CN109697464A - Method and system based on the identification of the precision target of object detection and signature search - Google Patents
Method and system based on the identification of the precision target of object detection and signature search Download PDFInfo
- Publication number
- CN109697464A CN109697464A CN201811540379.XA CN201811540379A CN109697464A CN 109697464 A CN109697464 A CN 109697464A CN 201811540379 A CN201811540379 A CN 201811540379A CN 109697464 A CN109697464 A CN 109697464A
- Authority
- CN
- China
- Prior art keywords
- feature
- group
- prediction block
- input picture
- original input
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4038—Scaling the whole image or part thereof for image mosaicing, i.e. plane images composed of plane sub-images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Abstract
The present invention provides a kind of method and system based on the identification of the precision target of object detection and signature search, wherein this method comprises: obtaining original input picture, obtains fisrt feature figure group by CNN convolutional neural networks；Resampling and concatenation are successively carried out to the fisrt feature figure group, obtain second feature figure group；The second feature figure group is mapped in the original input picture, the position detection of prediction block is carried out to the original input picture, obtains the prediction block of the original input picture；Judge whether the prediction block is target area, if so, obtaining the classification information in the target area；Extract the visual signature of the target area；The detailed classification information of visual signature and original input picture is matched.The method that the embodiment of the present invention is combined using object detection and signature search, realizes accurately identifying for target, can quickly add fresh target, be with good expansibility.
Description
Technical field
The present invention relates to target detection identification technology fields, are searched in particular to one kind based on object detection and feature
The method and system of the precision target identification of rope.
Background technique
In the prior art, the detection technique based on deep learning is used merely, can only identify the classification information of target, it can not
Accurately identifying target is some specific target.If in order to identify that target, can only be by target point to some specific target
It is more more careful that class is got, but arithmetic accuracy and speed can decreased significantly, and the data volume that target classification needs is huge, cost
It is high.Addition to fresh target can only add classification, and the detection identification model for needing new data training new, expanded period
It is long, scalability.
Summary of the invention
In view of this, the purpose of the present invention is to provide the sides of the precision target identification based on object detection and signature search
Method and system accomplish quickly to add new target, be with good expansibility, it can be achieved that target accurately identifies,.
In a first aspect, the embodiment of the invention provides a kind of based on the identification of the precision target of object detection and signature search
Method, wherein include:
Original input picture is obtained, obtains fisrt feature figure group by CNN convolutional neural networks；
Resampling and concatenation are successively carried out to the fisrt feature figure group, obtain second feature figure group；
The second feature figure group is mapped in the original input picture, the original input picture is predicted
The position detection of frame obtains the prediction block of the original input picture；
Judge whether the prediction block is target area, if so, obtaining the classification information in the target area；
Extract the visual signature of the target area；
The detailed classification information of visual signature and original input picture is matched.
With reference to first aspect, the embodiment of the invention provides the first possible embodiments of first aspect, wherein institute
Stating resampling includes up-sampling and down-sampling, successively carries out resampling and concatenation to the fisrt feature figure group, obtains the
Two characteristic pattern groups include:
Down-sampling is carried out to the fisrt feature figure group, obtains third feature figure group；
Characteristic pattern except out to out characteristic pattern in the third feature figure group is up-sampled, fourth feature is obtained
Figure group；
Characteristic pattern in the third feature figure group and the fourth feature figure group with same scale is spelled respectively
It connects, obtains the second feature figure group.
With reference to first aspect, the embodiment of the invention provides second of possible embodiments of first aspect, wherein institute
It states and the second feature figure group is mapped in the original input picture, prediction block position is carried out to the original input picture
Detection, the prediction block for obtaining the original input picture include:
Scale based on second feature figure group grid division on the original input picture；
Each characteristic pattern in the second feature figure group is respectively mapped in the grid of the original input picture；
A prediction block is at least generated on the grid；
The correction position that the prediction block is calculated based on bounding box regression model obtains the correction value of the prediction block；
Correction value based on the prediction block and the prediction block, obtains the center of the prediction block of the original input picture
Position.
With reference to first aspect, the embodiment of the invention provides the third possible embodiments of first aspect, wherein institute
It states and judges whether the prediction block is that target area includes:
Destination probability is generated in the prediction block；Judge the destination probability whether less than the first preset threshold, if so,
Then the prediction block is not the target area；
If it is not, then the prediction block is the target area.
With reference to first aspect, the embodiment of the invention provides the 4th kind of possible embodiments of first aspect, wherein institute
It states if so, the classification information obtained in the target area includes:
Generate the class probability of the target；
The destination probability is multiplied with the class probability, formation condition class probability；
The condition class probability maximum value is chosen, it will classification information corresponding with the condition class probability maximum value
It is input in database.
With reference to first aspect, the embodiment of the invention provides the 5th kind of possible embodiments of first aspect, wherein institute
It states and extracts the visual signature of the object detection area and include:
It will determine that the original input picture behind target area is input to Visual Feature Retrieval Process model as query image
In；
Extract the visual signature of the query image.
Second aspect is identified the embodiment of the invention also provides a kind of based on the precision target of object detection and signature search
System, wherein include:
Module of target detection obtains original input picture, obtains fisrt feature figure group by CNN convolutional neural networks；It is right
The fisrt feature figure group successively carries out resampling and concatenation, obtains second feature figure group；By the second feature figure group
It is mapped in the original input picture, the position detection of prediction block is carried out to the original input picture, is obtained described original
The prediction block of input picture；Judge whether the prediction block is target area, if so, obtaining the classification in the target area
Information；
Target's feature-extraction module, for extracting the visual signature of the target area；
Target signature search module, for the detailed classification information of visual signature and original input picture to match.
In conjunction with second aspect, the embodiment of the invention provides the first possible embodiments of second aspect, wherein institute
Stating resampling includes that up-sampling and down-sampling, the module of target detection are specifically used for:
Down-sampling is carried out to the fisrt feature figure group, obtains third feature figure group；
Characteristic pattern except out to out characteristic pattern in the third feature figure group is up-sampled, fourth feature is obtained
Figure group；
Characteristic pattern in the third feature figure group and the fourth feature figure with same scale is spliced respectively,
Obtain the second feature figure group.
In conjunction with second aspect, the embodiment of the invention provides second of possible embodiments of second aspect, wherein institute
State module of target detection also particularly useful for:
Scale based on second feature figure group grid division on the original input picture；
Each characteristic pattern in the second feature figure group is respectively mapped in the grid of the original input picture；
A prediction block is at least generated on the grid；
The correction position that the prediction block is calculated based on bounding box regression model obtains the correction value of the prediction block；
Correction value based on the prediction block and the prediction block, obtains the center of the prediction block of the original input picture
Position.
In conjunction with second aspect, the embodiment of the invention provides the third possible embodiments of second aspect, wherein institute
Target's feature-extraction module is stated to be specifically used for:
It will determine that the original input picture behind target area is input to Visual Feature Retrieval Process model as query image
In；
Extract the visual signature of the query image.
Method and system provided in an embodiment of the present invention based on the identification of the precision target of object detection and signature search, are adopted
The method combined with object detection and signature search, with detection technique of the simple use based on deep learning in the prior art
It compares, can not only identify the classification information of target, can also realize accurately identifying for target, can quickly add fresh target, have
There is good scalability.
To enable the above objects, features and advantages of the present invention to be clearer and more comprehensible, preferred embodiment is cited below particularly, and cooperate
Appended attached drawing, is described in detail below.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached
Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair
The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this
A little attached drawings obtain other relevant attached drawings.
Fig. 1 shows the side based on the identification of the precision target of object detection and signature search provided by the embodiment of the present invention
The flow chart of method；
Fig. 2 shows C.ReLU structural schematic diagrams provided by the embodiment of the present invention；
Fig. 3 shows fireModle+Inception structural schematic diagram provided by the embodiment of the present invention；
Fig. 4 shows the flow chart of resampling provided by the embodiment of the present invention；
Fig. 5 shows the flow chart of the position detection of prediction block provided by the embodiment of the present invention；
Fig. 6 show judge provided by the embodiment of the present invention prediction block whether be target area flow chart；
Fig. 7 shows the flow chart that the classification information in target area is obtained provided by the embodiment of the present invention；
Fig. 8 shows the flow chart that the visual signature of target area is extracted provided by the embodiment of the present invention；
Fig. 9 shows the structural representation of the system based on object detection and signature search provided by the embodiment of the present invention
Figure.
Main element symbol description: 1- module of target detection；2- target's feature-extraction module；3- target signature searches for mould
Block.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
Middle attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is only
It is a part of the embodiment of the present invention, instead of all the embodiments.The present invention being usually described and illustrated herein in the accompanying drawings is real
The component for applying example can be arranged and be designed with a variety of different configurations.Therefore, of the invention to what is provided in the accompanying drawings below
The detailed description of embodiment is not intended to limit the range of claimed invention, but is merely representative of selected reality of the invention
Apply example.Based on the embodiment of the present invention, those skilled in the art institute obtained without making creative work
There are other embodiments, shall fall within the protection scope of the present invention.
In view of being believed in the prior art using the classification that the detection technique based on deep learning can only recognize target merely
Breath, be based on this, the embodiment of the invention provides it is a kind of based on the precision target of object detection and signature search know method for distinguishing and
System is described below by embodiment.
Method for distinguishing is known based on the precision target of object detection and signature search the embodiment of the invention provides a kind of, is such as schemed
Shown in 1, following steps S101~S106 is specifically included:
S101 obtains original input picture, obtains fisrt feature figure group by CNN convolutional neural networks.
Convolutional neural networks are composed in series by the sub-network that multiple C.ReLU and fireModule+Inception are combined,
Character extraction is carried out to the original image of input, provides accurately target signature description for subsequent target detection.
As shown in Fig. 2, the calculation amount of neural network rises appreciably with the growth of detection classification number, detection speed is had
Certain decline occurs in pairs using low layer convolution kernel in C.ReLU structure in order to raise speed and guarantee detection accuracy and parameter is mutual
For the characteristic of opposite number, by simply connecting identical output and it being made to become double, in the number for reducing half output channel
It under amount, can still reach the quantity exported originally, both improve 2 times of arithmetic speed, in turn ensure the precision of detection network.And
And in order to provide the Nonlinear Modeling ability of neural network, activation primitive Relu joined in C.ReLU structure.
As shown in figure 3, including the convolutional layer that a convolution kernel size is 1 × 1 in fireModule structure, reach dimensionality reduction effect
Fruit guarantees the characteristic pattern number being input in Inception structure so that the characteristic pattern quantity in input feature vector figure group greatly reduces
Amount must be lacked as far as possible, reduce the calculation amount in Inception structure, therefore, fireModule structure plays the work of model compression
With realization reduces model parameter and improves the purpose of speed.It is mainly of different sizes comprising three convolution kernels in Inception structure
Convolutional layer, be 1 × 1 convolution kernel, 3 × 3 convolution kernels and two 3 × 3 convolution kernels respectively.In general, 1 can be chosen in network
× 1 convolution kernel, 3 × 3 convolution kernels and 5 × 5 convolution kernels, but in order to guarantee to promote network under conditions of with same perceived open country
Depth, and calculating parameter amount is reduced, 5 × 5 convolution kernels in embodiments of the present invention are by two 3 × 3 convolution nuclear subsitutions.
The last layer in Inception structure is 1 × 1 convolution kernel, has not only remained one layer of receptive field, but also increase input pattern
It is non-linear, slowed down it is some output features receptive fields growth, can accurately capture the target of small size.
Original input picture obtains output characteristic pattern group by multiple concatenated C.ReLU structures.Upper one layer of output is special
Figure group is levied as fireModule layers of input, by the input feature vector figure group that is obtained after 1 × 1 convolution kernel dimensionality reduction respectively with three
Different size of convolution nuclear convolution obtains the different output characteristic pattern group of three receptive fields, wherein uses 1 × 1 convolution nuclear convolution
The output characteristic pattern group obtained afterwards corresponds to sufficiently small receptive field, can accurately detect small target；Use 3 × 3 convolution kernels
And two 3 × 3 convolution kernels, the output characteristic pattern group obtained after convolution corresponds to big receptive field, to detect bigger mesh
Mark.Then these three characteristic pattern groups are spliced, the output characteristic pattern group of splicing is passed through into 1 × 1 convolution kernel, reach different scale spy
Levy the purpose of fusion.The sub-network series connection that multiple fireModule+Inception modules combine, mentions by the feature of multilayer
After taking, a good feature can be just obtained, as fisrt feature figure group.
S102 successively carries out resampling and concatenation to fisrt feature figure group, obtains second feature figure group.
Optionally, resampling includes up-sampling and down-sampling, as shown in figure 4, specifically including following steps in step S102
S201~S203:
S201 carries out down-sampling to fisrt feature figure group, obtains third feature figure group.
Down-sampling is carried out to fisrt feature figure group, obtains the characteristic pattern group of 40 × 40 scales；Continue to 40 × 40 scales
Characteristic pattern group carries out down-sampling, obtains the characteristic pattern group of 20 × 20 scales；Then the characteristic pattern group of 20 × 20 scales is carried out down
Sampling, obtains the characteristic pattern group of 10 × 10 scales, collectively constitutes third feature figure group by the characteristic pattern group of three kinds of scales.
S202 up-samples the characteristic pattern except out to out characteristic pattern in third feature figure group, obtains the 4th spy
Levy figure group.
The characteristic pattern group of 10 × 10 scales in third feature figure group is up-sampled, the characteristic pattern of 20 × 20 scales is obtained
Group；Then the characteristic pattern group of 20 × 20 scales is up-sampled, obtains the characteristic pattern group of 40 × 40 scales；It is obtained by up-sampling
The characteristic pattern of 20 × 20 scales and the characteristic pattern group of 40 × 40 scales collectively constitute fourth feature figure group.
S203 splices the characteristic pattern in third feature figure group and fourth feature figure group with same scale respectively,
Obtain second feature figure group.
By 20 × 20 scales in the characteristic pattern group of 20 × 20 scales in third feature figure group and fourth feature figure group
Characteristic pattern group is spliced, and the characteristic pattern group of 20 × 20 more scales of a characteristic pattern is obtained；By third feature figure group
In the characteristic pattern group of 40 × 40 scales and the characteristic pattern group of 40 × 40 scales in fourth feature figure group spliced, obtain one
The characteristic pattern group of 40 × 40 more scales of a characteristic pattern.Therefore, the characteristic pattern group of 10 × 10 scales, splicing obtain 20
The characteristic pattern group of × 20 scales and the characteristic pattern group of 40 × 40 scales together constitute multiple dimensioned second feature figure group.
Second feature figure group is mapped in original input picture by S103, and the position of prediction block is carried out to original input picture
Detection is set, the prediction block of original input picture is obtained.
Optionally, as shown in figure 5, specifically including following steps S301~S305 in step S103:
S301, the grid division on original input picture of the scale based on second feature figure group.
Original input picture comes out image characteristics extraction by convolutional layer, with going deep into for network, extracted image
Feature has more and more abstract semantic meaning, while the position meaning of these characteristics of image is but increasingly fuzzyyer, by will be special
Each pixel in sign figure group is mapped in original input picture, embodies its position meaning.
Based on the different scale of second feature figure group, the grid division on original input picture.That is, 10 × 10 rulers
The characteristic pattern group of degree can divide 10 × 10 grids on original input picture, and the characteristic pattern of 20 × 20 scales can be originally inputted
20 × 20 grids are divided on image, the characteristic pattern of 40 × 40 scales can divide 40 × 40 grids on original input picture,
Detection that big target is carried out in the characteristic pattern group of 10 × 10 scales, in the characteristic pattern group and 40 × 40 scales of 20 × 20 scales
Characteristic pattern group successively detects lesser target, meets that big target is few in this way and Small object can more actual conditions.
Each characteristic pattern in second feature figure group is respectively mapped in the grid of original input picture by S302.
By taking the characteristic pattern of 10 × 10 scales as an example, original input picture is divided into 10 × 10 grid, in characteristic pattern
Each feature and the small center of a lattice point of each of grid correspond, each of characteristic pattern feature just describes
In original input picture centered on corresponding to net region, it is greater than the feature in the region of the grid.
S303 at least generates a prediction block on grid.
K prediction block is generated on each grid by trained target detection model, each prediction block is one
A preset 4 dimensional vector (xi, yj, wk, hk), wherein (xi, yj) indicate grid element center point coordinate (i, j), (wk, hk) respectively indicate
The width and height of prediction block, and k=[0, K), indicate preset K prediction block.
S304 is calculated the correction position of prediction block based on bounding box regression model, obtains the correction value of prediction block.
The center of the not necessarily true frame in the center of prediction block is calculated pre- according to formula (1)~(4)
Survey correction value (the δ x of frameijk, δ yijk, δ wijk, δ hijk), respectively indicate the centre coordinate (x of default framei, yj) and prediction block width,
Height (wk, hk) correction value:
S305, the correction value based on prediction block and prediction block, obtains the center of the true frame of original input picture.
By the calculating parameter between training correction value and preset value, by the target's center position of prediction block to true frame
Center is approached, and the center of true frame is finally obtained, and the center of prediction block is essentially identical to true frame at this time
Center.
S104 judges whether prediction block is target area, if so, obtaining the classification information in target area.
Optionally, as shown in fig. 6, specifically including following steps S401~S403 in step S104:
S401 generates destination probability in prediction block.
A numerical value is generated in prediction block, range indicates the probability for having target in prediction block in [0,1].
S402 judges destination probability whether less than the first preset threshold, if so, prediction block is not target area.
First preset threshold is set, if the destination probability of the prediction block is lower than threshold value, indicates that the prediction block is a nothing
The target frame of effect, the prediction block are not the target areas of original input picture.
S403, if it is not, then prediction block is target area.
Optionally, as shown in fig. 7, specifically including following steps S501~S503 in step S403:
S501 generates the class probability of target.
Under the premise of prediction block is that frame is effectively predicted, pass through the deep learning neural network in target detection identification model
Calculate the class probability that the target belongs to each classification.There are 1000 major class in the embodiment of the present invention, by being calculated 1000
A class probability.
S502, destination probability are multiplied with class probability, formation condition class probability.
Destination probability is multiplied with 1000 class probabilities respectively, obtains 1000 condition class probabilities.
S503 chooses condition class probability maximum value, will classification letter corresponding with the condition class probability maximum value
Breath is input in database.
1000 condition class probabilities are arranged from high to low, choose condition class probability maximum value, it will be with condition classification
The corresponding classification information of maximum probability value is input in database.
S105 extracts the visual signature of target area.
Optionally, as shown in figure 8, specifically including following steps S601~S602 in step S105:
S601 will determine that the original input picture behind target area is input to Visual Feature Retrieval Process model as query image
In.
Based on CDVS (Compact Descriptors for VisualSearch, the compact description towards visual search
Son) technology building Visual Feature Retrieval Process model, which, which can be realized, extracts the visual signature of target in Android and IOS, special
It levies size and is less than 4KB.
S602 extracts the visual signature of query image.
The global characteristics and local feature of target area in original input picture are extracted simultaneously.Visual signature includes global special
It seeks peace local feature, it is the whole description to image that two kinds of features, which are to the different describing modes of image, and global characteristics are short, right
Image it is descriptive bad；Local feature is long, is the datail description to image, to the descriptive good of image.
S106 matches the detailed classification information of visual signature and original input picture.
The detailed classification information of global characteristics, local feature and corresponding original input picture is matched, is input to
In database.In order to cover more as far as possible video scene and commodity, the embodiment of the present invention is directed to all target and scene
Clustering has been carried out, the commodity major class list database of 1000 major class is summarized, has included the overall situation spy of commodity in the database
Sign, local feature, corresponding classification information and detailed classification information.
Original image is input in the target detection identification model built, obtains the classification information of original image, so
The global characteristics and local feature for extracting target area in original image afterwards, first with global characteristics in the commodity major class pre-established
It is retrieved on a large scale in list database, global characteristics and good global characteristics extracted in database is compared,
Global characteristics matching probability is obtained, if matching probability is higher than preset threshold, illustrates that original image is similar to the image in database；
It is general to obtain local feature matching for the comparison for carrying out the local feature and local feature corresponding with global characteristics of original image again
Rate is arranged by sequence from high to low, selects the highest local feature of local feature matching probability, obtains correspond in the database
Detailed classification information, i.e. the detailed classification information of original image.Due to storing 1000 class merchandise newss, quantity in database
It is huge, therefore construct one in server end and supported based on CDVS characteristics of image coding and extensive visual signature index technology
The extensive visual search engine of million scale image data amounts, engine are supported to carry out visual signature sieve first, in accordance with classification information
Choosing, view-based access control model feature carry out Secondary Match to the result after screening, obtain most close with target visual feature in affiliated major class
Promotion item information.
Since global characteristics are shorter relative to local feature, retrieval can be improved using global characteristics in extensive retrieval
Efficiency.The identical situation of the local feature matching probability maximum value of two or more local features is theoretically occurred, but practical
It is not so not in such case unless these images are all same original images in.The mode for coping with theoretical case is to press
The sequence of database is put into according to local feature, by two or more identical local features of local feature matching probability maximum value
It is ranked up, chooses and arrange detailed classification information corresponding to primary local feature.
In embodiments of the present invention, to the addition of fresh target, the visual signature of extraction need to be only put in storage, is not needed anti-
Training objective detects identification model again, and expanded period is short, and scalability is strong.
The embodiment of the invention provides a kind of based on the precision target of object detection and signature search knowledge method for distinguishing and is
System obtains original input picture, is obtained by CNN convolutional neural networks as shown in figure 9, the system includes: module of target detection 1
To fisrt feature figure group；Resampling and concatenation are successively carried out to fisrt feature figure group, obtain second feature figure group；By second
Characteristic pattern group is mapped in original input picture, and the position detection of prediction block is carried out to original input picture, is originally inputted
The prediction block of image；Judge whether prediction block is target area, if so, obtaining the classification information in target area；Target is special
Extraction module 2 is levied, for extracting the visual signature of target area；Target signature search module 3, for by visual signature with it is original
The detailed classification information of input picture matches.
Resampling includes up-sampling and down-sampling, and module of target detection 1 is specifically used for adopt to fisrt feature figure group
Sample obtains third feature figure group；Characteristic pattern except out to out characteristic pattern in third feature figure group is up-sampled, is obtained
Fourth feature figure group；Characteristic pattern in third feature figure group and fourth feature figure with same scale is spliced respectively, is obtained
To second feature figure group.
Module of target detection 1 divides net also particularly useful for the scale based on second feature figure group on original input picture
Lattice；Each characteristic pattern in second feature figure group is respectively mapped in the grid of original input picture；It is at least generated on grid
One prediction block；The correction position that prediction block is calculated based on bounding box regression model, obtains the correction value of prediction block；Based on prediction
The correction value of frame and prediction block obtains the center of the prediction block of original input picture.
Target's feature-extraction module 2 is specifically used for determine that the original input picture behind target area is defeated as query image
Enter into Visual Feature Retrieval Process model；Extract the visual signature of query image.
Based on above-mentioned analysis it is found that compared with simple use in the related technology is based on the detection technique of deep learning, this
The method and system method based on the identification of the precision target of object detection and signature search that inventive embodiments provide utilizes object
The method that detection and signature search combine, realizes accurately identifying for target, can quickly add fresh target, has good expand
Malleability.
Carried out provided by the embodiment of the present invention based on the precision target of object detection and signature search know method for distinguishing and
The computer program product of system, the computer readable storage medium including storing program code, said program code include
Instruction can be used for executing previous methods method as described in the examples, specific implementation can be found in embodiment of the method, herein no longer
It repeats.
Device based on the identification of the precision target of object detection and signature search provided by the embodiment of the present invention can be
Specific hardware or the software being installed in equipment or firmware etc. in equipment.Device provided by the embodiment of the present invention, in fact
Existing principle and the technical effect of generation are identical with preceding method embodiment, and to briefly describe, Installation practice part does not refer to it
Place, can refer to corresponding contents in preceding method embodiment.It is apparent to those skilled in the art that for description
It is convenienct and succinct, system, the specific work process of device and unit of foregoing description, during reference can be made to the above method embodiment
Corresponding process, details are not described herein.
In embodiment provided by the present invention, it should be understood that disclosed device and method, it can be by others side
Formula is realized.The apparatus embodiments described above are merely exemplary, for example, the division of the unit, only one kind are patrolled
Function division is collected, there may be another division manner in actual implementation, in another example, multiple units or components can combine or can
To be integrated into another system, or some features can be ignored or not executed.Another point, shown or discussed is mutual
Coupling, direct-coupling or communication connection can be INDIRECT COUPLING or communication link by some communication interfaces, device or unit
It connects, can be electrical property, mechanical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme
's.
In addition, each functional unit in embodiment provided by the invention can integrate in one processing unit, it can also
To be that each unit physically exists alone, can also be integrated in one unit with two or more units.
It, can be with if the function is realized in the form of SFU software functional unit and when sold or used as an independent product
It is stored in a computer readable storage medium.Based on this understanding, technical solution of the present invention is substantially in other words
The part of the part that contributes to existing technology or the technical solution can be embodied in the form of software products, the meter
Calculation machine software product is stored in a storage medium, including some instructions are used so that a computer equipment (can be a
People's computer, server or network equipment etc.) it performs all or part of the steps of the method described in the various embodiments of the present invention.
And storage medium above-mentioned includes: that USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), arbitrary access are deposited
The various media that can store program code such as reservoir (RAM, Random Access Memory), magnetic or disk.
It should also be noted that similar label and letter indicate similar terms in following attached drawing, therefore, once a certain Xiang Yi
It is defined in a attached drawing, does not then need that it is further defined and explained in subsequent attached drawing, in addition, term " the
One ", " second ", " third " etc. are only used for distinguishing description, are not understood to indicate or imply relative importance.
Finally, it should be noted that embodiment described above, only a specific embodiment of the invention, to illustrate the present invention
Technical solution, rather than its limitations, scope of protection of the present invention is not limited thereto, although with reference to the foregoing embodiments to this hair
It is bright to be described in detail, those skilled in the art should understand that: anyone skilled in the art
In the technical scope disclosed by the present invention, it can still modify to technical solution documented by previous embodiment or can be light
It is readily conceivable that variation or equivalent replacement of some of the technical features；And these modifications, variation or replacement, do not make
The essence of corresponding technical solution is detached from the spirit and scope of technical solution of the embodiment of the present invention.Should all it cover in protection of the invention
Within the scope of.Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. a kind of know method for distinguishing based on the precision target of object detection and signature search characterized by comprising
Original input picture is obtained, obtains fisrt feature figure group by CNN convolutional neural networks；
Resampling and concatenation are successively carried out to the fisrt feature figure group, obtain second feature figure group；
The second feature figure group is mapped in the original input picture, prediction block is carried out to the original input picture
Position detection obtains the prediction block of the original input picture；
Judge whether the prediction block is target area, if so, obtaining the classification information in the target area；
Extract the visual signature of the target area；
The detailed classification information of visual signature and original input picture is matched.
2. the method according to claim 1, wherein the resampling includes up-sampling and down-sampling, to described
Fisrt feature figure group successively carries out resampling and concatenation, obtains second feature figure group and includes:
Down-sampling is carried out to the fisrt feature figure group, obtains third feature figure group；
Characteristic pattern except out to out characteristic pattern in the third feature figure group is up-sampled, fourth feature figure is obtained
Group；
Characteristic pattern in the third feature figure group and the fourth feature figure group with same scale is spliced respectively, is obtained
To the second feature figure group.
3. the method according to claim 1, wherein it is described the second feature figure group is mapped to it is described original
In input picture, prediction block position detection is carried out to the original input picture, obtains the prediction block of the original input picture
Include:
Scale based on second feature figure group grid division on the original input picture；
Each characteristic pattern in the second feature figure group is respectively mapped in the grid of the original input picture；
A prediction block is at least generated on the grid；
The correction position that the prediction block is calculated based on bounding box regression model obtains the correction value of the prediction block；
Correction value based on the prediction block and the prediction block, obtains the centre bit of the prediction block of the original input picture
It sets.
4. judging whether the prediction block is target area packet the method according to claim 1, wherein described
It includes:
Destination probability is generated in the prediction block；The destination probability is judged whether less than the first preset threshold, if so, institute
Stating prediction block not is the target area；
If it is not, then the prediction block is the target area.
5. the method according to claim 1, wherein described if so, obtaining the classification in the target area
Information includes:
Generate the class probability of the target；
The destination probability is multiplied with the class probability, formation condition class probability；
The condition class probability maximum value is chosen, it will classification information input corresponding with the condition class probability maximum value
Into database.
6. the method according to claim 1, wherein the visual signature packet for extracting the object detection area
It includes:
It will determine that the original input picture behind target area is input in Visual Feature Retrieval Process model as query image；
Extract the visual signature of the query image.
7. a kind of system based on the identification of the precision target of object detection and signature search characterized by comprising
Module of target detection obtains original input picture, obtains fisrt feature figure group by CNN convolutional neural networks；To described
Fisrt feature figure group successively carries out resampling and concatenation, obtains second feature figure group；The second feature figure group is mapped
Into the original input picture, the position detection of prediction block is carried out to the original input picture, obtains described be originally inputted
The prediction block of image；Judge whether the prediction block is target area, if so, obtaining the classification letter in the target area
Breath；
Target's feature-extraction module, for extracting the visual signature of the target area；
Target signature search module, for the detailed classification information of visual signature and original input picture to match.
8. system according to claim 7, which is characterized in that the resampling includes up-sampling and down-sampling, the mesh
Mark detection module is specifically used for:
Down-sampling is carried out to the fisrt feature figure group, obtains third feature figure group；
Characteristic pattern except out to out characteristic pattern in the third feature figure group is up-sampled, fourth feature figure is obtained
Group；
Characteristic pattern in the third feature figure group and the fourth feature figure with same scale is spliced respectively, is obtained
The second feature figure group.
9. system according to claim 7, which is characterized in that the module of target detection also particularly useful for:
Scale based on second feature figure group grid division on the original input picture；
Each characteristic pattern in the second feature figure group is respectively mapped in the grid of the original input picture；
A prediction block is at least generated on the grid；
The correction position that the prediction block is calculated based on bounding box regression model obtains the correction value of the prediction block；
Correction value based on the prediction block and the prediction block, obtains the centre bit of the prediction block of the original input picture
It sets.
10. system according to claim 7, which is characterized in that the target's feature-extraction module is specifically used for:
It will determine that the original input picture behind target area is input in Visual Feature Retrieval Process model as query image；
Extract the visual signature of the query image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811540379.XA CN109697464A (en) | 2018-12-17 | 2018-12-17 | Method and system based on the identification of the precision target of object detection and signature search |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811540379.XA CN109697464A (en) | 2018-12-17 | 2018-12-17 | Method and system based on the identification of the precision target of object detection and signature search |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109697464A true CN109697464A (en) | 2019-04-30 |
Family
ID=66231792
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811540379.XA Pending CN109697464A (en) | 2018-12-17 | 2018-12-17 | Method and system based on the identification of the precision target of object detection and signature search |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109697464A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111159150A (en) * | 2019-12-19 | 2020-05-15 | 北京文安智能技术股份有限公司 | Data expansion method and device |
| CN111461145A (en) * | 2020-03-31 | 2020-07-28 | 中国科学院计算技术研究所 | Method for detecting target based on convolutional neural network |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101840518A (en) * | 2010-04-02 | 2010-09-22 | 中国科学院自动化研究所 | Biological vision mechanism-based object training and identifying method |
| US20130250050A1 (en) * | 2012-03-23 | 2013-09-26 | Objectvideo, Inc. | Video surveillance systems, devices and methods with improved 3d human pose and shape modeling |
| CN105574527A (en) * | 2015-12-14 | 2016-05-11 | 北京工业大学 | Quick object detection method based on local feature learning |
| US20160203614A1 (en) * | 2015-01-14 | 2016-07-14 | Samsung Electronics Co., Ltd. | Method and apparatus of detecting object using event-based sensor |
| CN106570522A (en) * | 2016-10-24 | 2017-04-19 | 中国科学院自动化研究所 | Object recognition model establishment method and object recognition method |
| CN107038448A (en) * | 2017-03-01 | 2017-08-11 | 中国科学院自动化研究所 | Target detection model building method |
| CN107122713A (en) * | 2017-03-27 | 2017-09-01 | 华南理工大学 | It is a kind of based on deep learning like physical property detection method |
| CN107316058A (en) * | 2017-06-15 | 2017-11-03 | 国家新闻出版广电总局广播科学研究院 | Improve the method for target detection performance by improving target classification and positional accuracy |
| CN107909093A (en) * | 2017-10-27 | 2018-04-13 | 浙江大华技术股份有限公司 | A kind of method and apparatus of Articles detecting |
| WO2018132961A1 (en) * | 2017-01-18 | 2018-07-26 | Nokia Technologies Oy | Apparatus, method and computer program product for object detection |
| CN108764247A (en) * | 2018-04-13 | 2018-11-06 | 中国科学院自动化研究所 | Deep learning object detecting method and device based on dense connection |
-
2018
- 2018-12-17 CN CN201811540379.XA patent/CN109697464A/en active Pending
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101840518A (en) * | 2010-04-02 | 2010-09-22 | 中国科学院自动化研究所 | Biological vision mechanism-based object training and identifying method |
| US20130250050A1 (en) * | 2012-03-23 | 2013-09-26 | Objectvideo, Inc. | Video surveillance systems, devices and methods with improved 3d human pose and shape modeling |
| US20160203614A1 (en) * | 2015-01-14 | 2016-07-14 | Samsung Electronics Co., Ltd. | Method and apparatus of detecting object using event-based sensor |
| CN105574527A (en) * | 2015-12-14 | 2016-05-11 | 北京工业大学 | Quick object detection method based on local feature learning |
| CN106570522A (en) * | 2016-10-24 | 2017-04-19 | 中国科学院自动化研究所 | Object recognition model establishment method and object recognition method |
| WO2018132961A1 (en) * | 2017-01-18 | 2018-07-26 | Nokia Technologies Oy | Apparatus, method and computer program product for object detection |
| CN107038448A (en) * | 2017-03-01 | 2017-08-11 | 中国科学院自动化研究所 | Target detection model building method |
| CN107122713A (en) * | 2017-03-27 | 2017-09-01 | 华南理工大学 | It is a kind of based on deep learning like physical property detection method |
| CN107316058A (en) * | 2017-06-15 | 2017-11-03 | 国家新闻出版广电总局广播科学研究院 | Improve the method for target detection performance by improving target classification and positional accuracy |
| CN107909093A (en) * | 2017-10-27 | 2018-04-13 | 浙江大华技术股份有限公司 | A kind of method and apparatus of Articles detecting |
| CN108764247A (en) * | 2018-04-13 | 2018-11-06 | 中国科学院自动化研究所 | Deep learning object detecting method and device based on dense connection |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111159150A (en) * | 2019-12-19 | 2020-05-15 | 北京文安智能技术股份有限公司 | Data expansion method and device |
| CN111461145A (en) * | 2020-03-31 | 2020-07-28 | 中国科学院计算技术研究所 | Method for detecting target based on convolutional neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110147743A (en) | Real-time online pedestrian analysis and number system and method under a kind of complex scene | |
| CN110378222A (en) | A kind of vibration damper on power transmission line target detection and defect identification method and device | |
| KR101930940B1 (en) | Apparatus and method for analyzing image | |
| CN109063649B (en) | Pedestrian re-identification method based on twin pedestrian alignment residual error network | |
| CN110705566B (en) | Multi-mode fusion significance detection method based on spatial pyramid pool | |
| CN110569814B (en) | Video category identification method, device, computer equipment and computer storage medium | |
| CN112884033B (en) | Household garbage classification detection method based on convolutional neural network | |
| CN109829065B (en) | Image retrieval method, device, equipment and computer readable storage medium | |
| CN113034506B (en) | Remote sensing image semantic segmentation method and device, computer equipment and storage medium | |
| CN112613356B (en) | Action detection method and device based on deep attention fusion network | |
| CN110751195A (en) | Fine-grained image classification method based on improved YOLOv3 | |
| CN112991364A (en) | Road scene semantic segmentation method based on convolution neural network cross-modal fusion | |
| CN105574545B (en) | The semantic cutting method of street environment image various visual angles and device | |
| CN114332889A (en) | Text box ordering method and text box ordering device for text image | |
| CN109657082A (en) | Remote sensing images multi-tag search method and system based on full convolutional neural networks | |
| CN109697464A (en) | Method and system based on the identification of the precision target of object detection and signature search | |
| CN115761627A (en) | Fire smoke flame image identification method | |
| CN115455171A (en) | Method, device, equipment and medium for mutual retrieval and model training of text videos | |
| CN108388904A (en) | A kind of dimension reduction method based on convolutional neural networks and covariance tensor matrix | |
| CN111507359A (en) | Self-adaptive weighting fusion method of image feature pyramid | |
| CN111368775A (en) | Complex scene dense target detection method based on local context sensing | |
| CN113094533B (en) | Image-text cross-modal retrieval method based on mixed granularity matching | |
| CN112668675B (en) | Image processing method and device, computer equipment and storage medium | |
| CN114168768A (en) | Image retrieval method and related equipment | |
| CN113723558A (en) | Remote sensing image small sample ship detection method based on attention mechanism |
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 | ||
| AD01 | Patent right deemed abandoned | ||
| AD01 | Patent right deemed abandoned |
Effective date of abandoning: 20210910 |