CN108921111A - Object detection post-processing approach and corresponding intrument - Google Patents
Object detection post-processing approach and corresponding intrument Download PDFInfo
- Publication number
- CN108921111A CN108921111A CN201810744184.0A CN201810744184A CN108921111A CN 108921111 A CN108921111 A CN 108921111A CN 201810744184 A CN201810744184 A CN 201810744184A CN 108921111 A CN108921111 A CN 108921111A
- Authority
- CN
- China
- Prior art keywords
- post
- processing
- detection
- detection information
- erroneous detection
- 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
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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 embodiment of the present invention provides a kind of object detection post-processing approach and corresponding intrument, is related to technical field of image processing.This method includes obtaining to treat the first post-processing object that process object obtain after tentatively post-processing, and the first post-processing object includes the first erroneous detection information;Depth post-processing is carried out to the first post-processing object using two disaggregated models, to filter the corresponding target detection object of the first erroneous detection information acquisition object to be detected;Two disaggregated models are obtained according to the training of the first training set, and the first training set includes the negative example being made of the second erroneous detection information and the positive example marked.Device is for executing the above method.The embodiment of the present invention is by carrying out depth post-processing to the first post-processing object using according to the negative example being made of the second erroneous detection information and two disaggregated models of the positive example marked building, the first erroneous detection information in the first post-processing object can be filtered out while guaranteeing recall rate, reduce the false detection rate treated when test object is detected.
Description
Technical field
The present invention relates to technical field of image processing, in particular to a kind of object detection post-processing approach and correspondence
Device.
Background technique
Object detection (object detection) is an important subject in computer vision.In particular with nothing
People drives, related needs is growing under video monitoring and wisdom security protection scene, and object detection obtains more and more concerns.
There are two types of current most popular object detection frames:One is with the area faster based on convolutional neural networks feature
Domain method (faster R-CNN) is that the generation of representative assumes that frame (region proposal) is then done the two-stage classified
(two-stage) method, another kind are the direct stages (one-stage) that detection block is returned from appearance using YOLO as representative
Method.Regardless of any method for checking object, there is the post-processing step of a deduplication frame before exporting final result, it is intended to
Avoid the detection block that multiple high superposeds are exported to the same object.
The post-processing of object detection at present is mainly based upon unsupervised non-maximum value restrainable algorithms (Non-maximum
Suppression,NMS).However, the threshold size in NMS method is to recall rate (Recall) and false detection rate (False
Positive, FP) have an impact, therefore, it is difficult to reach a good tradeoff between high recall rate and low false detection rate.
Summary of the invention
In view of this, the embodiment of the present invention is designed to provide a kind of object detection post-processing approach and corresponding intrument,
To solve the above technical problems.
In a first aspect, the embodiment of the invention provides a kind of object detection post-processing approach, including:
Obtain the first post-processing object treated and obtained after process object progress tentatively post-processes;Wherein, described to be processed
Object obtains after model treats test object progress object detection after testing, and the first post-processing object includes first
Erroneous detection information;
Depth post-processing is carried out to the first post-processing object using two disaggregated models, to filter first post-processing
The first erroneous detection information in object obtains the corresponding target detection object of the object to be detected;Wherein, two classification
Model is obtained according to the training of the first training set, and first training set includes the positive example marked and by the second erroneous detection information
The negative example constituted.
Further, before obtaining the first post-processing object treated and obtained after process object tentatively post-process,
The method further includes:
It obtains the detection model and the detection model is verified using verifying collection, obtain the second erroneous detection letter
Breath;
The negative example being made of the second erroneous detection information is obtained from first training set, and has marked object just
Example;
It is exercised supervision study according to the positive example and the negative example, obtains two disaggregated model.
Further, before obtaining the first post-processing object treated and obtained after process object tentatively post-process,
The method further includes:
The target value of threshold parameter defined in non-maximum value restrainable algorithms is obtained according to verifying collection;
Using the corresponding non-maximum value restrainable algorithms of the target value to the object to be processed in the detection model
Tentatively post-processed.
Further, the target value that threshold parameter defined in non-maximum value restrainable algorithms is obtained according to verifying collection,
Including:
Verifying collection is inputted in the detection model, and the verifying is tied using the non-maximum value restrainable algorithms
Fruit is post-processed;
The threshold parameter in the non-maximum value restrainable algorithms is adjusted, until the verifying collects corresponding recall rate and expires
Until sufficient preset value, the target value of the threshold parameter is obtained.
Further, before obtaining the first post-processing object treated and obtained after process object tentatively post-process,
The method further includes:
The object to be detected is input in the detection model and carries out object detection, the detection model is through too deep
Spend what learning method obtained.
Further, before obtaining the first post-processing object treated and obtained after process object tentatively post-process,
The method further includes:
The detection model is obtained according to the training of the second training set.
Second aspect, the embodiment of the invention provides a kind of object detection after-treatment devices, including:
Module is obtained, for obtaining the first post-processing object treated and obtained after process object progress tentatively post-processes;Its
In, the object to be processed obtains after model treats test object progress object detection after testing, after described first
Managing object includes the first erroneous detection information;
Depth post-processing module, for carrying out depth post-processing to the first post-processing object using two disaggregated models,
To filter the first erroneous detection information in the first post-processing object, the corresponding target detection of the object to be detected is obtained
Object;Wherein, two disaggregated model is obtained according to the training of the first training set, and first training set includes having marked
Positive example and the negative example being made of the second erroneous detection information.
The third aspect, the embodiment of the present invention provide a kind of electronic equipment, including:Processor and memory, wherein
The memory is stored with the program instruction that can be executed by the processor, and the processor calls described program to refer to
Enable the method and step for being able to carry out first aspect.
Fourth aspect, the embodiment of the present invention provide a kind of non-transient computer readable storage medium, including:
The non-transient computer readable storage medium stores computer instruction, and the computer instruction makes the computer
Execute the method and step of first aspect.
The embodiment of the present invention is trained by the negative example being made of using basis the second erroneous detection information and the positive example marked
Two disaggregated models carry out depth post-processing to the first post-processing object, so as to while guaranteeing recall rate, filter out the
The first erroneous detection information in one post-processing object, reduces the false detection rate treated when test object is detected.
Other features and advantages of the present invention will be illustrated in subsequent specification, also, partly be become from specification
It is clear that by implementing understanding of the embodiment of the present invention.The objectives and other advantages of the invention can be by written theory
Specifically noted structure is achieved and obtained in bright book, claims and attached drawing.
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 is electronic devices structure schematic diagram provided in an embodiment of the present invention;
Fig. 2 is a kind of object detection post-processing approach flow diagram provided in an embodiment of the present invention;
Fig. 3 is a kind of object detection post-processing approach entire flow schematic diagram provided in an embodiment of the present invention;
Fig. 4 is a kind of object detection after-treatment device structural schematic diagram provided in an embodiment of the present invention.
Specific embodiment
Below in conjunction with attached drawing in the embodiment of the present invention, technical solution in the embodiment of the present invention carries out clear, complete
Ground description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Usually exist
The component of the embodiment of the present invention described and illustrated in attached drawing can be arranged and be designed with a variety of different configurations herein.Cause
This, is not intended to limit claimed invention to the detailed description of the embodiment of the present invention provided in the accompanying drawings below
Range, but it is merely representative of selected embodiment of the invention.Based on the embodiment of the present invention, those skilled in the art are not doing
Every other embodiment obtained under the premise of creative work out, shall fall within the protection scope of the present invention.
It should 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.Meanwhile of the invention
In description, term " first ", " second " etc. are only used for distinguishing description, are not understood to indicate or imply relative importance.
Firstly, describing referring to Fig.1 for realizing the object detection post-processing approach of the embodiment of the present invention and showing for device
Example electronic equipment 100.
As shown in Figure 1, electronic equipment 100 include one or more processors 102, it is one or more storage device 104, defeated
Enter device 106, output device 108 and object after-treatment device 110, these components pass through bus system 112 and/or other shapes
Bindiny mechanism's (not shown) of formula interconnects.It should be noted that the component and structure of electronic equipment 100 shown in FIG. 1 are exemplary
, and not restrictive, as needed, the electronic equipment also can have other assemblies and structure.
The processor 102 can be central processing unit (CPU) or have data-handling capacity and/or instruction execution
The processing unit of the other forms of ability, and the other components that can control in the electronic equipment 100 are desired to execute
Function.
The storage device 104 may include one or more computer program products, and the computer program product can
To include various forms of computer readable storage mediums, such as volatile memory and/or nonvolatile memory.It is described easy
The property lost memory for example may include random access memory (RAM) and/or cache memory (cache) etc..It is described non-
Volatile memory for example may include read-only memory (ROM), hard disk, flash memory etc..In the computer readable storage medium
On can store one or more computer program instructions, processor 102 can run described program instruction, to realize hereafter institute
The client functionality (realized by processor) in the embodiment of the present invention stated and/or other desired functions.In the meter
Can also store various application programs and various data in calculation machine readable storage medium storing program for executing, for example, the application program use and/or
The various data etc. generated.
The input unit 106 can be the device that user is used to input instruction, and may include keyboard, mouse, wheat
One or more of gram wind and touch screen etc..
The output device 108 can export various information (for example, image or sound) to external (for example, user), and
It and may include one or more of display, loudspeaker etc..
The object after-treatment device 110 can post-process the object to be processed after object detection.
Illustratively, for realizing object detection post-processing approach according to an embodiment of the present invention and the exemplary electron of device
Equipment may be implemented as on the mobile terminals such as smart phone, tablet computer.
Fig. 2 is a kind of object detection post-processing approach flow diagram provided in an embodiment of the present invention, as shown in Fig. 2, should
Method, including:
Step 201:Obtain the first post-processing object treated and obtained after process object progress tentatively post-processes;Wherein, institute
It states object to be processed to obtain after model treats test object progress object detection after testing, the first post-processing object
Including the first erroneous detection information.
Specifically, have the post-processing step of a deduplication frame before exporting final result in object detection process,
It is intended to the detection block for avoiding exporting the same object multiple high superposeds.Post-processing step is divided into two by the embodiment of the present invention
Point, a part is tentatively to post-process, and second part is depth post-processing, obtains the first post-processing object first, this is after first
Reason object is to carry out preliminary post-processing to the object to be processed in detection model to obtain, also, after tentatively post-processing
It include the first erroneous detection information in the first post-processing object obtained.It should be noted that object to be processed is detection model pair
It is obtained after object progress object detection to be detected, and object to be processed is a data acquisition system, such as:It can be multiple figures
Picture.For different applications, different erroneous detection information is corresponded to, therefore, for object to be detected, has passed through progress
After preliminary post-processing, some first erroneous detection information undoubtedly can be also introduced.It is understood that mistake described in the embodiment of the present invention
Inspection information is FP (False Positive), also referred to as false positive example.Preliminary post-processing can be through non-maximum value restrainable algorithms
It carries out, can also be and obtained by other traditional post-treatment methods, the embodiment of the present invention is not specifically limited in this embodiment.Should also
Illustrate, when tentatively being post-processed, should require be handled according to required recall rate, therefore, the acquired
One post-processing object meets recall rate requirement, but may be such that false detection rate rises.
Step 202:Depth post-processing is carried out to the first post-processing object using two disaggregated models, to filter described the
The first erroneous detection information in one post-processing object obtains the corresponding target detection object of the object to be detected;Wherein, institute
Stating two disaggregated models is obtained according to the training of the first training set, and first training set includes the positive example marked and by second
The negative example that erroneous detection information is constituted.
Specifically, acquisition first process object after, in order to by first process object in the first erroneous detection information filtering
Fall, to reduce false detection rate, needs to carry out depth post-processing to the first post-processing object using two disaggregated models, thus by first
The first erroneous detection information processing in post-processing object is fallen.It it should be noted that two disaggregated models construct in advance, and is basis
The training of first training set obtains, and the first training set includes the negative example being made of the second erroneous detection information and the positive example marked.
It is understood that positive example and negative example are obtained from the first training set, for negative example, to the bread in image
It is detected, if two in image strip bread is easy to produce error detection when leaning on closer, exports while including two faces
The detection block of packet, here it is the second erroneous detection information, therefore, can be by by the friendship of two pieces of strip bread and ratio
(intersection over union, IoU) is above a certain threshold value (such as 0.7) " as heuristic rule, from first
Acquisition includes the sample of the second erroneous detection information as negative example in training set.The so-called positive example marked is to instruct in advance to first
Practice multiple images concentrated to be labeled, before mark, carries out object definition first, can be by object definition:People moves
Object, plant, food etc. can also be defined further, and the embodiment of the present invention is not specifically limited in this embodiment;It then will figure
Object as in is framed with rectangle frame, and carries out object binding to it, to complete the mark to image.Due to two classification moulds
Type is obtained according to the negative example being made of the second erroneous detection information and the positive example marked training, therefore, can effectively be identified
The first erroneous detection information in first post-processing object out.
The embodiment of the present invention is by utilizing trained two points of the negative example that is made of the second erroneous detection information and the positive example marked
Class model carries out depth post-processing to the first post-processing object, so as to while guaranteeing recall rate, after filtering out first
The first erroneous detection information in process object reduces the false detection rate treated when test object is detected.
On the basis of the above embodiments, after obtaining treat and obtain after process object tentatively post-process first
Before managing object, the method further includes:
It obtains the detection model and the detection model is verified using verifying collection, obtain the second erroneous detection letter
Breath;
The negative example being made of the second erroneous detection information is obtained from first training set, and has marked object just
Example;
It is exercised supervision study according to the positive example and the negative example, obtains two disaggregated model.
Specifically, needing to construct two disaggregated models before treating test object and carrying out object detection.By verifying collection input
It is verified into detection model, according to performance of the detection model on verifying collection, obtains the feature of the second erroneous detection information, then
Corresponding second erroneous detection information is generated by designing some heuristic rules.It is understood that the second erroneous detection information can be
It is multiple.Still by taking the image with bread as an example, which is input in detection model, is then carried out just using NMS algorithm
Step post-processing, the image post-processed, it can be seen that two root long bread have been considered being a bread from this image
Output, for this kind of second erroneous detection information, we, which can set, " is above a certain threshold value (ratio with the IoU of two pieces of strip bread
As 0.7) " heuristic rule.It should be noted that different application scenarios can correspond to different heuristic rules, tool
Body needs to gain enlightenment according to the actual situation, and formula is regular, and the embodiment of the present invention does not limit heuristic rule concrete condition herein
It is fixed.
The object being made of the second erroneous detection information is obtained from the first training set as negative example, will have been marked in the first training set
Object is infused as positive example, it is to be understood that the number of the negative example and the number of positive example are multiple, and the number of positive example is more than
The number of negative example.Then the positive example got and negative example, training obtains two disaggregated models in the way of supervised learning.It answers
When explanation, there are many algorithms of supervised learning, such as decision Tree algorithms, Bayesian Classification Arithmetic, support vector machines etc.,
The embodiment of the present invention is not specifically limited in this embodiment.
The embodiment of the present invention obtains the second erroneous detection information by collecting according to verifying, then obtains from the first training set by second
The negative example that erroneous detection information is constituted finally exercises supervision according to negative example and positive example and learns to obtain two classifiers, which can
Effectively filter out the first erroneous detection information introduced in preliminary post-processing.
On the basis of the above embodiments, after obtaining to the object to be detected progress tentatively post-processing in detection model
Before the first post-processing object arrived, the method further includes:
The object to be detected in the detection model is tentatively post-processed.
Specifically, needing to obtain to after object detection after treating test object by detection model and carrying out object detection
Object to be processed post-processed, it is necessary first to tentatively post-processed, when tentatively being post-processed it is used after place
Reason method can be NMS algorithm, be also possible to other post-processing algorithms, and the embodiment of the present invention is not specifically limited in this embodiment.It should
Illustrate, detection model is also to construct in advance, and construction method can be using faster R-CNN or YOLO etc., the present invention
Embodiment is not specifically limited in this embodiment.
The embodiment of the present invention is by utilizing trained two points of the negative example that is made of the second erroneous detection information and the positive example marked
Class model carries out depth post-processing to the first post-processing object, so as to while guaranteeing recall rate, after filtering out first
The first erroneous detection information in process object reduces the false detection rate treated when test object is detected.
On the basis of the above embodiments, after obtaining treat and obtain after process object tentatively post-process first
Before managing object, the method further includes:
The target value of threshold parameter defined in non-maximum value restrainable algorithms is obtained according to verifying collection;
Using the corresponding non-maximum value restrainable algorithms of the target value to the object to be processed in the detection model
Tentatively post-processed.
Specifically, the thresholding parameter value in NMS is set when tentatively being post-processed using non-maximum value restrainable algorithms NMS
Fixed is higher, then using verifying collection to verifying, judges whether the recall rate on verifying collection meets the requirements, if discontented
Foot, then thresholding parameter value is set it is some higher, until verifying collection on recall rate meet the requirements until, to obtain corresponding
The target value of threshold parameter.Process object is treated using the corresponding NMS of the target value for meeting recall rate tentatively to be post-processed.
The value of threshold parameter in NMS by being improved the requirement for meeting recall rate by the embodiment of the present invention, therefrom not
Exempt from that some first erroneous detection information can be introduced, therefore, recycles two classifiers to be filtered the first erroneous detection information, to reach
Not only meet the requirement of recall rate, but also reduce false detection rate.
On the basis of the above embodiments, described that the ginseng of threshold value defined in non-maximum value restrainable algorithms is obtained according to verifying collection
Several target values, including:
Verifying collection is inputted in the detection model, and the verifying is tied using the non-maximum value restrainable algorithms
Fruit is post-processed;
The threshold parameter in the non-maximum value restrainable algorithms is adjusted, until the verifying collects corresponding recall rate and expires
Until sufficient preset value, the target value of the threshold parameter is obtained.
In the specific implementation process, when determining the value of the threshold parameter in NMS algorithm, one can be preset
Then verifying collection is input in detection model by value, detection model carries out object detection to verifying collection and generates testing result, then
Testing result is post-processed using NMS, if post-processing corresponding recall rate is unsatisfactory for preset value, is adjusted in NMS
Threshold parameter, and repeat the above steps, it is known that until the corresponding recall rate of verifying collection meets preset value, obtains recall rate and meet in advance
If when value, the value of threshold parameter in NMS, using the value as target value.
The embodiment of the present invention adjusts the threshold parameter in NMS by verifying collection, so that target value meets recall rate
Requirement, can therefrom introduce some first erroneous detection information unavoidably, therefore, recycle two classifiers the first erroneous detection information was carried out
Filter, to reach the requirement for not only meeting recall rate, but also reduces false detection rate.
On the basis of the above embodiments, after obtaining treat and obtain after process object tentatively post-process first
Before process object, the method further includes:
The object to be detected is input in the detection model and carries out object detection, the detection model is through too deep
Spend what learning method obtained.
Specifically, by deep learning method construct detection model, wherein the detection model can be by R-CNN,
The methods of fast R-CNN, faster R-CNN or YOLO training obtain, and object to be detected is input in detection model and is carried out
Object detection can be acquired after object detection and is framed to the object in image with rectangle frame.It should be noted that pair
Object in image can also use round or other polygon circles after being detected, and the embodiment of the present invention, which does not do this, to be had
Body limits, in addition, may include multiple repeat blocks in the picture, therefore, it is necessary to carry out the last handling process of duplicate removal.
The embodiment of the present invention passes through using according to the negative example of the second erroneous detection information and two classification of the positive example building marked
Model carries out depth post-processing to the first post-processing object, so as to while guaranteeing recall rate, after filtering out first
The first erroneous detection information in object is managed, the false detection rate treated when test object is detected is reduced.
On the basis of the above embodiments, after obtaining treat and obtain after process object tentatively post-process first
Before managing object, the method further includes:
The detection model is obtained according to the training of the second training set.
Specifically, obtaining the second training set, detection model is obtained using traditional algorithm training by the second training set,
In, traditional algorithm can be faster R-CNN, can also be other methods, and the embodiment of the present invention does not limit this specifically
It is fixed.It should be noted that the data sample in the first training set and the second training set be it is the same, only to the mark of data sample
Injecting method is different, is data sample have been carried out to the mark of positive and negative example, and train in the first training set of two disaggregated models of training
Second training set of detection model only marks the type of object but without the mark of positive and negative example.
The embodiment of the present invention is by utilizing trained two points of the negative example that is made of the second erroneous detection information and the positive example marked
Class model carries out depth post-processing to the first post-processing object, so as to while guaranteeing recall rate, after filtering out first
The first erroneous detection information in process object reduces the false detection rate treated when test object is detected.
Fig. 3 is a kind of object detection post-processing approach entire flow schematic diagram provided in an embodiment of the present invention, such as Fig. 3 institute
Show, including:
Firstly, the deep learning method can by the second training set using deep learning method one detection model of training
Think faster R-CNN;
Object detection is carried out secondly, object to be detected is input in detection model, obtains object to be processed;
Again, the threshold parameter in NMS model is set, detection model is verified by verifying collection, whether judges it
The recall rate of meet demand continues raising threshold parameter if being unsatisfactory for and utilizes the NMS model until meeting recall rate
It treats process object tentatively to be post-processed, obtains the first post-processing object, meanwhile, it can also introduce additional the first erroneous detection letter
Breath;
Again, the verifying according to verifying collection to detection model, obtains the second erroneous detection information, in general specifically detects mould
Type adds specific application scenarios, and testing result also has apparent deviation, i.e., typical erroneous detection information;
Again, obtained from training set according to the second erroneous detection information include the second erroneous detection information object as negative example,
Simultaneously using the object for being labeled and having classified in training set as positive example, one two classification of mode training of supervised learning is used
Model;
Finally, carrying out depth post-processing, mistake to the first post-processing object that preliminary post-processing obtains by two disaggregated models
Filter the first erroneous detection information that NMS model tentatively post-processes introducing.
The embodiment of the present invention passes through using according to the negative example of the second erroneous detection information and two classification of the positive example building marked
Model carries out depth post-processing to the first post-processing object, so as to while guaranteeing recall rate, after filtering out first
The first erroneous detection information in object is managed, the false detection rate treated when test object is detected is reduced.
Fig. 4 is a kind of object detection after-treatment device structural schematic diagram provided in an embodiment of the present invention, as shown in figure 4, should
Device includes:Obtain module 401 and depth post-processing module 402, wherein
It obtains module 401 and carries out the obtain after tentatively post-processing first post-processing object for object to be processed;Wherein, institute
It states object to be processed to obtain after model treats test object progress object detection after testing, the first post-processing object
Including the first erroneous detection information;
After depth post-processing module 402 is used to carry out depth to the first post-processing object using two disaggregated models
Reason obtains the corresponding target of the object to be detected to filter the first erroneous detection information in the first post-processing object
Test object;Wherein, two disaggregated model is obtained according to the training of the first training set, and first training set includes having marked
The positive example of note and the negative example being made of the second erroneous detection information.
On the basis of the above embodiments, described device further includes:
Authentication module is obtained for obtaining the detection model and being verified using verifying collection to the detection model
The second erroneous detection information;
Sample acquisition module, for obtaining the negative example being made of the second erroneous detection information from first training set,
And the positive example of object is marked;
Two disaggregated models construct module, for being exercised supervision study according to the positive example and the negative example, acquisition described two
Disaggregated model.
On the basis of the above embodiments, described device further includes:
Preliminary post-processing module, for tentatively being post-processed to the object to be detected in the detection model.
On the basis of the above embodiments, the preliminary post-processing module, is specifically used for:
The target value of threshold parameter defined in non-maximum value restrainable algorithms is obtained according to verifying collection;
Using the corresponding non-maximum value restrainable algorithms of the target value to the object to be processed in the detection model
Tentatively post-processed.
On the basis of the above embodiments, the preliminary post-processing module, is specifically used for:
Verifying collection is inputted in the detection model, and the verifying is tied using the non-maximum value restrainable algorithms
Fruit is post-processed;
The threshold parameter in the non-maximum value restrainable algorithms is adjusted, until the verifying collects corresponding recall rate and expires
Until sufficient preset value, the target value of the threshold parameter is obtained.
On the basis of the above embodiments, described device further includes:
Obj ect detection module carries out object detection for the object to be detected to be input in the detection model, institute
Stating detection model is obtained by deep learning method.
On the basis of the above embodiments, described device further includes:
Detection model training module, for obtaining the detection model according to the training of the second training set.
It is apparent to those skilled in the art that for convenience and simplicity of description, the device of foregoing description
Specific work process, no longer can excessively be repeated herein with reference to the corresponding process in preceding method.
In conclusion the embodiment of the present invention is by using according to the negative example that is made of the second erroneous detection information and having marked just
Two disaggregated models of example building carry out depth post-processing to the first post-processing object, so as to while guaranteeing recall rate,
The first erroneous detection information in the first post-processing object is filtered out, the false detection rate treated when test object is detected is reduced.
The present embodiment discloses a kind of computer program product, and the computer program product includes being stored in non-transient calculating
Computer program on machine readable storage medium storing program for executing, the computer program include program instruction, when described program instruction is calculated
When machine executes, computer is able to carry out method provided by above-mentioned each method embodiment, for example including:Process object is treated in acquisition
Carry out the obtain after tentatively post-processing first post-processing object;Wherein, the object to be processed be after testing model to be checked
It is obtained after survey object progress object detection, the first post-processing object includes the first erroneous detection information;Utilize two disaggregated models
Depth post-processing is carried out to the first post-processing object, to filter first erroneous detection letter in the first post-processing object
Breath obtains the corresponding target detection object of the object to be detected;Wherein, two disaggregated model is according to the first training training
It gets, first training set includes the positive example marked and the negative example that is made of the second erroneous detection information.The present embodiment mentions
For a kind of non-transient computer readable storage medium, the non-transient computer readable storage medium stores computer instruction, institute
Stating computer instruction makes the computer execute method provided by above-mentioned each method embodiment, for example including:Place is treated in acquisition
Reason object carries out the obtain after tentatively post-processing first post-processing object;Wherein, the object to be processed is model after testing
It treats after test object carries out object detection and obtains, the first post-processing object includes the first erroneous detection information;Utilize two points
Class model carries out depth post-processing to the first post-processing object, to filter described first in the first post-processing object
Erroneous detection information obtains the corresponding target detection object of the object to be detected;Wherein, two disaggregated model is according to the first instruction
Practice training to get, first training set includes the positive example marked and the negative example that is made of the second erroneous detection information.
In several embodiments provided herein, it should be understood that disclosed device and method can also pass through
Other modes are realized.The apparatus embodiments described above are merely exemplary, for example, flow chart and block diagram in attached drawing
Show the device of multiple embodiments according to the present invention, the architectural framework in the cards of method and computer program product,
Function and operation.In this regard, each box in flowchart or block diagram can represent the one of a module, section or code
Part, a part of the module, section or code, which includes that one or more is for implementing the specified logical function, to be held
Row instruction.It should also be noted that function marked in the box can also be to be different from some implementations as replacement
The sequence marked in attached drawing occurs.For example, two continuous boxes can actually be basically executed in parallel, they are sometimes
It can execute in the opposite order, this depends on the function involved.It is also noted that every in block diagram and or flow chart
The combination of box in a box and block diagram and or flow chart can use the dedicated base for executing defined function or movement
It realizes, or can realize using a combination of dedicated hardware and computer instructions in the system of hardware.
In addition, each functional module in each embodiment of the present invention can integrate one independent portion of formation together
Point, it is also possible to modules individualism, an independent part can also be integrated to form with two or more modules.
It, can be with if the function is realized and when sold or used as an independent product in the form of software function module
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: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.
The foregoing is only a preferred embodiment of the present invention, is not intended to restrict the invention, for the skill of this field
For art personnel, the invention may be variously modified and varied.All within the spirits and principles of the present invention, made any to repair
Change, equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.It should be noted that:Similar label and letter exist
Similar terms are indicated in following attached drawing, therefore, once being defined in a certain Xiang Yi attached drawing, are then not required in subsequent attached drawing
It is further defined and explained.
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
Those familiar with the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all contain
Lid is within protection scope of the present invention.Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
It should be noted that, in this document, relational terms such as first and second and the like are used merely to a reality
Body or operation are distinguished with another entity or operation, are deposited without necessarily requiring or implying between these entities or operation
In any actual relationship or order or sequence.Moreover, the terms "include", "comprise" or its any other variant are intended to
Non-exclusive inclusion, so that the process, method, article or equipment including a series of elements is not only wanted including those
Element, but also including other elements that are not explicitly listed, or further include for this process, method, article or equipment
Intrinsic element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that
There is also other identical elements in process, method, article or equipment including the element.
Claims (9)
1. a kind of object detection post-processing approach, which is characterized in that including:
Obtain the first post-processing object treated and obtained after process object progress tentatively post-processes;Wherein, the object to be processed
Model is treated after test object carries out object detection and is obtained after testing, and the first post-processing object includes the first erroneous detection
Information;
Depth post-processing is carried out to the first post-processing object using two disaggregated models, to filter the first post-processing object
In the first erroneous detection information, obtain the corresponding target detection object of the object to be detected;Wherein, two disaggregated model
It is to be obtained according to the training of the first training set, first training set includes the positive example marked and is made of the second erroneous detection information
Negative example.
2. being obtained the method according to claim 1, wherein being treated after process object tentatively post-process in acquisition
Before the first post-processing object arrived, the method further includes:
It obtains the detection model and the detection model is verified using verifying collection, obtain the second erroneous detection information;
The negative example being made of the second erroneous detection information is obtained from first training set, and has marked the positive example of object;
It is exercised supervision study according to the positive example and the negative example, obtains two disaggregated model.
3. being obtained the method according to claim 1, wherein being treated after process object tentatively post-process in acquisition
Before the first post-processing object arrived, the method further includes:
The target value of threshold parameter defined in non-maximum value restrainable algorithms is obtained according to verifying collection;
The object to be processed in the detection model is carried out using the target value corresponding non-maximum value restrainable algorithms
Preliminary post-processing.
4. according to the method described in claim 3, it is characterized in that, described obtain in non-maximum value restrainable algorithms according to verifying collection
The target value of the threshold parameter of definition, including:
Verifying collection is inputted in the detection model, and using the non-maximum value restrainable algorithms to the verification result into
Row post-processing;
The threshold parameter in the non-maximum value restrainable algorithms is adjusted, until the verifying collects corresponding recall rate and meets in advance
If obtaining the target value of the threshold parameter until value.
5. being obtained the method according to claim 1, wherein being treated after process object tentatively post-process in acquisition
Before the first post-processing object arrived, the method further includes:
The object to be detected is input in the detection model and carries out object detection, the detection model is by depth
What learning method obtained.
6. according to the method described in claim 2, being obtained it is characterized in that, being treated after process object tentatively post-process in acquisition
Before the first post-processing object arrived, the method further includes:
The detection model is obtained according to the training of the second training set.
7. a kind of object detection after-treatment device, which is characterized in that including:
Module is obtained, for obtaining the first post-processing object treated and obtained after process object progress tentatively post-processes;Wherein, institute
It states object to be processed to obtain after model treats test object progress object detection after testing, the first post-processing object
Including the first erroneous detection information;
Depth post-processing module, for carrying out depth post-processing to the first post-processing object using two disaggregated models, with mistake
The first erroneous detection information in the first post-processing object is filtered, the corresponding target detection pair of the object to be detected is obtained
As;Wherein, two disaggregated model is obtained according to the training of the first training set, and first training set includes having marked just
Example and the negative example being made of the second erroneous detection information.
8. a kind of electronic equipment, which is characterized in that including:Processor and memory, wherein
The memory is stored with the program instruction that can be executed by the processor, and the processor calls described program to instruct energy
Enough execute as the method according to claim 1 to 6.
9. a kind of non-transient computer readable storage medium, which is characterized in that the non-transient computer readable storage medium is deposited
Computer instruction is stored up, the computer instruction makes the computer execute as the method according to claim 1 to 6.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810744184.0A CN108921111A (en) | 2018-07-06 | 2018-07-06 | Object detection post-processing approach and corresponding intrument |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810744184.0A CN108921111A (en) | 2018-07-06 | 2018-07-06 | Object detection post-processing approach and corresponding intrument |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108921111A true CN108921111A (en) | 2018-11-30 |
Family
ID=64425598
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810744184.0A Pending CN108921111A (en) | 2018-07-06 | 2018-07-06 | Object detection post-processing approach and corresponding intrument |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108921111A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109801643A (en) * | 2019-01-30 | 2019-05-24 | 龙马智芯(珠海横琴)科技有限公司 | The treating method and apparatus of Reverberation Rejection |
CN110110723A (en) * | 2019-05-07 | 2019-08-09 | 艾瑞迈迪科技石家庄有限公司 | A kind of method and device that objective area in image automatically extracts |
CN110135456A (en) * | 2019-04-08 | 2019-08-16 | 图麟信息科技(上海)有限公司 | A kind of training method and device of target detection model |
CN111861966A (en) * | 2019-04-18 | 2020-10-30 | 杭州海康威视数字技术股份有限公司 | Model training method and device and defect detection method and device |
EP4035070B1 (en) * | 2019-09-24 | 2024-01-17 | Motorola Solutions, Inc. | Method and server for facilitating improved training of a supervised machine learning process |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
CN106156808A (en) * | 2015-04-23 | 2016-11-23 | 佳能株式会社 | The on-line study method and apparatus of object classifier and detection method and equipment |
CN106295666A (en) * | 2015-05-14 | 2017-01-04 | 佳能株式会社 | Grader generates, updates and method for checking object and device and image processing equipment |
US9613296B1 (en) * | 2015-10-07 | 2017-04-04 | GumGum, Inc. | Selecting a set of exemplar images for use in an automated image object recognition system |
CN107193836A (en) * | 2016-03-15 | 2017-09-22 | 腾讯科技(深圳)有限公司 | A kind of recognition methods and device |
CN107967773A (en) * | 2017-12-01 | 2018-04-27 | 旗瀚科技有限公司 | A kind of supermarket self-help purchase method of view-based access control model identification |
-
2018
- 2018-07-06 CN CN201810744184.0A patent/CN108921111A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
CN106156808A (en) * | 2015-04-23 | 2016-11-23 | 佳能株式会社 | The on-line study method and apparatus of object classifier and detection method and equipment |
CN106295666A (en) * | 2015-05-14 | 2017-01-04 | 佳能株式会社 | Grader generates, updates and method for checking object and device and image processing equipment |
US9613296B1 (en) * | 2015-10-07 | 2017-04-04 | GumGum, Inc. | Selecting a set of exemplar images for use in an automated image object recognition system |
CN107193836A (en) * | 2016-03-15 | 2017-09-22 | 腾讯科技(深圳)有限公司 | A kind of recognition methods and device |
CN107967773A (en) * | 2017-12-01 | 2018-04-27 | 旗瀚科技有限公司 | A kind of supermarket self-help purchase method of view-based access control model identification |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109801643A (en) * | 2019-01-30 | 2019-05-24 | 龙马智芯(珠海横琴)科技有限公司 | The treating method and apparatus of Reverberation Rejection |
CN109801643B (en) * | 2019-01-30 | 2020-12-04 | 龙马智芯(珠海横琴)科技有限公司 | Processing method and device for reverberation suppression |
CN110135456A (en) * | 2019-04-08 | 2019-08-16 | 图麟信息科技(上海)有限公司 | A kind of training method and device of target detection model |
CN111861966A (en) * | 2019-04-18 | 2020-10-30 | 杭州海康威视数字技术股份有限公司 | Model training method and device and defect detection method and device |
CN111861966B (en) * | 2019-04-18 | 2023-10-27 | 杭州海康威视数字技术股份有限公司 | Model training method and device and defect detection method and device |
CN110110723A (en) * | 2019-05-07 | 2019-08-09 | 艾瑞迈迪科技石家庄有限公司 | A kind of method and device that objective area in image automatically extracts |
EP4035070B1 (en) * | 2019-09-24 | 2024-01-17 | Motorola Solutions, Inc. | Method and server for facilitating improved training of a supervised machine learning process |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108921111A (en) | Object detection post-processing approach and corresponding intrument | |
Chen et al. | Median filtering forensics based on convolutional neural networks | |
CN110728209B (en) | Gesture recognition method and device, electronic equipment and storage medium | |
CN108351961A (en) | Image and characteristic mass merge ocular angiogenesis and face and/or sub- facial information for the image enhancement and feature extraction of ocular angiogenesis and face recognition and for biological recognition system | |
CN110728330A (en) | Object identification method, device, equipment and storage medium based on artificial intelligence | |
WO2021189364A1 (en) | Method and device for generating adversarial image, equipment, and readable storage medium | |
CN109583449A (en) | Character identifying method and Related product | |
CN107316029B (en) | A kind of living body verification method and equipment | |
CN111079639A (en) | Method, device and equipment for constructing garbage image classification model and storage medium | |
CN109756458B (en) | Identity authentication method and system | |
EP3693876A1 (en) | Biometric authentication, identification and detection method and device for mobile terminal and equipment | |
CN106845406A (en) | Head and shoulder detection method and device based on multitask concatenated convolutional neutral net | |
CN107274543B (en) | A kind of recognition methods of bank note, device, terminal device and computer storage medium | |
CN109543548A (en) | A kind of face identification method, device and storage medium | |
CN106650615B (en) | A kind of image processing method and terminal | |
CN108416343B (en) | Face image recognition method and device | |
CN105046882B (en) | Fall down detection method and device | |
CN108108731A (en) | Method for text detection and device based on generated data | |
US20220309635A1 (en) | Computer vision-based anomaly detection method, device and electronic apparatus | |
CN110991444A (en) | Complex scene-oriented license plate recognition method and device | |
CN107423306A (en) | A kind of image search method and device | |
CN110321892A (en) | A kind of picture screening technique, device and electronic equipment | |
CN114358278A (en) | Training method and device of neural network model | |
CN110765843A (en) | Face verification method and device, computer equipment and storage medium | |
CN111860078A (en) | Face silence living body detection method and device, readable storage medium and equipment |
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: 20181130 |
|
RJ01 | Rejection of invention patent application after publication |