in Cache buffer memory, system loads the 2nd and detects image local area that under yardstick, check point Pa is corresponding to Cache buffer memory from internal memory, image local area corresponding to check point P1 originally leaving in Cache buffer memory is replaced, there is the 2nd Cache Miss under this detection yardstick, by that analogy, at the 2nd, detect under yardstick, the number of times that X check point is carried out to the Cache Miss that occurs in testing process is the numerical value after the business of X and M rounds up.
By that analogy, at 3rd～N, detect under yardstick X check point carried out in testing process, existing detection method can cause each to detect under yardstick, the number of times that X check point is carried out to the Cache Miss that occurs in testing process is the numerical value after the business of X and M rounds up, like this, the numerical value after the business that N number of times that detects the Cache Miss that under yardstick, X check point is carried out occurring in testing process is X and M rounds up and the product of N.
Wherein, detect the detection window size of selecting when yardstick represents to carry out pointwise traverse scanning; Check point is the pixel position on image to be detected.
In sum, in the existing Adaboost detection method based on multiple dimensioned, after there is Cache Miss, the CPU that system comprises can forward to and in external memory, search the image local area relevant to check point, reduced the handling property of detection system, affected the detection efficiency of Adaboost, and it is larger by the size of altimetric image, the frequency that in multiple scale detecting process, the cache miss of surveyed area occurs is higher, the handling property of whole system is lower, and Adaboost detection efficiency need further raising.
Summary of the invention
In view of this, the object of the present invention is to provide a kind of Adaboost detection method based on multiple dimensioned, the method can improve Adaboost detection efficiency, improves the handling property of system.
The object of the present invention is to provide a kind of Adaboost detection system based on multiple dimensioned, this system can improve Adaboost detection efficiency, improves the handling property of system.
For achieving the above object, technical scheme of the present invention is specifically achieved in that
An Adaboost detection method based on multiple dimensioned, the method comprises:
A, according to position skew, from the measuring point queue to be checked of the information that comprises X*N measuring point to be checked, obtain the information of a measuring point to be checked; The information of described measuring point to be checked at least comprises position skew, detects yardstick rank and identification information; Described N and X are the integer that is greater than 1;
Whether step-length corresponding to current detection yardstick that B, judgement are obtained from the step-length array that comprises N array element is 0, if so, and execution step D, otherwise execution step C; Described step-length is the number of the measuring point to be checked at interval between adjacent two measuring points to be checked that need detect through Weak Classifier;
C, according to identification information judgment measuring point to be checked, whether be positioned at first place and put, if so, execution step D, otherwise, upgrade the step-length corresponding to current detection yardstick of preserving in step-length array, execution step A;
D, judge whether Cache buffer memory has preserved the image local area corresponding with measuring point to be checked, if do not preserved, from internal memory, load image local area that measuring point to be checked is corresponding to Cache buffer memory, perform step afterwards E, otherwise directly perform step E;
E, utilize the Weak Classifier under current detection yardstick to detect the image local area corresponding with measuring point to be checked of preserving in Cache buffer memory, testing result is write to described step-length array as step-length corresponding to current detection yardstick, execution step A.
Preferably, before described steps A, further comprise:
A ', the step-length array that generates N detection yardstick and X measuring point to be checked are N the measuring point queue to be checked detecting under yardstick.
Preferably, before described steps A, further comprise:
According to sample image and Adaboost algorithm, generate with N and detect the Weak Classifier that in yardstick, each detection yardstick is corresponding.
In said method, steps A ' described generation N the step-length array that detects yardstick comprises:
The step-length array that generation comprises N array element, an array element detects step-length corresponding to yardstick in order to preserve one; The initial value of the step-length that described arbitrary array element is preserved is 0.
In said method, steps A ' the to be checked measuring point queue of a described generation X measuring point to be checked under N detection yardstick comprise:
A1 ', an array that comprises X*N array element of generation;
A2 ', under each detects yardstick, testing image is scanned, obtain each check point under current detection yardstick position skew, detect yardstick rank and identification information;
In an array element of the array that A3 ', each check point that scanning is obtained generate at N position skew, detection yardstick rank and the identification information write step A1 ' detecting under yardstick;
A4 ', the array that has write the information of X*N measuring point to be checked in steps A 3 ' is arranged according to position skew ascending order, the array after arranging according to position skew ascending order is as measuring point queue to be checked;
The skew of described position is in order to represent coordinate or the positional information of check point in image to be detected; Described detection yardstick rank is in order to determine the relative size that detects yardstick; Described identification information is put in order to represent the first place whether check point is positioned at testing image.
In said method, described in step C the first be set to detect line by line in the row first place of a line put or detect by column in the row first place of row put.
In said method, upgrade the step-length corresponding to current detection yardstick of preserving in step-length array described in step C and comprise:
Step-length using the difference of step-length corresponding to current detection yardstick and 1 after upgrading, writes the step-length after upgrading in array element corresponding with current detection yardstick in step-length array.
An Adaboost detection system based on multiple dimensioned, this system connects internal memory and Cache buffer memory, and this system comprises:
Detect control module, connect internal memory, Cache buffer memory and N Weak Classifier, according to position skew, from the measuring point queue to be checked of the information that comprises the individual measuring point to be checked of X*N, obtain the information of a measuring point to be checked; Described N and described X are the integer that is greater than 1; The information of described measuring point to be checked at least comprises position skew, detects yardstick rank and identification information;
Step-length corresponding to current detection yardstick that described detection control module is obtained the step-length array of judging from comprising N array element is 0, or judge that the non-zero and measuring point to be checked of step-length that the current detection yardstick obtain is corresponding is when put first place from the step-length array that comprises N array element, by image local area corresponding to measuring point to be checked being loaded on from internal memory Cache buffer memory, or image local area corresponding to measuring point to be checked in Cache buffer memory exports the Weak Classifier under current detection yardstick to, when judging that the non-zero and measuring point to be checked of step-length that current detection yardstick is corresponding is not positioned at first place and puts, upgrade step-length corresponding to array element corresponding with current detection yardstick in temporary step-length array, described step-length is the number of the measuring point to be checked at interval between adjacent two measuring points to be checked that need detect through Weak Classifier,
In a described N Weak Classifier, arbitrary Weak Classifier, according to the image local area corresponding with measuring point to be checked of input, generates the testing result of presentation video similarity degree, exports testing result to detection control module as the step-length corresponding with current detection yardstick.
Preferably, described detection control module also generates N step-length array and the detection queue of the individual measuring point to be checked of X under N detection yardstick that detects yardstick.
In said system, described detection control module comprises:
Scanning element, according to N and the X of the output of check point monitoring unit, generate an array that comprises X*N array element, at each, detect under yardstick testing image is scanned, obtain position skew, detection yardstick rank and the identification information of each check point under current detection yardstick and write in the array of generation, the array after arranging according to position skew ascending order is as measuring point queue write storage unit to be checked;
Step-length control module, generates in order to preserve N step-length array that detects yardstick according to the N of check point monitoring unit output, and by step-length array write storage unit; The first triggering according to check point monitoring unit, utilizes current detection yardstick rank from storage unit, to obtain the step-length corresponding with current detection yardstick, and exports check point monitoring unit to; According to the second triggering of check point monitoring unit, in the step-length that renewal current detection yardstick is corresponding the step-length array of write storage unit; According to the 3rd triggering of check point monitoring unit, the testing result that check point monitoring unit is exported is in the step-length array of step-length write storage unit corresponding to current detection yardstick;
Check point monitoring unit, the position of reading a measuring point to be checked from storage unit is offset, detects yardstick rank and identification information, and output first is toggled to step-length control module, whether the step-length corresponding with current detection yardstick that judges step-length control module feedback is 0, if step-length is 0, by image local area corresponding to measuring point to be checked being loaded on from internal memory Cache buffer memory, or image local area corresponding to measuring point to be checked in Cache buffer memory exports the Weak Classifier under current detection yardstick to, if step-length is non-zero and determine that according to identification information measuring point to be checked is positioned at first place and puts, by image local area corresponding to measuring point to be checked being loaded on from internal memory Cache buffer memory, or image local area corresponding to measuring point to be checked in Cache buffer memory exports the Weak Classifier under current detection yardstick to, if step-length is non-zero and determine that according to identification information measuring point to be checked is not positioned at first place and puts, output second is toggled to step-length and controls single,
Described check point monitoring unit is transmitted to step-length control module by the testing result of Weak Classifier output, and exports the 3rd and be toggled to step-length control module;
Storage unit, in order to preserve measuring point queue to be checked and step-length array.
As seen from the above technical solutions, the invention provides a kind of Adaboost detection method based on multiple dimensioned, in the method, according to position skew, from the measuring point queue to be checked of the information that comprises X*N measuring point to be checked, obtain the information of a measuring point to be checked, whether step-length corresponding to current detection yardstick that judgement is obtained from the step-length array that comprises N array element is 0, if step-length is 0, by image local area corresponding to measuring point to be checked being loaded on from internal memory Cache buffer memory, or the image that detects as Weak Classifier of image local area corresponding to the measuring point to be checked of having preserved in Cache buffer memory, if step-length is non-zero and measuring point to be checked is positioned at first place and puts, by image local area corresponding to measuring point to be checked being loaded on from internal memory Cache buffer memory, or the image that detects as Weak Classifier of image local area corresponding to the measuring point to be checked of having preserved in Cache buffer memory, if step-length is non-zero and measuring point to be checked is not positioned at first place and puts, upgrade the step-length corresponding to current detection yardstick of preserving in step-length array.The present invention also provides a kind of Adaboost detection system based on multiple dimensioned.Adopt method and system of the present invention, can improve Adaboost detection efficiency, improve the handling property of system.
Accompanying drawing explanation
Fig. 1 is the process flow diagram that the present invention is based on multiple dimensioned Adaboost detection method.
Fig. 2 is the structural representation that the present invention is based on multiple dimensioned Adaboost detection system.
Embodiment
For making object of the present invention, technical scheme and advantage clearer, referring to the accompanying drawing embodiment that develops simultaneously, the present invention is described in more detail.
The invention provides a kind of Adaboost detection method and system based on multiple dimensioned, the present invention is in order to improve Adaboost detection efficiency, when testing image being carried out to the detection of N detection yardstick, the method that no longer adopts existing N traversal to detect to the X of testing image measuring point to be checked, but once in traversal detection, X measuring point to be checked carried out to the detection of Weak Classifier, and judge whether to carry out Weak Classifier while detecting for arbitrary check point, according to previous step-length of having carried out the dot generation to be detected of Weak Classifier detection under current detection yardstick, determine whether the measuring point current to be checked under current detection yardstick to carry out the judgement of Weak Classifier detection, in the present invention, except being positioned at the arbitrary to be checked measuring point of first place putting and carrying out the detection number of times of Weak Classifier detection, be less than N time, because measuring point to be checked is at the Cache Miss that carries out can causing when Weak Classifier detects image local area, above-mentioned detection method of the present invention can effectively reduce image local area Cache Miss quantity, such as, in existing system of carrying out object detection, if will there is Cache Miss one time in average every 4 adjacent check points, the number of times that Cache Miss occurs while adopting the existing Adaboost detection method based on multiple dimensioned to detect N testing image that detects yardstick is numerical value after the business of X and 4 rounds up and the product of N, and while adopting the Adaboost detection method based on multiple dimensioned of the present invention to detect N testing image that detects yardstick, the number of times that Cache Miss occurs is the numerical value that the business of X and 4 rounds up, the Cache Miss of image local area is much smaller than existing method, effectively improved the handling property of detection efficiency and system.
Fig. 1 is the process flow diagram that the present invention is based on multiple dimensioned Adaboost detection method.Now, in conjunction with Fig. 1, to the present invention is based on multiple dimensioned Adaboost detection method, describe, specific as follows:
Step 101: generate N step-length array that detects yardstick;
In this step, step-length array comprises N array element, and each array element is in order to preserve the step-length under this detection yardstick, under current detection yardstick the previous real check point of current measuring point to be checked to the step-length between next real check point.Wherein, step-length is the number of the check point at interval between adjacent two check points that need detect through Weak Classifier, in order to represent the number of the not check point at interval between adjacent two real check points; Real check point is the check point through having judged that Weak Classifier detects; Check point is not without the check point that carries out Weak Classifier detection through judgement; Measuring point to be checked is the check point that process Weak Classifier does not detect and current wait judges whether to carry out Weak Classifier detection.
This step is specially: the N comprising according to testing image numerical value that detects yardstick, generates in order to preserve the step-length array of step-length under arbitrary detection yardstick.
The size of the detection window of selecting when detection yardstick of the present invention represents to detect, can represent by the form of pixel product the size of detection window, such as 64 pixel * 64 pixels, and 64 pixel * 128 pixels etc.; Check point of the present invention is the pixel position on image to be detected, to the testing process of this check point, is exactly to the similarity comparison process between the corresponding image local area of this check point and sample image; N and X are the integer that is greater than 1.
In this step, the initial value of the step-length that in step-length array, arbitrary detection yardstick is corresponding is 0, first measuring point to be checked of giving tacit consent under arbitrary detection yardstick all needs to carry out Weak Classifier detection, numerical value in subsequent step recycling step-length numeral and of the present invention each measuring point to be checked is carried out to N detection method that detects yardstick one by one, whether treat check point needs to carry out Weak Classifier detection and screens; Step-length under arbitrary detection yardstick can be carried out dynamic change according to the testing result of Weak Classifier, and the image local area that measuring point to be checked is corresponding is more similar to corresponding region in sample image, and step-length is just less.
Step 102: obtain X measuring point to be checked at N position skew, detection yardstick rank and the identification information detecting under yardstick, and generate measuring point queue to be checked;
In this step, position is offset in order to represent coordinate or the positional information of measuring point to be checked in testing image, and measuring point to be checked is with respect to the position offset of first pixel of testing image; The position of first pixel of testing image can be the edge, left side of testing image and upper side edge along the pixel intersecting; Detect yardstick rank in order to determine the relative size that detects yardstick, such as, default detection yardstick rank is 1 o'clock, the size of the detection window that this detection yardstick is corresponding is 64 pixel * 64 pixels, detecting yardstick rank is 2 o'clock, the size of the detection window that this detection yardstick is corresponding is 128 pixel * 128 pixels, according to detecting yardstick rank, can indirectly determine the relative size of detection window; Identification information is put in order to represent the surveyed area first place whether this measuring point to be checked is positioned at testing image, if testing image is detected line by line, surveyed area first place is set to the row head of certain a line, such as being positioned at capable first place with 1 value representation, put, with 0 value representation, be positioned at non-row first place and put, if testing image is detected by column, surveyed area first place is set to the row head of a certain row, such as be positioned at row first place with 1 value representation, put, with 0 value representation, be positioned at non-row first place and put.
This step comprises: step 1021, generates an array that comprises X*N array element; Step 1022, detects under yardstick testing image is scanned at each, obtains position skew, detection yardstick rank and the identification information of each measuring point to be checked under current detection yardstick; Step 1023, in the array that each measuring point to be checked that scanning is obtained generates N position skew, detection yardstick rank and the identification information write step 1021 detecting under yardstick; Step 1024, by the array that has write position skew in step 1023, detected yardstick rank and identification information, according to the arrangement of position skew ascending order, the array after arranging according to position skew ascending order is as measuring point queue to be checked.
Wherein, step 1022 can adopt existing image detecting method to scan testing image, to obtain position skew, detection yardstick rank and the identification information of each measuring point to be checked.
Step 103: obtain a measuring point Pi to be checked according to position skew from measuring point queue to be checked;
This step is specially: the position skew, detection yardstick rank and the identification information that in the measuring point queue to be checked from arranging according to position skew ascending order, obtain a check point Pi.
I in this step is more than or equal to 1 integer.
Step 104: obtain the step-length corresponding with current detection yardstick from step-length array;
The step-length corresponding with current detection yardstick of mentioning in this step is the step-length generating according to the front testing result that once completes Weak Classifier detection under current detection yardstick or the step-length that the step-length generating according to the front testing result that once completes Weak Classifier detection under current detection yardstick has been carried out to renewal.
This step is specially: according to current detection yardstick rank, obtain the array element corresponding with detecting yardstick, using the numerical value of preserving in array element as the step-length corresponding with current detection yardstick from step-length array.
Step 105: judge whether the step-length under current detection yardstick is 0, if so, execution step 108, otherwise execution step 106;
This step specifically judges whether the numerical value of preserving in array element corresponding with current detection yardstick in step-length array is 0, if, represent that current measuring point Pi to be checked need to carry out Weak Classifier detection, otherwise, represent that current measuring point Pi to be checked needs further judgement, to determine that it becomes still check point not of real check point.
Step 106: according to identification information judgment measuring point Pi to be checked, whether be positioned at first place and put, if so, execution step 108, otherwise execution step 107;
This step and step 105 combination, to determine whether Pi to be detected carries out Weak Classifier detection, particularly, when the step-length non-zero corresponding with current detection yardstick, as long as measuring point Pi to be checked is positioned at, put the row first place of a line or put the row first place of row, just carry out Weak Classifier detection, otherwise skip to according to step-length the measuring point to be checked that the next one may become real check point, the measuring point to be checked that may be called real check point to the next one judges.
For detecting line by line, this step is the judgement whether first measuring point to be checked of row that the step-length impact that produces after Weak Classifier detects for fear of the measuring point to be checked that is positioned at the end of line of previous row is positioned at current line carries out Weak Classifier detection, in other words, the measuring point to be checked of putting for row first place, adjacent measuring point to be checked after only existing, there is no front adjacent measuring point to be checked, also the step-length that does not just exist previous real check point to form in this row, and in order to guarantee testing result accuracy, the present invention carries out Weak Classifier detection to the first check point of the row of every a line.
Step 107: upgrade step-length corresponding to current detection yardstick, perform step afterwards 103;
This step comprises: the step-length using the difference of step-length corresponding to current detection yardstick and 1 after upgrading, the step-length after upgrading is write in array element corresponding with current detection yardstick in step-length array, and perform step afterwards 103.
Step 108: the Weak Classifier under use current detection yardstick is treated check point Pi and detected;
This step comprises: step 1081, according to current detection yardstick, obtain with current detection yardstick under Weak Classifier; Step 1082, obtains the image local area corresponding with measuring point Pi to be checked; Step 1083, utilizes Weak Classifier to treat the characteristics of image that image local area that check point Pi is corresponding comprises and carries out similarity detection.
In step 1082, if image local area corresponding to measuring point Pi to be checked is stored in Cache buffer memory, there is not Cache Miss, in step 1083, export image local area corresponding to measuring point Pi to be checked of preserving in Cache buffer memory to Weak Classifier; In step 1082, if image local area corresponding to measuring point Pi to be checked is not stored in Cache buffer memory, there is Cache Miss 1 time, image local area corresponding to measuring point Pi to be checked in internal memory is loaded in Cache buffer memory, so that in step 1083, export image local area corresponding to measuring point Pi to be checked of preserving in Cache buffer memory to Weak Classifier.
Step 109: using testing result as step-length corresponding to current detection yardstick, perform step afterwards 103;
This step is specially: the testing result of measuring point Pi to be checked being carried out to Weak Classifier detection writes in array element corresponding with current detection yardstick in step-length array as step-length.
After this step, measuring point Pi to be checked becomes real check point.
Wherein, treat the similarity degree of respective image feature in characteristics of image that testing result that check point Pi carries out Weak Classifier detection represented that image local area corresponding to measuring point Pi to be checked comprises and sample image, such as can be by the uncorrelated feature of the similarity degree of the two, the weak correlated characteristic of part and strong correlation character representation, more approach strong correlation feature, the testing result that measuring point Pi to be checked detects through Weak Classifier represents more similar, the step-length that current detection yardstick is corresponding is less, such as, when if the testing result that under current detection yardstick, the Weak Classifier of measuring point Pi to be checked detects represents strong correlation feature, the step-length that current detection yardstick is corresponding is 0, if the testing result that under current detection yardstick, the Weak Classifier of measuring point Pi to be checked detects represents the weak correlated characteristic of part, the step-length that current detection yardstick is corresponding is 1, if the testing result that under current detection yardstick, the Weak Classifier of measuring point Pi to be checked detects represents uncorrelated feature, the step-length that current detection yardstick is corresponding is 3.
The similarity degree that the step-length that the present invention mentions and testing result represent is inversely proportional to, be actually for fear of when the follow-up selection next one carries out the measuring point to be checked of Weak Classifier detection, because step-length is selected the improper correctness that affects testing result, in other words, for general testing image, it between check point, is Existential Space correlativity, if a check point is very similar to sample, around it, contiguous image local area corresponding to check point also can be more similar to sample, now step-length should arrange smaller, in order to avoid skip those check points the most similar to sample, and affect the correctness of testing result, therefore, the similarity degree that the present invention's step-length is set to represent with testing result is inversely proportional to.
Preferably, before step 101 of the present invention, further comprise: step 100, generates with each and detect the Weak Classifier that yardstick is corresponding according to sample image; In this step, Weak Classifier under a certain detection yardstick is according to the training set of the sample image under this detection yardstick, utilize Adaboost algorithm to calculate to obtain, can adopt existing Adaboost algorithm to carry out interative computation generation for the Weak Classifier of a certain detection yardstick, at this, no longer the concrete computation process of Weak Classifier be repeated.
If in system will there is Cache Miss one time in average every M check point, adopt in the process that the Adaboost detection method based on multiple dimensioned of the present invention detects testing image, the number of times that Cache Miss occurs is the numerical value that the business of X and M rounds up.The M of now only take is 3 for example as 4, X is greater than 4, N, and the Adaboost detection method based on multiple dimensioned of the present invention is described, specific as follows:
In the present embodiment, M is 4, and Cache Miss can occur one time average every 4 check points, that is to say in the image local area that measuring point P1 to be checked is corresponding and can comprise measuring point P2 to be checked, measuring point P3 to be checked and image local area corresponding to measuring point P4 to be checked; In the present embodiment, suppose that check point P1 and regional area corresponding to check point Pa are not stored in Cache buffer memory; Wherein, a is 4j+1; J is more than or equal to 1 integer.
For position skew be 0 in the first measuring point P1 to be checked of row, need to carry out N Weak Classifier that detects yardstick detects, N testing result of the characteristics of image that image local area corresponding to acquisition this measuring point P1 to be checked of expression comprises and the characteristics of image similarity of sample image, carries out assignment according to N testing result to N array element corresponding with detecting yardstick in step-length array.
Because image local area corresponding to measuring point P1 to be checked is not stored in Cache buffer memory, there is 1 time Cache Miss, before detecting through Weak Classifier, from internal memory, load the image local area corresponding with measuring point P1 to be checked to Cache buffer memory, utilize Weak Classifier to treat the image local area that check point P1 is corresponding and carry out 3 detections that detect yardstick, according to testing result, detecting yardstick rank, be 1 and to detect yardstick rank be that 2 o'clock corresponding step-lengths are 0, detecting yardstick rank, be that 3 o'clock corresponding step-lengths are 2, P1 to be detected becomes the front once real check point of measuring point P2 to be checked.
Position skew be 1 and not in the first measuring point P2 to be checked of row because real check point P 1 is 1 and to detect yardstick rank be that the step-length of 2 o'clock is 0 detecting yardstick rank, measuring point P2 to be checked is 1 and to detect yardstick rank be all to need to carry out Weak Classifier detection at 2 o'clock detecting yardstick rank, utilize image local area corresponding to measuring point P1 to be checked of having preserved in Cache buffer memory to carry out Weak Classifier detection, now can there is not Cache Miss, the testing result that measuring point P2 to be checked is 1 o'clock in detection yardstick rank is weak dependence, the step-length that detects yardstick rank in step-length array and be at 1 o'clock is updated to 1, the testing result that measuring point P2 to be checked is 2 o'clock in yardstick rank is irrelevance, the step-length that detects yardstick rank in step-length array and be at 2 o'clock is updated to 3, measuring point P2 to be checked is 1 and to detect yardstick rank be to become real check point at 2 o'clock detecting yardstick rank, the measuring point P2 to be checked in row head is because real check point P1 is that 3 o'clock corresponding step-lengths are 2 detecting yardstick rank, the characteristics of image that image local area corresponding to measuring point P2 to be checked comprises may be corresponding from sample image characteristics of image exist certain different, the measuring point P2 to be checked that the present invention is no longer 3 o'clock to detection yardstick rank carries out Weak Classifier detection, upgrading in step-length array with detecting yardstick rank is 3 o'clock corresponding array elements, be about to array element value corresponding to this detection yardstick and subtract 1, step value is updated to 1 by 2.
Position skew is 2 and in the first measuring point P3 to be checked of row, detecting yardstick rank, is not 1 o'clock, because the testing result of real check point P2 is non-zero, does not carry out Weak Classifier detection, and is that 1 step-length is updated to 0 by detecting yardstick rank in step-length array; Measuring point P3 to be checked is 2 o'clock detecting yardstick rank, because the testing result of real check point P2 is non-zero, does not carry out Weak Classifier detection, and is that 2 step-length is updated to 2 by detecting yardstick rank in step-length array; Measuring point P3 to be checked is 3 o'clock detecting yardstick rank, owing to being that 3 corresponding array elements are 1 with detecting yardstick rank in step-length array, be that step-length is 1, measuring point P3 to be checked does not carry out Weak Classifier detection, and upgrading in step-length array with detecting yardstick rank is that 3 corresponding step-lengths are 0.
Position skew is 4 and not in the first check point P4 of row, owing to detecting yardstick, be that 1 o'clock corresponding step-length is 0, carry out Weak Classifier detection, utilize image local area corresponding to measuring point P1 to be checked of having preserved in Cache buffer memory to carry out Weak Classifier detection, now still can there is not Cache Miss, owing to detecting yardstick, be that 2 o'clock corresponding step-lengths are non-zero, under this detection yardstick also without carrying out Weak Classifier detection, by detecting yardstick rank in step-length array, be that 2 step-length is updated to 1, owing to being that 3 o'clock step-lengths are 0 detecting yardstick rank, need to carry out Weak Classifier detection, utilize image local area corresponding to measuring point P1 to be checked of having preserved in Cache buffer memory to carry out Weak Classifier detection, now still can there is not Cache Miss.
And for the measuring point Pa to be checked after measuring point P4 to be checked, such as measuring point P5 to be checked, image local area corresponding to measuring point P1 to be checked in Cache buffer memory do not comprise image local area corresponding to measuring point P5 to be checked, image local area corresponding to measuring point P5 to be checked need to be loaded into Cache buffer memory from internal memory, replace the image local area of measuring point P 1 correspondence to be checked of preserving in Cache buffer memory, now there is Cache Miss 1 time, the image local area corresponding to measuring point P5 to be checked of buffer memory can provide measuring point P5 to be checked to detect required image to the Weak Classifier of measuring point P8 to be checked.
Known according to foregoing, when utilizing method of the present invention to carry out the detection of N detection yardstick to X measuring point to be checked, if every M check point occurs once, the number of times of whole testing process generation Cache Miss is the numerical value after the business of X and M rounds up.
Fig. 2 is the structural representation that the present invention is based on multiple dimensioned Adaboost detection system.Now, in conjunction with Fig. 2, to the present invention is based on the structure of multiple dimensioned Adaboost detection system, describe, specific as follows:
The present invention is based on multiple dimensioned Adaboost detection system comprises: detect control module 21 and N Weak Classifier 22.Wherein, detect control module 21 and connect N Weak Classifier 22, internal memory (not shown in Fig. 2) and Cache buffer memory 23.
The step-length array corresponding with the individual measuring point to be checked of X that detects N detection yardstick of control module 21 generation is also temporary, obtain X measuring point to be checked that testing image comprises and at N, detect position skew under yardstick, detect yardstick rank and identification information, generate measuring point queue to be checked and keep in.
Detect control module 21 and according to position skew, from measuring point queue to be checked, obtain the position skew of a measuring point Pi to be checked, detect yardstick rank and identification information, obtain the step-length that current detection yardstick is corresponding, judging that the step-length that current detection yardstick is corresponding is 0, or judging that the non-zero and measuring point Pi to be checked of step-length that current detection yardstick is corresponding is when put first place, by image local area corresponding to measuring point Pi to be checked being loaded on from internal memory Cache buffer memory 23, or image local area corresponding to measuring point Pi to be checked in Cache buffer memory 23 exports the Weak Classifier 22 under current detection yardstick to, when judging that the non-zero and measuring point Pi to be checked of step-length that current detection yardstick is corresponding is not positioned at first place and puts, upgrade step-length corresponding to array element corresponding with current detection yardstick in temporary step-length array.Wherein, put the row first place that first place is set in detecting line by line, or put the row first place in detecting by column.
The characteristics of image of arbitrary Weak Classifier 22 in order to comprise according to the image local area corresponding with measuring point Pi to be checked of input in N Weak Classifier 22, the characteristics of image that under generation expression current detection yardstick, image local area corresponding to measuring point Pi to be checked comprises and the testing result of the corresponding characteristics of image similarity degree of sample image, export to testing result as the step-length corresponding with current detection yardstick and detect control module 21.Wherein, a N of the present invention Weak Classifier 22 can generate according to existing Adaboost algorithm and sample image, at this, no longer the structure of Weak Classifier is repeated.
The Cache buffer memory 23 that the present invention mentions comprises the image local area corresponding with measuring point to be checked in order to preserve; Wherein, in the image local area corresponding with measuring point to be checked, can comprise a plurality of measuring points to be checked, should can by external memory, be loaded in Cache buffer memory 23 and preserve by detecting control module 21 with a corresponding image local area to be detected, with the measuring point a plurality of to be checked that this image local area is comprised, carry out Weak Classifier detection.
Wherein, detect control module 21 and comprise scanning element 210, step-length control module 211, check point monitoring unit 212 and storage unit 213.
Scanning element 210 is according to the numerical value of N of the detection yardstick of check point monitoring unit 212 outputs and the measuring point number X to be checked that testing image comprises, generate an array that comprises X*N array element, at each, detect under yardstick testing image is scanned, obtain the position skew of each measuring point to be checked under current detection yardstick, detect yardstick rank and identification information, each measuring point to be checked that scanning is obtained is N the position skew detecting under yardstick, detecting yardstick rank and identification information writes in the array of generation, to write position skew, element in the array of detection yardstick rank and identification information is arranged according to position skew ascending order, array after arranging according to position skew ascending order is in measuring point queue write storage unit 213 to be checked.
Step-length control module 211, according to the numerical value of N of the detection yardstick of check point monitoring unit 212 output, generates in order to preserve N step-length array that detects yardstick, and initial value of setting the step-length of each detection yardstick is 0, by step-length array write storage unit 213.
Step-length control module 211 triggers according to first of check point monitoring unit 212, utilizes current detection yardstick rank from storage unit 213, to obtain the step-length corresponding with current detection yardstick, and exports check point monitoring unit 212 to; Trigger according to second of check point monitoring unit 212, using the difference of the step-length corresponding with current detection yardstick and 1 in upgrading afterwards step-length write storage unit 213 with array element corresponding to current detection yardstick in; Trigger according to the 3rd of check point monitoring unit 212 the, the testing result from Weak Classifier 22 that check point monitoring unit 212 is forwarded is in step-length write storage unit 213 corresponding to current detection yardstick in the array element corresponding with current detection yardstick.
Check point monitoring unit 212 is when starting initialization, and the measuring point number X to be checked that the numerical value of N of output detections yardstick and testing image comprise is to scanning element 210 and step-length control module 211; Check point monitoring unit 212 is according to the measuring point queue to be checked generating according to the arrangement of position skew ascending order in storage unit 213, the position skew, detection yardstick rank and the identification information that read a measuring point Pi to be checked, output first is toggled to step-length control module 211.
Check point monitoring unit 212 judges whether the step-length corresponding with current detection yardstick of step-length control module 211 feedbacks is 0, if step-length is 0, judge and in Cache buffer memory 23, whether preserved the image local area corresponding with measuring point Pi to be checked, if preserved, export image local area that the measuring point Pi to be checked that reads is corresponding to the Weak Classifier 22 under current detection yardstick from Cache buffer memory 23, if do not preserved, from internal memory, load the image local area corresponding with measuring point Pi to be checked to Cache buffer memory 23, image local area corresponding to measuring point Pi to be checked that output is read from Cache buffer memory 23 is to the Weak Classifier 22 under current detection yardstick.
Check point monitoring unit 212 is after judging that step-length is non-zero, further according to the identification information judgment of measuring point Pi to be checked measuring point Pi to be checked, whether being positioned at first place puts, if being positioned at first place puts, judge and in Cache buffer memory 23, whether preserved the image local area corresponding with measuring point Pi to be checked, if preserved, export image local area that the measuring point Pi to be checked that reads is corresponding to the Weak Classifier 22 under current detection yardstick from Cache buffer memory 23, if do not preserved, from internal memory, load the image local area corresponding with measuring point Pi to be checked to Cache buffer memory 23, image local area corresponding to measuring point Pi to be checked that output is read from Cache buffer memory 23 is to the Weak Classifier 22 under current detection yardstick, if do not put in first place, output second is toggled to step-length control module 211.
Check point monitoring unit 212 is transmitted to step-length control module 211 by the testing result of Weak Classifier 22 output, and exports the 3rd and be toggled to step-length control module 211.
Storage unit 213 is in order to preserve measuring point queue to be checked and step-length array.
Preferably, system of the present invention also comprises a sorter generation module (not shown in Fig. 2), and this sorter generation module utilizes Adaboost algorithm and sample image, generates with N and detects N the Weak Classifier 22 that yardstick is corresponding.
In above-mentioned preferred embodiment of the present invention, existing multi-dimension testing method does not detect when yardstick detects and travels through N time N for another example, but completed by a traverse scanning detection that detects yardstick to N of X measuring point to be checked, like this, the number of times that adopts method of the present invention to carry out occurring in testing process Cache Miss is the numerical value after the business of X and M rounds up, compared to adopting existing multi-dimension testing method to carry out Miss the number of Cache occurring in testing process, be numerical value after the business of X and M rounds up and the product of N, greatly reduced the number of times from internal memory load image regional area to Cache buffer memory, Adaboost detection efficiency and system performance have been improved.
The foregoing is only preferred embodiment of the present invention, in order to limit the present invention, within the spirit and principles in the present invention not all, any modification of making, be equal to replacement, improvement etc., within all should being included in the scope of protection of the invention.
Claims (7)
1. the Adaboost detection method based on multiple dimensioned, is characterized in that, the method comprises:
A ', the step-length array that generates N detection yardstick and X measuring point to be checked are N the measuring point queue to be checked detecting under yardstick; The described step-length array that generates N detection yardstick comprises: generate the step-length array that comprises N array element, an array element detects step-length corresponding to yardstick in order to preserve one; The initial value of the step-length that arbitrary array element is preserved is 0; The information that described queue to be detected comprises X*N measuring point to be checked, the information of described measuring point to be checked at least comprises position skew, detects yardstick rank and identification information; Described N and X are the integer that is greater than 1; The skew of described position is in order to represent coordinate or the positional information of check point in image to be detected; Described detection yardstick rank is in order to determine the relative size that detects yardstick; Described identification information is put in order to represent the first place whether check point is positioned at testing image;
A, according to position skew, from the measuring point queue to be checked of the information that comprises X*N measuring point to be checked, obtain the information of a measuring point to be checked;
Whether step-length corresponding to current detection yardstick that B, judgement are obtained from the step-length array that comprises N array element is 0, if so, and execution step D, otherwise execution step C; Described step-length is the number of the measuring point to be checked at interval between adjacent two measuring points to be checked that need detect through Weak Classifier;
C, according to identification information judgment measuring point to be checked, whether be positioned at first place and put, if so, execution step D, otherwise, upgrade the step-length corresponding to current detection yardstick of preserving in step-length array, execution step A;
D, judge whether Cache buffer memory has preserved the image local area corresponding with measuring point to be checked, if do not preserved, from internal memory, load image local area that measuring point to be checked is corresponding to Cache buffer memory, perform step afterwards E, otherwise directly perform step E;
E, utilize the Weak Classifier under current detection yardstick, the image local area corresponding with measuring point to be checked of preserving in Cache buffer memory detected, testing result is write to described step-length array as step-length corresponding to current detection yardstick.
2. method according to claim 1, is characterized in that, before described steps A, further comprises:
According to sample image and Adaboost algorithm, generate with N and detect the Weak Classifier that in yardstick, each detection yardstick is corresponding.
3. method according to claim 1, is characterized in that, steps A ' the to be checked measuring point queue of described generation X measuring point to be checked under N detection yardstick comprise:
A1 ', an array that comprises X*N array element of generation;
A2 ', under each detects yardstick, testing image is scanned, obtain each check point under current detection yardstick position skew, detect yardstick rank and identification information;
In an array element of the array that A3 ', each check point that scanning is obtained generate at N position skew, detection yardstick rank and the identification information write step A1 ' detecting under yardstick;
A4 ', the array that has write the information of X*N measuring point to be checked in steps A 3 ' is arranged according to position skew ascending order, the array after arranging according to position skew ascending order is as measuring point queue to be checked.
4. method according to claim 1 and 2, is characterized in that, described in step C the first be set to detect line by line in the row first place of a line put or detect by column in the row first place of row put.
5. method according to claim 1 and 2, is characterized in that, upgrades the step-length corresponding to current detection yardstick of preserving in step-length array described in step C to comprise:
Step-length using the difference of step-length corresponding to current detection yardstick and 1 after upgrading, writes the step-length after upgrading in array element corresponding with current detection yardstick in step-length array.
6. the Adaboost detection system based on multiple dimensioned, this system connects internal memory and Cache buffer memory, it is characterized in that, this system inclusion test monitoring module and N Weak Classifier, described detection control module connects internal memory, Cache buffer memory and N Weak Classifier, and described detection control module comprises scanning element, step-length control module, check point monitoring unit and storage unit;
Scanning element, generates X measuring point to be checked N the detection queue detecting under yardstick, by the measuring point queue write storage unit to be checked of the information that comprises X*N measuring point to be checked; The information of described measuring point to be checked at least comprises position skew, detects yardstick rank and identification information; Described N and X are the integer that is greater than 1; The skew of described position is in order to represent coordinate or the positional information of check point in image to be detected; Described detection yardstick rank is in order to determine the relative size that detects yardstick; Described identification information is put in order to represent the first place whether check point is positioned at testing image;
Step-length control module, according to the N of check point monitoring unit output, generate in order to preserve N step-length array that detects yardstick, and by step-length array write storage unit, described generation comprises in order to preserve N step-length array that detects yardstick: generate the step-length array that comprises N array element, an array element detects step-length corresponding to yardstick in order to preserve one, and the initial value of the step-length that arbitrary array element is preserved is 0; The first triggering according to check point monitoring unit, utilizes current detection yardstick rank from storage unit, to obtain the step-length corresponding with current detection yardstick, and exports check point monitoring unit to; According to the second triggering of check point monitoring unit, in the step-length that renewal current detection yardstick is corresponding the step-length array of write storage unit; According to the 3rd triggering of check point monitoring unit, the testing result that check point monitoring unit is exported is in the step-length array of step-length write storage unit corresponding to current detection yardstick;
Check point monitoring unit, the position of reading a measuring point to be checked from storage unit is offset, detects yardstick rank and identification information, and output first is toggled to step-length control module, whether the step-length corresponding with current detection yardstick that judges step-length control module feedback is 0, if step-length is 0, by image local area corresponding to measuring point to be checked being loaded on from internal memory Cache buffer memory, or image local area corresponding to measuring point to be checked in Cache buffer memory exports the Weak Classifier under current detection yardstick to, if step-length is non-zero and determine that according to identification information measuring point to be checked is positioned at first place and puts, by image local area corresponding to measuring point to be checked being loaded on from internal memory Cache buffer memory, or image local area corresponding to measuring point to be checked in Cache buffer memory exports the Weak Classifier under current detection yardstick to, if step-length is non-zero and determine that according to identification information measuring point to be checked is not positioned at first place and puts, output second is toggled to step-length control module,
Described check point monitoring unit is transmitted to step-length control module by the testing result of Weak Classifier output, and exports the 3rd and be toggled to step-length control module;
Storage unit, in order to preserve measuring point queue to be checked and step-length array;
In a described N Weak Classifier, arbitrary Weak Classifier, according to the image local area corresponding with measuring point to be checked of input, generates the testing result of presentation video similarity degree, exports testing result to detection control module as the step-length corresponding with current detection yardstick.
7. system according to claim 6, it is characterized in that, described scanning element, according to N and the X of the output of check point monitoring unit, generate an array that comprises X*N array element, at each, detect under yardstick testing image is scanned, obtain position skew, detection yardstick rank and the identification information of each check point under current detection yardstick and write in the array of generation, the array after arranging according to position skew ascending order is as measuring point queue write storage unit to be checked.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210169536.7A CN102722723B (en) | 2012-05-24 | 2012-05-24 | Multi-scale-based Adaboost detection method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210169536.7A CN102722723B (en) | 2012-05-24 | 2012-05-24 | Multi-scale-based Adaboost detection method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102722723A CN102722723A (en) | 2012-10-10 |
| CN102722723B true CN102722723B (en) | 2014-11-05 |
Family
ID=46948473
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210169536.7A Active CN102722723B (en) | 2012-05-24 | 2012-05-24 | Multi-scale-based Adaboost detection method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102722723B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273923B (en) * | 2017-06-02 | 2020-09-29 | 浙江理工大学 | Construction method of textile fabric friction sound wave discriminator |
| CN109961079B (en) * | 2017-12-25 | 2021-06-04 | 北京君正集成电路股份有限公司 | Image detection method and device |
| CN109490904B (en) * | 2018-11-15 | 2021-09-21 | 上海炬佑智能科技有限公司 | Time-of-flight sensor and detection method thereof |
| CN112686115A (en) * | 2020-12-24 | 2021-04-20 | 中国矿业大学(北京) | Pedestrian target detection method in automatic driving scene based on scale perception |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101350063A (en) * | 2008-09-03 | 2009-01-21 | 北京中星微电子有限公司 | Method and apparatus for locating human face characteristic point |
| CN101645135A (en) * | 2009-07-13 | 2010-02-10 | 北京中星微电子有限公司 | Method and system for human detection |
| CN102024149A (en) * | 2009-09-18 | 2011-04-20 | 北京中星微电子有限公司 | Method of object detection and training method of classifier in hierarchical object detector |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7664328B2 (en) * | 2005-06-24 | 2010-02-16 | Siemens Corporation | Joint classification and subtype discovery in tumor diagnosis by gene expression profiling |
| TW200842733A (en) * | 2007-04-17 | 2008-11-01 | Univ Nat Chiao Tung | Object image detection method |
-
2012
- 2012-05-24 CN CN201210169536.7A patent/CN102722723B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101350063A (en) * | 2008-09-03 | 2009-01-21 | 北京中星微电子有限公司 | Method and apparatus for locating human face characteristic point |
| CN101645135A (en) * | 2009-07-13 | 2010-02-10 | 北京中星微电子有限公司 | Method and system for human detection |
| CN102024149A (en) * | 2009-09-18 | 2011-04-20 | 北京中星微电子有限公司 | Method of object detection and training method of classifier in hierarchical object detector |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102722723A (en) | 2012-10-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101517612B (en) | Image reading device and image reading method | |
| CN106716441B (en) | Event-based computer vision computing | |
| CN102722723B (en) | Multi-scale-based Adaboost detection method and system | |
| KR101546590B1 (en) | Hough transform for circles | |
| CN107209942B (en) | Object detection method and image retrieval system | |
| CN102870399A (en) | Segmentation of a word bitmap into individual characters or glyphs during an OCR process | |
| US20130170749A1 (en) | Method and apparatus for document image indexing and retrieval using multi-level document image structure and local features | |
| US20090245648A1 (en) | Ridge direction extracting device, ridge direction extracting program, and ridge direction extracting method | |
| US10261849B1 (en) | Preventative remediation of services | |
| US10325130B2 (en) | Predictive anomaly detection | |
| CN109086336A (en) | Paper date storage method, device and electronic equipment | |
| CN101536032A (en) | Image recognition apparatus and method | |
| CN105139508B (en) | A kind of method and device of detection bank note | |
| JP2019212157A (en) | Commodity specification device, program, and learning method | |
| KR102378086B1 (en) | Event detecting device | |
| CN108197622A (en) | A kind of detection method of license plate, device and equipment | |
| JPWO2011114474A1 (en) | Determination apparatus, determination system, determination method, and computer program | |
| WO2019088223A1 (en) | Detection device and detection program | |
| CN112215271A (en) | Anti-occlusion target detection method and device based on multi-head attention mechanism | |
| EP2884467B1 (en) | Image identification system and image storage control method | |
| CN111047579B (en) | Feature quality assessment method and image feature uniform extraction method | |
| JP6252296B2 (en) | Data identification method, data identification program, and data identification apparatus | |
| CN108182677B (en) | Prepress register detection method, prepress register detection device and computer readable storage medium | |
| US8428353B2 (en) | Calculation apparatus for amount of characteristic and discrimination apparatus | |
| JP2022014793A (en) | Information processing device, information processing method, and program |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |