CN107368832A - Target detection and sorting technique based on image - Google Patents
Target detection and sorting technique based on image Download PDFInfo
- Publication number
- CN107368832A CN107368832A CN201710617028.3A CN201710617028A CN107368832A CN 107368832 A CN107368832 A CN 107368832A CN 201710617028 A CN201710617028 A CN 201710617028A CN 107368832 A CN107368832 A CN 107368832A
- Authority
- CN
- China
- Prior art keywords
- wave filter
- well
- marked target
- score
- detected
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
This disclosure relates to a kind of target detection and sorting technique based on image, including:Image to be detected is pre-processed;Import the grader of training in advance;Feature extraction is carried out to each well-marked target in described image to be detected using the grader;Optimal location of each well-marked target in described image to be detected is determined according to feature extraction result;Marking area according to where each optimal location determines each well-marked target.The target detection and sorting technique based on image provided by the disclosure, using the target measurement handled based on threshold binary image, has been filtered and has been free of target area, accelerated detection process.Solve the problems, such as the translationization because of target object illumination, it is sized caused by image the complicated change such as shade of the color value of pixel and target object all increase the difficulty of target detection.Also, this method can effectively can be applied in Practical Project with real time execution.
Description
Technical field
This disclosure relates to big data analysis field, more particularly to a kind of target detection and sorting technique based on image.
Background technology
Well-marked target detection is the basic function of human vision, and one of Main Topics of computer vision.Mesh
Mark detection is always the importance of computer vision research, and well-marked target detection is driven in image retrieval, target identification, auxiliary
Sail many fields such as system and video monitoring and suffer from important application.Due in actual image data, target object is always not
It is evitable to be under relative complex background, therefore the target detection under complex background has been subjected to more and more computers and regarded
The researcher in feel field payes attention to, and as a most active field in computer vision research.On well-marked target
The detection classification of object mainly includes the method for the method and machine learning of non-machine learning.
The method of non-machine learning mainly includes threshold binary image processing detection, Hough transformation detection, stencil matching detection etc.,
This kind of method is mainly based upon rule and fixed template is detected, and detection method is excessively simple, therefore general only adaptation
In the project that oneself can control color and brightness.
The method of machine learning is detected by building different graders to target object, although complexity can be handled
Target classification and detection under scene, but in actual applications, because photo resolution is higher, in the process scanned to full figure
In, it is necessary to consume a longer time, therefore target detection is less efficient.Further, since the change of observation viewpoint, same target object
It can occur significantly to change in the picture, these changes may both show as yardstick, rotation, the difference of gradient, it is also possible to table
It is now the difference of perspective projection, causes the effect of detection poor.Finally, in most natural images, target object is not
It is present in simple background, on the contrary, various other objects may be included in background.The presence of complex background causes standard
Really and rapidly detection target becomes very difficult.Performance difference normally results between performance difference and small class in big class
The robustness of object detection method reduces.Performance difference refers to similar interindividual change in class, for example, the difference of people
Body has differences in color, texture, shape, posture etc..Due to illumination, background, posture, the change of viewpoint and the shadow that blocks
Ring, even if also appeared in different images can be very different for same person so that structure possesses the apparent mould of generalization ability
Type is extremely difficult.
The content of the invention
In view of this, the present disclosure proposes a kind of target detection and sorting technique based on image.
According to the one side of the disclosure, there is provided a kind of target detection and sorting technique based on image, including:To be checked
Altimetric image is pre-processed;Import the grader of training in advance;Using the grader to each aobvious in described image to be detected
Write target and carry out feature extraction;Determine that each well-marked target is optimal in described image to be detected according to feature extraction result
Position;Marking area according to where each optimal location determines each well-marked target.
In a kind of possible implementation, described image to be detected is pre-processed, including:To the to be detected of acquisition
Image carries out redundancy and removes processing and/or noise removal process;Each well-marked target in described image to be detected is obtained, and
Determine the initial position of each well-marked target.
In a kind of possible implementation, the grader includes root wave filter and multiple part wave filters, the profit
Feature extraction is carried out to each well-marked target in described image to be detected with the grader, including:Utilize described wave filter
Determine the root position of each well-marked target;Using each part wave filter, each part in each well-marked target is determined
Position;According to the score of each part wave filter, each part wave filter position relative to described position change
Shape cost and shift term determine the score of each well-marked target.
In a kind of possible implementation, according to the score of each part wave filter, each part wave filter institute
Spent in position relative to the deformation of described position and shift term determines the score of each well-marked target, including:
The score of the well-marked target is calculated using formula 1:
Wherein, score (p0,…,pn) for the score of well-marked target, b is shift term;For i-th
Part wave filter FiScore, Fi' it is that the vectorization of i-th part wave filter represents,For i-th of part wave filter
Fi' characteristic vector, pi=(xi,yi,li) represent i-th of wave filter where layer liWith horizontal position coordinate xiAnd upright position
Coordinate yi, H is characterized pyramid, and n represents the quantity of part wave filter;
Spent for i-th of part wave filter position relative to the deformation of described position;diI-th
Individual part wave filter position is spent relative to the deformation of itself anchor point position,
I-th of part wave filter relative to anchor point position displacementCalculated using formula 2:
The deformation behaviour of i-th of part wave filterUse formula 3 calculate:
Wherein, (x0,y0) be root wave filter layer where it coordinate, viIt is a bivector, represents i-th of part
The anchor point position of wave filter relative to root position coordinate,WithHorizontal displacement and the level of part wave filter are represented respectively
Square of displacement,WithSquare of vertical displacement and the vertical displacement of part wave filter is represented respectively.
In a kind of possible implementation, determine each well-marked target described to be detected according to feature extraction result
Optimal location in image, including:Obtain the optimal location for each part wave filter for detecting the well-marked target;According to each
The best position calculation of the part wave filter detects the comprehensive score of the root position of the root wave filter of the well-marked target;According to
The comprehensive score of described wave filter determines that described wave filter is optimal in described image to be detected in the well-marked target
Position.
In a kind of possible implementation, the comprehensive score according to described wave filter determines described wave filter
In optimal location of the well-marked target in described image to be detected, including:
The root position of each described wave filter is calculated in l using formula 80Comprehensive score score (the x of layer0,y0,l0), institute
State comprehensive score score (x0,y0,l0) in the case of highest, using the place root position of described wave filter of acquisition as described in
The optimal location of root wave filter:
Wherein, l0For the layer where described wave filter, x0, y0It is described wave filter in l0The position coordinates of layer, λ are
To obtain l0Layer twice of resolution ratio and need the number of plies walked downwards in feature pyramid H,
Wherein, l takes l0When calculate
Calculated by formula 7:
Wherein, Di,lWhen (x, y) value represents for the anchor point position of i-th of part wave filter to be placed on position (x, y) of l layers, the
I part wave filter takes l to the maximum contribution value of root position score, l0When calculate
In a kind of possible implementation, the comprehensive score according to described wave filter determines described wave filter
In optimal location of the well-marked target in described image to be detected, in addition to:
In the case where described wave filter is in the optimal location, is calculated using formula 9 and obtain each part filtering
The optimal location of device:
Wherein, the pi,l(x, y) is the position function of each wave filter.
It is described to be determined according to the optimal location and export each well-marked target institute in a kind of possible implementation
Marking area;Including:According to the optimal location, using bounding box Forecasting Methodology, each well-marked target place is determined
Each estimation range;The estimation range repeated is removed using non-maxima suppression method, where determining each well-marked target
Marking area;Notable area where according to the contextual information amendment of each well-marked target and exporting each well-marked target
Domain.
In a kind of possible implementation, methods described also includes:Notable mesh is not present in described image to be detected
In the case of target, described image to be detected is reduced, and image to be detected after diminution is pre-processed.
According to another aspect of the present disclosure, there is provided a kind of non-volatile computer readable storage medium storing program for executing, when the storage
When instruction in medium is by terminal and/or the computing device of server so that terminal and/or server are able to carry out a kind of base
In the target detection and sorting technique of image, methods described includes:Image to be detected is pre-processed;Import training in advance
Grader;Feature extraction is carried out to each well-marked target in described image to be detected using the grader;According to feature extraction
As a result optimal location of each well-marked target in described image to be detected is determined;Each institute is determined according to each optimal location
State the marking area where well-marked target.
The target detection and sorting technique based on image provided by the disclosure, utilizes the mesh handled based on threshold binary image
Mapping amount, filter and be free of target area, accelerated detection process.Solves the translation because of target object illumination, sized
Cause the problem of complicated change such as shade of the color value of pixel and target object all increases the difficulty of target detection in image.
Also, this method can effectively can be applied in Practical Project with real time execution.
According to below with reference to the accompanying drawings becoming to detailed description of illustrative embodiments, the further feature and aspect of the disclosure
It is clear.
Brief description of the drawings
Comprising in the description and the accompanying drawing of a part for constitution instruction and specification together illustrate the disclosure
Exemplary embodiment, feature and aspect, and for explaining the principle of the disclosure.
Fig. 1 shows the flow chart of a kind of target detection and sorting technique based on image according to the embodiment of the disclosure one;
Fig. 2 shows the flow of another target detection and sorting technique based on image according to the embodiment of the disclosure one
Figure.
Embodiment
Describe various exemplary embodiments, feature and the aspect of the disclosure in detail below with reference to accompanying drawing.It is identical in accompanying drawing
Reference represent the same or analogous element of function.Although the various aspects of embodiment are shown in the drawings, remove
Non-specifically point out, it is not necessary to accompanying drawing drawn to scale.
Special word " exemplary " is meant " being used as example, embodiment or illustrative " herein.Here as " exemplary "
Illustrated any embodiment should not necessarily be construed as preferred or advantageous over other embodiments.
In addition, in order to better illustrate the disclosure, numerous details is given in embodiment below.
It will be appreciated by those skilled in the art that without some details, the disclosure can equally be implemented.In some instances, for
Method, means, element and circuit well known to those skilled in the art are not described in detail, in order to highlight the purport of the disclosure.
Embodiment 1
Fig. 1 shows the flow chart of a kind of target detection and sorting technique based on image according to the embodiment of the disclosure one.
As shown in figure 1, the method comprising the steps of 11 to step 15.
Step 11, image to be detected is pre-processed.
It is described that image to be detected is pre-processed in a kind of possible mode, including:To the described to be detected of acquisition
Image carries out redundancy and removes processing and/or noise removal process;Each well-marked target in described image to be detected is obtained, and
Determine the initial position of each well-marked target.
In the present embodiment, pre-detection can be carried out to image to be detected using third party's method, in being highlighted with acquisition
Each well-marked target, and the initial position of each well-marked target is determined, it is easy to when subsequent classifier carries out feature extraction, according to initial
Position accurately determines the position of a well-marked target.
Step 12, the grader of training in advance is imported.
In a kind of possible implementation, the grader includes root wave filter and multiple part wave filters, root filtering
Device is used for the information for obtaining well-marked target, and part wave filter is used for the information for obtaining each part in well-marked target.
In the present embodiment, root wave filter can cover the region where whole well-marked target, and its resolution ratio is relatively low;And portion
Part wave filter only covers small parts in well-marked target region, and its resolution ratio is higher.Part wave filter is placed on root institute
Under the λ layers of layer, the resolution ratio of this layer of feature is twice of the feature of layer where root.
Step 13, feature extraction is carried out to each well-marked target in described image to be detected using the grader.
In a kind of possible implementation, it is described using the grader to each notable mesh in described image to be detected
Mark carries out feature extraction, including:The root position of each well-marked target is determined using described wave filter;Utilize each part
Wave filter, determine the position of each part in each well-marked target;According to the score of each part wave filter, each part
Wave filter position is spent relative to the deformation of described position and shift term determines the score of each well-marked target.
In a kind of possible implementation, the score according to each part wave filter, each part filter
Device position is spent relative to the deformation of described position and shift term determines the score of each well-marked target, including:
Displacement of i-th of part wave filter position relative to anchor point position is calculated using formula 2Calculated i-th using formula 3
The deformation behaviour of part wave filterScore score (the p of the well-marked target are calculated using formula 10,…,pn)。
Step 14, optimal position of each well-marked target in described image to be detected is determined according to feature extraction result
Put.
In a kind of possible implementation, step 14 can include:Obtain each portion for detecting the well-marked target
The optimal location of part wave filter;The root that the well-marked target is detected according to the best position calculation of each part wave filter filters
The comprehensive score of the root position of device;Determine described wave filter in the well-marked target according to the comprehensive score of described wave filter
Optimal location in described image to be detected.
In a kind of possible implementation, the root position of each described wave filter is calculated in l using formula 80The synthesis of layer
Score score (x0,y0,l0), in the comprehensive score score (x0,y0,l0) in the case of highest, by described filter of acquisition
Optimal location of the place root position of ripple device as described wave filter.
In a kind of possible implementation, the comprehensive score according to described wave filter determines described wave filter
In optimal location of the well-marked target in described image to be detected, in addition to:It is in described optimal in described wave filter
In the case of position, the optimal location for obtaining each part wave filter is calculated using formula 9.
Step 15, the marking area according to where each optimal location determines each well-marked target.
In a kind of possible implementation, step 15 can include:According to the optimal location, predicted using bounding box
Method, determine each estimation range where each well-marked target;The Target area repeated is removed using non-maxima suppression method
Domain, determine the marking area where each well-marked target;According to the contextual information amendment of each well-marked target and export
Marking area where each well-marked target.
In a kind of possible implementation, the above method also includes:Notable mesh is not present in described image to be detected
In the case of target, the size of described image to be detected is reduced, obtains each well-marked target in described image to be detected, and utilize institute
State grader and feature extraction is carried out to each well-marked target in image to be detected.
It should be noted that in the present embodiment formula used and its calculating process refer in embodiment 2 each formula and
Its calculating process.
Using concrete application scene as example, the above method is illustrated.Such as:In the information of storage multitude of video data
In the medium scene of security centre or cell monitoring room, its well-marked target acquisition process can be:Intercept what is locally preserved first
Crucial video image in video data or real-time video stream data, analyzes such vedio data, and data are entered
Row pretreatment, including but not limited to noise reduction, thresholding etc. operate.Then, pretreated view data is entered by grader
Row classification is handled, if detecting well-marked target, extracts the characteristic vector of each well-marked target, and according to characteristic vector pickup knot
Fruit determines the optimal location of a well-marked target.Finally, well-marked target place marking area is determined according to optimal location and exported.
It should be noted that although a kind of target detection based on image and classification are described using embodiment 1 as example
Method as above, it is understood by one of ordinary skill in the art that the disclosure answers not limited to this.In fact, user completely can be according to individual
Hobby and/or the flexible setting procedure flow of practical application scene, as long as meeting disclosed technique scheme.
The target detection and sorting technique based on image that the disclosure is provided, utilize the target handled based on threshold binary image
Measurement, has filtered and has been free of target area, accelerated detection process.Solve the translation because of target object illumination, sized make
The problem of complicated change such as shade of the color value of pixel and target object all increases the difficulty of target detection into image.And
And this method can effectively can be applied in Practical Project with real time execution.
Embodiment 2
Fig. 2 shows the flow chart according to the target detection of the disclosure one embodiment based on image and sorting technique.Such as Fig. 2
Shown, the method comprising the steps of 31 to step 37.
Step 31, image to be detected is pre-processed.
In the present embodiment, because image to be detected has more redundancy and noise, in order to optimizing detection
Efficiency, the scanning frequency of image to be detected is reduced, in step 31, can be detected in advance with other third party's methods to be detected
Each well-marked target in the presence of image, and determine the initial position of each well-marked target.Meanwhile in the initial bit of each well-marked target
Put after having good positioning, some rotation, regulations of Pan and Zoom are carried out to default object module so that the direction of object module,
Size is probably consistent with the target in image.
Step 32, the good each grader of training in advance is imported, to be identified, classify to each well-marked target.Due to each aobvious
The species for writing target there may be difference, it would therefore be desirable to carry out classification and Detection using different graders.
Step 33, feature extraction is carried out to each well-marked target in image to be detected using the grader.
In the present embodiment, object module of the Star Model of use as grader, the Star Model by one substantially
The high-resolution of small parts in the root wave filter and covering well-marked target of the coarse low resolution of the upper whole well-marked target of covering
The part wave filter of rate is formed.Root wave filter defines detection window (pixel for the feature space part that wave filter is covered).
Where part wave filter is placed on root wave filter under the λ layers of layer, the resolution ratio of this layer of feature is layer where root wave filter
Twice of feature.Utilize high-resolution features, part filtering most important to obtaining high recognition performance come definition component wave filter
Device can be caught relative to the more pinpoint feature of root wave filter.Such as:The object module of face is established, its root wave filter is caught
What is caught is the rough edge information such as face border, and part wave filter can catch the details letters such as eyes in face, nose, mouth
Breath.
In the present embodiment, (n+2) tuple can be defined as in form for the object module containing n part:
(F0,pi,…pn, b), F0It is root wave filter, piIt is the model of i-th of part, b is the real number value for representing deviation.Each part mould
Type is defined with a triple:(Fi,vi,di), FiIt is the part wave filter of i-th of part.viIt is a bivector, specifies the
The anchor point position (normal place when not deforming upon) of i part wave filter is relative to root position (where root wave filter
Position) coordinate;diIt is four dimensional vectors, specifies the parameter of a quadratic function, this quadratic function represents the every of part
Individual possible position is spent relative to the deformation of anchor point position.
Each well-marked target specifies position of each wave filter in feature pyramid H in model:Z=(p0,…,
pn), wherein, pi=(xi,yi,li) represent i-th of wave filter where layer and position coordinates, xi,yi,liI-th of filter is represented respectively
Horizontal coordinate, vertical coordinate and the place layer of ripple device.And p0=(x0,y0,l0) be root wave filter where layer and position coordinates.
The feature resolution of layer where each part is twice of the feature resolution of layer where root wave filter, i.e. liPoint of layer feature
Resolution is l0Twice of the resolution ratio of layer feature, and li=l0- λ (i > 0).
The feature extraction result of well-marked target is score, and the score of each well-marked target is equal to each part wave filter respective
The score of position subtracts each part wave filter position (from the point of view of data) and spent relative to the deformation of the root position of root wave filter
For expense along with deviation (also known as shift term) (from the point of view of space), the calculation formula of the score of well-marked target can be formula 1:
Wherein, score (p0,…,pn) for the score of well-marked target, b is shift term;For i-th
Part wave filter FiScore, Fi' it is that the vectorization of i-th part wave filter represents,For i-th of part wave filter Fi'
Characteristic vector, pi=(xi,yi,li) represent i-th of wave filter where layer liWith horizontal position coordinate xiSat with upright position
Mark yi, H is characterized pyramid, and n represents the quantity of part wave filter.
In formula 1 i-th of part wave filter relative to anchor point position displacementCalculation can be formula 2:
Wherein, (x0,y0) be root wave filter layer where it coordinate, need to multiply to part wave filter place layer in order to unified
With 2.viBe part i anchor point position relative to the coordinate offset of root, so 2 (x0,y0)+viPart i when expression does not deform upon
Absolute coordinate (absolute coordinate of anchor point position).
In formula 1,For the deformation behaviour of i-th of part wave filter, its calculation can be formula 3:
Wherein,WithSquare of horizontal displacement and the horizontal displacement of part wave filter is represented respectively,WithTable respectively
Show square of vertical displacement and the vertical displacement of part wave filter.
Such as:D=(0,0,1,1), then it is exactly its position and anchor point position that the deformation of i-th of part wave filter, which is spent,
Square of distance.Under normal circumstances, deformation cost is any detachable quadratic function of displacement.It is to incite somebody to action to introduce deviation
During multiple model composition mixed models, make the score of multiple models there is comparativity.
In the present embodiment, well-marked target z score can be expressed as the form of dot product:β ψ (H, z), β are model ginsengs
Number vector, its calculation formula can be formula 4:
β=(F0',···Fn',d1,···,dn, b) and formula 4
ψ (H, z) is characteristic vector, and its calculation formula can be formula 5:
Step 34, judge to whether there is well-marked target in image to be detected, well-marked target in image to be detected be present
In the case of, perform step 35.In the case of well-marked target is not present in image to be detected, image to be detected is reduced, is returned again to
Step 31, the image after diminution is pre-processed.
In the present embodiment, the size that diminution obtains image to be detected every time can be image to be detected size before reducing
Half, can carry out twice reduce detection not yet obtain well-marked target in the case of, it is believed that do not deposited in the image to be detected
In well-marked target, terminate the detection to the image to be detected.
Step 35, feature extraction result is obtained, and determines each marking area in image to be detected according to feature extraction result
In optimal location.
When detecting well-marked target in the picture, according to the comprehensive score of each root position of the best position calculation of all parts
score(p0), such as following formula 6:
Wherein,Represent four dimensional vectors.
The root position of high score defines a well-marked target detection, and the component locations for producing high score root position define
One complete well-marked target, optimal position of the well-marked target in image to be detected can be determined according to the root position of high score
Put.
By defining the comprehensive score (overall score) of each root position, we can detect the more of well-marked target
Individual example (assuming that most examples on each root position).This method is relevant with sliding window detector, because can be with
Think score (p0) it is score of the detection window in designated root position.
In the present embodiment, Dynamic Programming (dynamic programming) and generalized distance transformation (min- can be used
Convolution the optimal location (being the function of root position) of calculating unit) is carried out.
IfIt is arrays of the storage means i in the response of feature pyramid l layers.
Matching algorithm can calculate these responses first.Pay attention to Ri,1It is wave filter FiWith the crosscorrelation of feature pyramid l layers.
After having calculated these wave filter responses, it is changed to allow with spatial location laws by formula 7:
Wherein, Di,lIt is obtained to root position when (x, y) value represents for the anchor point of i-th of part to be placed on position (x, y) of l layers
The maximum contribution value divided.This conversion can expand to wave filter high score close position, while also take deformation cost into account,
L takes l0When can calculate
Array D after conversioni,lCan according to generalized distance transformation algorithm in linear session from array Ri,1It is calculated.
The comprehensive score of each layer of root position can be expressed as this layer of root wave filter response and be adopted plus by conversion and son
The response of the part wave filter of sample, calculation can be formula 8,
Wherein, λ is to need the number of plies walked downwards in pyramid to obtain twice of resolution ratio of a certain layer.
It should be noted that for a fixed root position, the optimum position of each part of selection that can be independent, because
Not account for the interaction between part wave filter in the calculating of the score of well-marked target.Array D after conversioni,lRepresent
For i-th of part to the contribution margin of root position Synthesis score, it is the function of part wave filter anchor point position.So by by root
Wave filter response is added with the contribution margin of all parts wave filter, has just obtained the comprehensive score of one root position of l layers, wherein
The contribution margin of each part wave filter is precalculated to be stored in array Di,1-λIn.
In addition, calculating Di,lDuring algorithm can also calculate optimal location (namely the anchor point coordinate of each part
Function), such as formula 9:
Find the root position (x of a high score0,y0,l0) after, Ke YiCorresponding to middle lookup
Part optimal location.
Step 36, according to the optimal location, it is predicted using bounding volume method, determines at least the one of each well-marked target
The score of individual estimation range, the corresponding bounding box in each estimation range and the bounding box.
In the present embodiment, using the complete configuration of well-marked target, z=(p0,…,pn), to predict the bag of each well-marked target
Enclose box.This is that characteristic vector g (z) is mapped as into bounding box upper left angle point (x by one1,y1) and bottom right angle point (x2,y2)
Function is realized.For a model containing n part, g (z) is the vector of a 2n+3 dimension, comprising in units of pixel
Root filter width (pointing out dimensional information) and each wave filter (including root wave filter and part wave filter) in image upper left
The position coordinates of angle point.Because the positioning precision of part wave filter is greater than root wave filter, multiple dimensioned deformable component mould is used
Type can obtain potential valuable information.
Step 37, the estimation range repeated is removed using non-maxima suppression method, is determined aobvious where each well-marked target
Write region.
Matching process in step 36 often obtains multiple overlapping detections of each well-marked target example, i.e., each aobvious
Corresponding multiple bounding boxs can be obtained by writing target.With a greedy non-maxima suppression algorithm (Non-maximum
Suppression, NMS) program eliminates repetition detection, obtain in multiple bounding boxs corresponding to each well-marked target score most
High bounding box, and by the marking area where the bounding box of highest scoring well-marked target the most.Non-maxima suppression algorithm
Essence be search local maximum, suppress non-maximum element.
In the present embodiment, the bounding box prediction algorithm introduced using step 36, can obtain the notable mesh of certain class in image
One estimation range set D of target.The corresponding bounding box in each estimation range and a score in D.By score to the inspection in D
Sort result is surveyed, avidly selection has the estimation range of top score and skips the bounding box of the estimation range selected before
The result more than 50% is covered, to obtain the marking area of each well-marked target highest scoring.
Step 38, the marking area of each well-marked target is modified using contextual information.
With contextual information (Contextual Information) to the marking area where each well-marked target of acquisition
Carry out the simple program of scoring (secondary scoring) again.If (D1,···Dk) it is to be schemed with the model of k different target classification
As the marking area obtained in I.Each marking area (B, s) ∈ DiBy a bounding box B=(x1,y1,x2,y2) and a score
S is defined.We are with k dimensional vectors c (I)=(σ (s1),···,σ(sk)) define image I context, wherein siIt is
DiThe fraction of middle top score marking area,It is a logic letter that renormalization is carried out to fraction
Number.
In order to be scored again the marking area (B, s) in image I, original detection score, a left side for bounding box are utilized
Upper angle and bottom right angular coordinate and image context construct one 25 dimension (c (I) is 20 dimensions) characteristic vector g, such as following formula 10:
G=(σ (s), x1,y1,x2,y2, c (I)) and formula 10
Coordinate x1,y1,x2,y2∈ [0,1] is normalized with the width and height of image.With a special grader to this
New characteristic vector is scored, and obtains the new score of marking area.The contextual information that this grader combination g is defined will be correct
Marking area is distinguished and exported from wrong report marking area.
It should be noted that although a kind of target detection based on image and classification are described using embodiment 2 as example
Method as above, it is understood by one of ordinary skill in the art that the disclosure answers not limited to this.In fact, user completely can be according to individual
Hobby and/or practical application scene flexibly set the specific implementation of each step, as long as meeting the think of of disclosed technique scheme
Road.
The target detection and sorting technique based on image that the disclosure is provided, utilize the target handled based on threshold binary image
Measurement, has filtered and has been free of target area, accelerated detection process.Solve the translation because of target object illumination, sized make
The problem of complicated change such as shade of the color value of pixel and target object all increases the difficulty of target detection into image.And
And this method can effectively can be applied in Practical Project with real time execution.
The disclosure can be system, method and/or computer program product.Computer program product can include computer
Readable storage medium storing program for executing, containing for making processor realize the computer-readable program instructions of various aspects of the disclosure.
Computer-readable recording medium can keep and store to perform the tangible of the instruction that uses of equipment by instruction
Equipment.Computer-readable recording medium for example can be-- but be not limited to-- storage device electric, magnetic storage apparatus, optical storage
Equipment, electromagnetism storage device, semiconductor memory apparatus or above-mentioned any appropriate combination.Computer-readable recording medium
More specifically example (non exhaustive list) includes:Portable computer diskette, hard disk, random access memory (RAM), read-only deposit
It is reservoir (ROM), erasable programmable read only memory (EPROM or flash memory), static RAM (SRAM), portable
Compact disk read-only storage (CD-ROM), digital versatile disc (DVD), memory stick, floppy disk, mechanical coding equipment, for example thereon
It is stored with punch card or groove internal projection structure and the above-mentioned any appropriate combination of instruction.Calculating used herein above
Machine readable storage medium storing program for executing is not construed as instantaneous signal in itself, the electromagnetic wave of such as radio wave or other Free propagations, leads to
Cross the electromagnetic wave (for example, the light pulse for passing through fiber optic cables) of waveguide or the propagation of other transmission mediums or transmitted by electric wire
Electric signal.
Computer-readable program instructions as described herein can be downloaded to from computer-readable recording medium it is each calculate/
Processing equipment, or outer computer or outer is downloaded to by network, such as internet, LAN, wide area network and/or wireless network
Portion's storage device.Network can include copper transmission cable, optical fiber is transmitted, is wirelessly transferred, router, fire wall, interchanger, gateway
Computer and/or Edge Server.Adapter or network interface in each calculating/processing equipment receive from network to be counted
Calculation machine readable program instructions, and the computer-readable program instructions are forwarded, for the meter being stored in each calculating/processing equipment
In calculation machine readable storage medium storing program for executing.
For perform the disclosure operation computer program instructions can be assembly instruction, instruction set architecture (ISA) instruction,
Machine instruction, machine-dependent instructions, microcode, firmware instructions, condition setup data or with one or more programming languages
The source code or object code that any combination is write, programming language of the programming language including object-oriented-such as
Smalltalk, C++ etc., and conventional procedural programming languages-such as " C " language or similar programming language.Computer
Readable program instructions fully can on the user computer perform, partly perform on the user computer, be only as one
Vertical software kit performs, part performs or completely in remote computer on the remote computer on the user computer for part
Or performed on server.In the situation of remote computer is related to, remote computer can pass through network-bag of any kind
LAN (LAN) or wide area network (WAN)-be connected to subscriber computer are included, or, it may be connected to outer computer (such as profit
Pass through Internet connection with ISP).In certain embodiments, by using computer-readable program instructions
Status information carry out personalized customization electronic circuit, such as PLD, field programmable gate array (FPGA) or can
Programmed logic array (PLA) (PLA), the electronic circuit can perform computer-readable program instructions, so as to realize each side of the disclosure
Face.
Referring herein to the method, apparatus (system) according to the embodiment of the present disclosure and the flow chart of computer program product and/
Or block diagram describes various aspects of the disclosure.It should be appreciated that each square frame and flow chart of flow chart and/or block diagram and/
Or in block diagram each square frame combination, can be realized by computer-readable program instructions.
These computer-readable program instructions can be supplied to all-purpose computer, special-purpose computer or other programmable datas
The processor of processing unit, so as to produce a kind of machine so that these instructions are passing through computer or other programmable datas
During the computing device of processing unit, work(specified in one or more of implementation process figure and/or block diagram square frame is generated
The device of energy/action.These computer-readable program instructions can also be stored in a computer-readable storage medium, these refer to
Order causes computer, programmable data processing unit and/or other equipment to work in a specific way, so as to be stored with instruction
Computer-readable medium then includes a manufacture, and it is included in one or more of implementation process figure and/or block diagram square frame
The instruction of the various aspects of defined function/action.
Computer-readable program instructions can also be loaded into computer, other programmable data processing units or other
In equipment so that series of operation steps is performed on computer, other programmable data processing units or miscellaneous equipment, with production
Raw computer implemented process, so that performed on computer, other programmable data processing units or miscellaneous equipment
Instruct function/action specified in one or more of implementation process figure and/or block diagram square frame.
Flow chart and block diagram in accompanying drawing show the system, method and computer journey of multiple embodiments according to the disclosure
Architectural framework in the cards, function and the operation of sequence product.At this point, each square frame in flow chart or block diagram can generation
One module of table, program segment or a part for instruction, the module, program segment or a part for instruction include one or more use
In the executable instruction of logic function as defined in realization.At some as the function of in the realization replaced, being marked in square frame
Can be with different from the order marked in accompanying drawing generation.For example, two continuous square frames can essentially be held substantially in parallel
OK, they can also be performed in the opposite order sometimes, and this is depending on involved function.It is also noted that block diagram and/or
The combination of each square frame and block diagram in flow chart and/or the square frame in flow chart, function or dynamic as defined in performing can be used
The special hardware based system made is realized, or can be realized with the combination of specialized hardware and computer instruction.
It is described above the presently disclosed embodiments, described above is exemplary, and non-exclusive, and
It is not limited to disclosed each embodiment.In the case of without departing from the scope and spirit of illustrated each embodiment, for this skill
Many modifications and changes will be apparent from for the those of ordinary skill in art field.The selection of term used herein, purport
The principle of each embodiment, practical application or technological improvement to the technology in market are best being explained, or is leading this technology
Other those of ordinary skill in domain are understood that each embodiment disclosed herein.
Claims (9)
1. a kind of target detection and sorting technique based on image, it is characterised in that including:
Image to be detected is pre-processed;
Import the grader of training in advance;
Feature extraction is carried out to each well-marked target in described image to be detected using the grader;
Optimal location of each well-marked target in described image to be detected is determined according to feature extraction result;
Marking area according to where each optimal location determines each well-marked target.
2. according to the method for claim 1, it is characterised in that described image to be detected is pre-processed, including:
Redundancy removal processing and/or noise removal process are carried out to image to be detected of acquisition;
Each well-marked target in described image to be detected is obtained, and determines the initial position of each well-marked target.
3. method according to claim 1 or 2, it is characterised in that the grader includes root wave filter and multiple parts
Wave filter, it is described that feature extraction is carried out to each well-marked target in described image to be detected using the grader, including:
The root position of each well-marked target is determined using described wave filter;
Using each part wave filter, the position of each part in each well-marked target is determined;
According to the score of each part wave filter, each part wave filter position relative to described position deformation
Cost and shift term determine the score of each well-marked target.
4. according to the method for claim 3, it is characterised in that according to the score of each part wave filter, each portion
Part wave filter position is spent relative to the deformation of described position and shift term determines the score of each well-marked target,
Including:
The score of the well-marked target is calculated using formula 1:
Wherein, score (p0,…,pn) for the score of well-marked target, b is shift term;Filtered for i-th of part
Ripple device FiScore, Fi' it is that the vectorization of i-th part wave filter represents,For i-th of part wave filter Fi' spy
Sign vector, pi=(xi,yi,li) represent i-th of wave filter where layer liWith horizontal position coordinate xiWith vertical position coordinate yi,
H is characterized pyramid, and n represents the quantity of part wave filter;
Spent for i-th of part wave filter position relative to the deformation of described position;diI-th
Part wave filter position is spent relative to the deformation of itself anchor point position,
I-th of part wave filter relative to anchor point position displacementCalculated using formula 2:
The deformation behaviour of i-th of part wave filterUse formula 3 calculate:
Wherein, (x0,y0) be root wave filter layer where it coordinate, viIt is a bivector, represents i-th of part filtering
The anchor point position of device relative to root position coordinate,WithThe horizontal displacement and horizontal displacement of part wave filter are represented respectively
Square,WithSquare of vertical displacement and the vertical displacement of part wave filter is represented respectively.
5. method according to any one of claim 1 to 4, it is characterised in that each institute is determined according to feature extraction result
Optimal location of the well-marked target in described image to be detected is stated, including:
Obtain the optimal location for each part wave filter for detecting the well-marked target;
The comprehensive of the root position of the root wave filter of the well-marked target is detected according to the best position calculation of each part wave filter
Close score;
Determine described wave filter in the well-marked target in described image to be detected according to the comprehensive score of described wave filter
In optimal location.
6. according to the method for claim 4, it is characterised in that the comprehensive score according to described wave filter determines institute
Optimal location of the root wave filter in the well-marked target in described image to be detected is stated, including:
The root position of each described wave filter is calculated in l using formula 80Comprehensive score score (the x of layer0,y0,l0), in the synthesis
Score score (x0,y0,l0) in the case of highest, using the place root position of described wave filter of acquisition as described filtering
The optimal location of device:
Wherein, l0For the layer where described wave filter, x0, y0It is described wave filter in l0The position coordinates of layer, λ are to obtain
l0Layer twice of resolution ratio and need the number of plies walked downwards in feature pyramid H,Wherein,
L takes l0When calculate
Calculated by formula 7:
Wherein, Di,lWhen (x, y) value represents for the anchor point position of i-th of part wave filter to be placed on position (x, y) of l layers, i-th
Part wave filter takes l to the maximum contribution value of root position score, l0When calculate
7. according to the method for claim 6, it is characterised in that the comprehensive score according to described wave filter determines institute
Optimal location of the root wave filter in the well-marked target in described image to be detected is stated, in addition to:
In the case where described wave filter is in the optimal location, is calculated using formula 9 and obtain each part wave filter
Optimal location:
Wherein, the pi,l(x, y) is the position function of each wave filter.
8. according to the method for claim 1, it is characterised in that it is described according to the optimal location determine and export it is each described in
Marking area where well-marked target;Including:
According to the optimal location, using bounding box Forecasting Methodology, each estimation range where each well-marked target is determined;
The estimation range repeated is removed using non-maxima suppression method, determines the marking area where each well-marked target;
Marking area where according to the contextual information amendment of each well-marked target and exporting each well-marked target.
9. according to the method for claim 1, it is characterised in that methods described also includes:
In the case of well-marked target is not present in described image to be detected, reduce described image to be detected, and to diminution after
Image to be detected is pre-processed.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710617028.3A CN107368832A (en) | 2017-07-26 | 2017-07-26 | Target detection and sorting technique based on image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710617028.3A CN107368832A (en) | 2017-07-26 | 2017-07-26 | Target detection and sorting technique based on image |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107368832A true CN107368832A (en) | 2017-11-21 |
Family
ID=60307205
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710617028.3A Pending CN107368832A (en) | 2017-07-26 | 2017-07-26 | Target detection and sorting technique based on image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107368832A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109145947A (en) * | 2018-07-17 | 2019-01-04 | 昆明理工大学 | A kind of Fashionable women dress image fine grit classification method based on component detection and visual signature |
CN110490196A (en) * | 2019-08-09 | 2019-11-22 | Oppo广东移动通信有限公司 | Subject detection method and apparatus, electronic equipment, computer readable storage medium |
CN111656356A (en) * | 2018-03-16 | 2020-09-11 | 罗克韦尔柯林斯德国公司 | Object recognition system using distributed neural network |
CN111680698A (en) * | 2020-04-21 | 2020-09-18 | 北京三快在线科技有限公司 | Image recognition method and device and training method and device of image recognition model |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104200236A (en) * | 2014-08-22 | 2014-12-10 | 浙江生辉照明有限公司 | Quick target detection method based on DPM (deformable part model) |
CN104636749A (en) * | 2013-11-14 | 2015-05-20 | 中国移动通信集团公司 | Target object detection method and device |
US9042601B2 (en) * | 2013-03-14 | 2015-05-26 | Nec Laboratories America, Inc. | Selective max-pooling for object detection |
CN204946028U (en) * | 2015-06-12 | 2016-01-06 | 北京富兴隆科技有限公司 | Object detecting device |
CN105956632A (en) * | 2016-05-20 | 2016-09-21 | 浙江宇视科技有限公司 | Target detection method and device |
CN106067022A (en) * | 2016-05-28 | 2016-11-02 | 北方工业大学 | Remote sensing image harbor ship detection false alarm eliminating method based on DPM algorithm |
-
2017
- 2017-07-26 CN CN201710617028.3A patent/CN107368832A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9042601B2 (en) * | 2013-03-14 | 2015-05-26 | Nec Laboratories America, Inc. | Selective max-pooling for object detection |
CN104636749A (en) * | 2013-11-14 | 2015-05-20 | 中国移动通信集团公司 | Target object detection method and device |
CN104200236A (en) * | 2014-08-22 | 2014-12-10 | 浙江生辉照明有限公司 | Quick target detection method based on DPM (deformable part model) |
CN204946028U (en) * | 2015-06-12 | 2016-01-06 | 北京富兴隆科技有限公司 | Object detecting device |
CN105956632A (en) * | 2016-05-20 | 2016-09-21 | 浙江宇视科技有限公司 | Target detection method and device |
CN106067022A (en) * | 2016-05-28 | 2016-11-02 | 北方工业大学 | Remote sensing image harbor ship detection false alarm eliminating method based on DPM algorithm |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111656356A (en) * | 2018-03-16 | 2020-09-11 | 罗克韦尔柯林斯德国公司 | Object recognition system using distributed neural network |
CN111656356B (en) * | 2018-03-16 | 2023-08-08 | 罗克韦尔柯林斯德国公司 | Object recognition system using distributed neural network |
US11816567B2 (en) | 2018-03-16 | 2023-11-14 | Rockwell Collins Deutschland Gmbh | System for identifying objects by means of distributed neural networks |
CN109145947A (en) * | 2018-07-17 | 2019-01-04 | 昆明理工大学 | A kind of Fashionable women dress image fine grit classification method based on component detection and visual signature |
CN109145947B (en) * | 2018-07-17 | 2022-04-12 | 昆明理工大学 | Fashion women's dress image fine-grained classification method based on part detection and visual features |
CN110490196A (en) * | 2019-08-09 | 2019-11-22 | Oppo广东移动通信有限公司 | Subject detection method and apparatus, electronic equipment, computer readable storage medium |
CN111680698A (en) * | 2020-04-21 | 2020-09-18 | 北京三快在线科技有限公司 | Image recognition method and device and training method and device of image recognition model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110059694B (en) | Intelligent identification method for character data in complex scene of power industry | |
CN112348815B (en) | Image processing method, image processing apparatus, and non-transitory storage medium | |
US10346720B2 (en) | Rotation variant object detection in Deep Learning | |
Islam et al. | Label refinement network for coarse-to-fine semantic segmentation | |
CN110287960A (en) | The detection recognition method of curve text in natural scene image | |
CN101406390B (en) | Method and apparatus for detecting part of human body and human, and method and apparatus for detecting objects | |
CN107368832A (en) | Target detection and sorting technique based on image | |
CN107624061A (en) | Machine vision with dimension data reduction | |
CN110929593A (en) | Real-time significance pedestrian detection method based on detail distinguishing and distinguishing | |
CN110287798B (en) | Vector network pedestrian detection method based on feature modularization and context fusion | |
CN110517270B (en) | Indoor scene semantic segmentation method based on super-pixel depth network | |
CN112242002B (en) | Object identification and panoramic roaming method based on deep learning | |
CN109360179A (en) | A kind of image interfusion method, device and readable storage medium storing program for executing | |
CN112700529A (en) | Method and system for generating three-dimensional model according to standard document | |
CN111507353B (en) | Chinese field detection method and system based on character recognition | |
CN115082941A (en) | Form information acquisition method and device for form document image | |
Mi et al. | Research on a fast human-detection algorithm for unmanned surveillance area in bulk ports | |
CN109740527B (en) | Image processing method in video frame | |
Qiu et al. | CrossDet++: Growing crossline representation for object detection | |
CN114494786A (en) | Fine-grained image classification method based on multilayer coordination convolutional neural network | |
Melzer et al. | Stroke detection of brush strokes in portrait miniatures using a semi-parametric and a model based approach | |
CN111767919A (en) | Target detection method for multi-layer bidirectional feature extraction and fusion | |
CN110781851A (en) | Method for identifying decoration house type graph based on picture | |
CN114972335A (en) | Image classification method and device for industrial detection and computer equipment | |
Varkentin et al. | Development of an application for car license plates recognition using neural network technologies |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171121 |
|
RJ01 | Rejection of invention patent application after publication |