CN108875901A - Neural network training method and generic object detection method, device and system - Google Patents
Neural network training method and generic object detection method, device and system Download PDFInfo
- Publication number
- CN108875901A CN108875901A CN201711161464.0A CN201711161464A CN108875901A CN 108875901 A CN108875901 A CN 108875901A CN 201711161464 A CN201711161464 A CN 201711161464A CN 108875901 A CN108875901 A CN 108875901A
- Authority
- CN
- China
- Prior art keywords
- convolution
- block
- result
- classifier
- characteristic pattern
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of neural network training method and generic object detection methods, device and system, are related to detection technique field, are applied to generic object and detect, which includes：Process of convolution is carried out to input picture, obtain the characteristic pattern of input picture, process of convolution at least twice is carried out to characteristic pattern, utilize at least two classifiers, generic object detection is carried out respectively to the intermediate features figure after each process of convolution, obtain at least two testing results, at least two testing results are merged, obtain final detection result, anti-pass loss is carried out to one or more classifiers according to final detection result, update the parameter in one or more classifiers, solve the problems, such as that will lead to precision not high using convolutional neural networks in the prior art, while guaranteeing computational efficiency, rate of false alarm can be reduced, and then improve detection accuracy.
Description
Technical field
The present invention relates to technical field of image detection, general more particularly, to a kind of neural network training method and one kind
Object detecting method, device and system.
Background technique
Neural network is a kind of extensive, multi-parameters optimization tool.By a large amount of training data, neural network can
Learn the hiding feature for being difficult to summarize in data out, so that the task of multinomial complexity is completed, for example, Face datection, picture classification,
Object detection, movement tracking, natural language translation etc., in short, neural network is widely applied by artificial intelligence circle.And compared to
The detection of the particular categories object such as Face datection or pedestrian detection, generic object detection is detection field problem the most extensive.
The reason is that, the object-oriented classification of generic object detection institute is more (may be up to tens classes or hundreds of classes), it more difficult to which it is accurate to obtain
Positioning.
Summary of the invention
In view of this, the purpose of the present invention is to provide neural network training method and generic object detection methods, dress
It sets and system, to reduce the rate of false alarm in generic object detection process.
In a first aspect, neural network is used for generic object the embodiment of the invention provides a kind of neural network training method
Detection；
Method includes：
Process of convolution is carried out to input picture, obtains the characteristic pattern of input picture；
Process of convolution at least twice is carried out to characteristic pattern, using at least two classifiers, in after each process of convolution
Between characteristic pattern carry out convolution detection processing respectively, obtain at least two testing results；
At least two testing results are merged, final detection result is obtained；
Anti-pass loss is carried out to one or more classifiers according to final detection result, is updated in one or more classifiers
Parameter.
With reference to first aspect, the embodiment of the invention provides the first possible embodiments of first aspect, wherein right
Characteristic pattern carries out process of convolution at least twice, using at least two classifiers, to the intermediate features figure after each process of convolution point
Not carry out convolution detection processing, obtain at least two testing result intermediate features figures, specifically include：
Characteristic pattern is divided into multiple blocks；
Process of convolution at least twice is carried out to characteristic pattern；
Using at least two classifiers, convolution detection processing is carried out respectively to the intermediate features figure after each process of convolution,
Obtain at least two testing results of each block.
The possible embodiment of with reference to first aspect the first, the embodiment of the invention provides second of first aspect
Possible embodiment, wherein at least two testing results are merged, final detection result is obtained, specifically includes：
For each block, if each testing result of the block is to be received, the final detection result of the block
To be received；
If at least one testing result of the block is to be rejected, the final detection result of the block is to be rejected.
The possible embodiment of second with reference to first aspect, the embodiment of the invention provides the third of first aspect
Possible embodiment, wherein according to final detection result to one or more classifiers carry out anti-pass loss, update one or
Parameter in multiple classifiers, specifically includes：
For each block, the final detection result of the block is compared with the legitimate reading of the block, is somebody's turn to do
The comparison result of block；
Anti-pass loss is carried out to one or more classifiers according to the comparison result of the block, updates one or more classification
Parameter in device.
The third possible embodiment with reference to first aspect, the embodiment of the invention provides the 4th kind of first aspect
Possible embodiment, wherein anti-pass loss is carried out to one or more classifiers according to the comparison result of the block, updates one
Parameter in a or multiple classifiers, specifically includes：
If the legitimate reading of the block is that should be rejected, final detection result is to be rejected, then to refusing the block at first
Classifier anti-pass loss, update the parameter refused in the classifier of the block at first；
If the legitimate reading of the block is that should be rejected, final detection result is to be received, then closest to testing result
The classifier anti-pass of confidence level is lost, and updates testing result closest to the parameter in the classifier of confidence level；
If the legitimate reading of the block is that should be received, final detection result is to be received, then closest to testing result
The classifier anti-pass of confidence level is lost, and updates testing result closest to the parameter in the classifier of confidence level；
If the legitimate reading of the block is that should be received, final detection result is to be rejected, then is lower than to testing result and sets
The classifier anti-pass of reliability is lost, and updates testing result lower than the parameter in the classifier of confidence level.
Second aspect, the embodiment of the invention also provides a kind of generic object detection methods, including：Input picture is carried out
Process of convolution obtains the characteristic pattern of input picture；
Process of convolution at least twice is carried out to characteristic pattern, using at least two classifiers, in after each process of convolution
Between characteristic pattern carry out convolution detection processing respectively, obtain at least two testing results；
At least two testing results are merged, final detection result is obtained.
In conjunction with second aspect, the embodiment of the invention provides the first possible embodiments of second aspect, wherein right
Characteristic pattern carries out process of convolution at least twice, using at least two classifiers, to the intermediate features figure after each process of convolution point
Not carry out convolution detection processing, intermediate features figure obtain at least two testing results, specifically include：
Characteristic pattern is divided into multiple blocks；
Process of convolution at least twice is carried out to characteristic pattern；
Using at least two classifiers, convolution detection processing is carried out respectively to the intermediate features figure after each process of convolution,
Obtain at least two testing results of each block.
In conjunction with the first possible embodiment of second aspect, the embodiment of the invention provides second of second aspect
Possible embodiment, wherein at least two testing results are merged, final detection result is obtained, specifically includes：
For each block, if each testing result of the block is to be received, the final detection result of the block
To be received；
If at least one testing result of the block is to be rejected, the final detection result of the block is to be rejected.
The third aspect, the embodiment of the invention also provides a kind of generic object detection devices, including：
Characteristic pattern obtains module, for carrying out process of convolution to input picture, obtains the characteristic pattern of input picture；
Detection module, for carrying out process of convolution at least twice to characteristic pattern, using at least two classifiers, to every secondary volume
Product treated intermediate features figure carries out respectively convolution detection processing, obtains at least two testing results；
As a result merging module obtains final detection result for merging at least two testing results.
Fourth aspect, the embodiment of the invention also provides a kind of generic object detection systems, including：Image collecting device,
Processor and storage device；
Image collecting device, for acquiring image information to be identified；
Computer program is stored on storage device, computer program executes any of the above-described when being run by processor
Method.
5th aspect, the embodiment of the invention also provides a kind of computer readable storage medium, computer-readable storage mediums
The step of being stored with computer program in matter, the method for any of the above-described executed when computer program is run by processor.
The embodiment of the present invention brings following beneficial effect：In neural network training method provided in an embodiment of the present invention,
Classifier is increased, also, above-mentioned classifier multistage is arranged.It is to input figure first when needing to carry out generic object detection
Piece carries out process of convolution, obtains characteristic pattern, later, carries out process of convolution at least twice to characteristic pattern, process of convolution here is logical
It is often continuous, also, utilizes at least two classifiers, needs exist for illustrating, number and the progress process of convolution of classifier
Number is consistent, and carries out convolution detection processing respectively to the intermediate features figure after each process of convolution, obtains at least two inspections
It surveys as a result, then, at least two testing results are merged, to obtain final detection result, finally, according to final detection
As a result anti-pass loss is carried out to one or more classifiers, updates the parameter in one or more classifiers.Therefore, the present invention is real
Apply example offer neural network training method, especially suitable for the detection of generic object, i.e., by by convolutional neural networks and point
Class device is combined, and is reduced the simple phenomenon high using convolutional neural networks rate of false alarm, is improved the detection to generic object
Precision.
Other features and advantages of the present invention will illustrate in the following description, also, partly become from specification
It obtains it is clear that understand through the implementation of the invention.The objectives and other advantages of the invention are in specification, claims
And specifically noted structure is achieved and obtained in attached drawing.
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
It, below will be to specific in order to illustrate more clearly of the specific embodiment of the invention or technical solution in the prior art
Embodiment or attached drawing needed to be used in the description of the prior art be briefly described, it should be apparent that, it is described below
Attached drawing is some embodiments of the present invention, for those of ordinary skill in the art, before not making the creative labor
It puts, is also possible to obtain other drawings based on these drawings.
Fig. 1 is the schematic diagram for the electronic equipment that the embodiment of the present invention one provides；
Fig. 2 is the flow chart of neural network training method provided by Embodiment 2 of the present invention；
Fig. 3 is the flow chart of the second step of neural network training method provided by Embodiment 2 of the present invention；
Fig. 4 is the flow chart of the third step of neural network training method provided by Embodiment 2 of the present invention；
Fig. 5 is the flow chart for the generic object detection method that the embodiment of the present invention three provides；
Fig. 6 is the flow chart of the second step for the generic object detection method that the embodiment of the present invention three provides；
Fig. 7 is the flow chart of the third step for the generic object detection method that the embodiment of the present invention three provides；
Fig. 8 is the schematic diagram for the generic object detection device that the embodiment of the present invention four provides.
Icon：
1- characteristic pattern obtains module；2- detection module；3- result merging module.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with attached drawing to the present invention
Technical solution be clearly and completely described, it is clear that described embodiments are some of the embodiments of the present invention, rather than
Whole embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art are not making creative work premise
Under every other embodiment obtained, shall fall within the protection scope of the present invention.
Currently, generic object detection depends on convolutional neural networks, a large amount of wrong reports can be generated during processing, examined
It is low to survey precision, is based on this, a kind of neural network training method provided in an embodiment of the present invention and generic object detection method, dress
It sets and system, rate of false alarm can be reduced while guaranteeing computational efficiency, promote detection accuracy.
For convenient for understanding the present embodiment, first to a kind of neural metwork training side disclosed in the embodiment of the present invention
The exemplary electronic device of method describes in detail.
Embodiment one：
Firstly, electric to describe a kind of example of neural network training method for realizing the embodiment of the present invention referring to Fig.1
Sub- 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 image acquisition device 110, these components pass through bus system 112 and/or other forms
The interconnection of bindiny mechanism's (not shown).It should be noted that the component and structure of electronic equipment 100 shown in FIG. 1 are only exemplary, and
Unrestricted, as needed, electronic equipment also can have other assemblies and structure.
Processor 102 can be central processing unit (CPU) or have data-handling capacity and/or instruction execution capability
Other forms processing unit, and can control other components in electronic equipment 100 to execute desired function.
Storage device 104 may include one or more computer program products, and computer program product may include each
The computer readable storage medium of kind form, such as volatile memory and/or nonvolatile memory.Volatile memory example
It such as may include random access memory (RAM) and/or cache memory (cache).Nonvolatile memory is for example
It may include read-only memory (ROM), hard disk, flash memory etc..It can store one or more on computer readable storage medium
Computer program instructions, processor 102 can run program instruction, to realize in the embodiment of the present invention hereafter (by processor
Realize) client functionality and/or other desired functions.It can also store in a computer-readable storage medium various
Application program and various data, such as application program use and/or the various data generated etc..
Input unit 106 can be the device that user is used to input instruction, and may include keyboard, mouse, microphone
One or more of with touch screen etc..
Output device 108 can export various information (for example, image or sound) to external (for example, user), and can
To include one or more of display, loudspeaker etc..
Image acquisition device 110 can acquire image information to be identified, and acquired image information is stored in
For the use of other components in storage device 104.
It illustratively, can be with for realizing the exemplary electronic device of neural network training method according to an embodiment of the present invention
It is implemented as on the mobile terminals such as smart phone, tablet computer.
Embodiment two：
The embodiment of the invention provides a kind of neural network training methods.
According to embodiments of the present invention, a kind of embodiment of neural network training method is provided, it should be noted that attached
The step of process of figure illustrates can execute in a computer system such as a set of computer executable instructions, though also,
So logical order is shown in flow charts, but in some cases, it can be to be different from shown by sequence execution herein
Or the step of description.
As shown in Fig. 2, the neural network training method of the embodiment of the present invention is detected for generic object, specifically include as follows
Step：
S101：Process of convolution is carried out to input picture, obtains the characteristic pattern of input picture.
Specifically, the picture that input tape object space marks in full convolutional neural networks is rolled up when being trained
Product processing, to obtain features described above figure.It is in order to by final detection result that the purpose with object space mark is provided on picture
It is compared with the mark, to carry out anti-pass loss to classifier, achievees the purpose that trained.
S102：Process of convolution at least twice is carried out to characteristic pattern, using at least two classifiers, after each process of convolution
Intermediate features figure carry out convolution detection processing respectively, obtain at least two testing results.
Specifically, features described above figure is carried out continuous several times process of convolution, here, the number of progress process of convolution is limited extremely
Few is twice, it is, of course, also possible to include more than number twice.In addition, carrying out the number of process of convolution and the number of classifier
It is consistent.In this way, after to a characteristic pattern process of convolution of every progress, the intermediate features figure that will be obtained using a classifier
Convolution detection processing is carried out, a testing result is obtained, later, the step is repeated, to obtain more testing results.Alternatively,
Multiple convolution processing continuously is carried out to characteristic pattern, obtaining continuous multiple intermediate features figures will obtain using multiple classifiers later
Multiple intermediate features figures carry out convolution detection processing, obtain multiple testing results.
As shown in figure 3, above-mentioned steps S102 carries out process of convolution at least twice to characteristic pattern, at least two classification are utilized
Device carries out convolution detection processing to the intermediate features figure after each process of convolution respectively, obtains at least two testing results, specifically
Including：
S1021：Characteristic pattern is divided into multiple blocks.
By taking 9 blocks that characteristic pattern is divided into 3*3 as an example, by above-mentioned block according to from top to bottom, from left to right suitable
Sequence is successively marked as 1,2,3 ... 8,9.
S1022：Process of convolution at least twice is carried out to characteristic pattern.
When it is implemented, be as unit of block, to characteristic pattern carry out process of convolution, i.e., respectively to each block carry out to
Few process of convolution twice.By taking the block marked as 1 as an example, cubic convolution processing is carried out to it, respectively obtain intermediate features Figure 11,
Intermediate features Figure 12 and intermediate features Figure 13.
S1023：Using at least two classifiers, convolution detection is carried out respectively to the intermediate features figure after each process of convolution
Processing, obtains at least two testing results of each block.
Since the block marked as 1 has only carried out cubic convolution processing, so three classifiers of setting, are denoted as point respectively
Class device 11, classifier 12 and classifier 13.When implementing, carried out at convolution detection using 11 pairs of classifier intermediate characteristic patterns 11
Reason, obtains testing result 11, carries out convolution detection processing using 12 pairs of classifier intermediate characteristic patterns 12, obtains testing result 12,
Generic object detection is carried out using 13 pairs of classifier intermediate characteristic patterns 13, obtains testing result 13.
S103：At least two testing results are merged, final detection result is obtained.
Specifically, the number of testing result is two or more by testing result obtained in step S102, it will be upper
It states multiple testing results to merge, obtains final detection result, and exported.Above-mentioned testing result 11, detection are tied
Fruit 12 and testing result 13 merge, and obtain final detection result 1.
As shown in figure 4, above-mentioned steps S103 merges at least two testing results, final detection result, tool are obtained
Body includes：
S1031：For each block, if each testing result of the block is to be received, the most final inspection of the block
Surveying result is to be received.
Specifically, testing result 12 is to be received, and testing result 13 is if the testing result 11 of block 1 is to be received
Received, in this case, the final detection result of the block is to be received.
S1032：If at least one testing result of the block is to be rejected, the final detection result of the block is to be refused
Absolutely.
Specifically, if the testing result 11 of block 1 is to be rejected, alternatively, testing result 12 is to be rejected, alternatively, inspection
Surveying result 13 is to be rejected, as long as any one testing result of i.e. block 1 is to be rejected, in this case, the block
Final detection result is to be rejected.
S104：Anti-pass loss is carried out to one or more classifiers according to final detection result.
Specifically, being described in detail in two kinds of situation.
The first situation carries out anti-pass loss to one or more classifiers according to final detection result, specifically includes：
(1) for each block, the final detection result of the block is compared with the legitimate reading of the block, is obtained
The comparison result of the block.
The legitimate reading of the block is the result directly read by the mark on picture with object space, it is assumed that block 1
It is to be received, and above-mentioned final detection result is to be rejected, and is compared by the result that the position mark on picture is directly read
It was found that the comparison result of the block 1 is inconsistent.Specifically, testing result 11 is to be received, testing result 12 is to be received,
Testing result 13 is to be rejected.Obviously, testing result 13 is problematic.
(2) anti-pass loss is carried out to one or more classifiers according to the comparison result of the block.
When implementation, classifier 13 corresponding to testing result 13 is found out, anti-pass loss is carried out to classifier 13, updates one
Or the parameter in multiple classifiers.
Second situation carries out anti-pass loss to one or more classifiers according to the comparison result of the block, updates one
Parameter in a or multiple classifiers specifically includes：
(1) it should be rejected if the legitimate reading of the block is, final detection result is to be rejected, then to refusing the area at first
The classifier anti-pass of block is lost, and the parameter in the classifier for refusing the block at first is updated.
I.e. block 1 is the final detection knot being rejected, and above-mentioned by the result that the position mark on picture is directly read
Fruit is to be rejected, specifically, classifier 11 is first refusal 1 classifier of block, classifier 12 is that second refusal block 1 divides
Class device, classifier 13 are that third refuses 1 classifier of block, it should be noted that a statement of first, second and third
The statement made such as the sequence that executes of classifier is illustrated only, and should not be construed the position to classifier, importance.This
When, then it is lost to 11 anti-pass of classifier for refusing the block at first, updates the parameter in classifier 11.
(2) it should be rejected if the legitimate reading of the block is, final detection result is to be received, then most connects to testing result
The classifier anti-pass loss of nearly confidence level, updates testing result closest to the parameter in the classifier of confidence level.
I.e. block 1 is the final detection knot being rejected, and above-mentioned by the result that the position mark on picture is directly read
Fruit is to be received, specifically, classifier 11 is to be received, and confidence level is 60%, classifier 12 is is received, and confidence
Degree is 70%, and classifier 13 is to be received, and confidence level is 80%, it should be noted that first, second and third
Statement the table made such as illustrate only the sequence that executes of classifier, and should not be construed the position to classifier, importance
It states.At this moment, then it is lost to testing result closest to 11 anti-pass of classifier of confidence level i.e. 50%, updates the ginseng in classifier 11
Number.
(3) should be received if the legitimate reading of the block is, final detection result is to be received, then most connects to testing result
The classifier anti-pass loss of nearly confidence level, updates testing result closest to the parameter in the classifier of confidence level.
I.e. block 1 is the final detection knot received, and above-mentioned by the result that the position mark on picture is directly read
Fruit is to be received, specifically, classifier 11 is to be received, and confidence level is 60%, classifier 12 is is received, and confidence
Degree is 70%, and classifier 13 is to be received, and confidence level is 80%, it should be noted that first, second and third
Statement the table made such as illustrate only the sequence that executes of classifier, and should not be construed the position to classifier, importance
It states.At this moment, then the parameter in classifier 11 is updated to 11 anti-pass of the classifier loss closest to confidence level 50%.
(4) should be received if the legitimate reading of the block is, final detection result is to be rejected, then is lower than to testing result
The classifier anti-pass of confidence level is lost, and updates testing result lower than the parameter in the classifier of confidence level.
I.e. block 1 is the final detection knot received, and above-mentioned by the result that the position mark on picture is directly read
Fruit is to be rejected, specifically, classifier 11 is to be rejected, and confidence level is 60%, classifier 12 is is rejected, and confidence
Degree is 70%, and classifier 13 is to be rejected, and confidence level is 40%, it should be noted that first, second and third
Statement the table made such as illustrate only the sequence that executes of classifier, and should not be construed the position to classifier, importance
It states.At this moment, then 13 anti-pass of classifier to testing result lower than confidence level 50% is lost, and updates the parameter in classifier 13.
Classifier is increased during process of convolution in neural network training method provided in an embodiment of the present invention, and
And above-mentioned classifier multistage setting.When carrying out generic object detection, it is that process of convolution is carried out to input picture first, obtains
It inputting after the characteristic pattern of picture, process of convolution at least twice is carried out to characteristic pattern, process of convolution here is usually continuous,
Also, at least two classifiers are utilized, convolution detection processing is carried out respectively to the intermediate features figure after each process of convolution, are obtained
Then at least two testing results merge at least two testing results, to obtain final detection result, finally, root
Anti-pass loss is carried out to one or more classifiers according to final detection result, updates the parameter in one or more classifiers.Cause
This, neural network training method provided in an embodiment of the present invention, i.e., by being combined convolutional neural networks and classifier,
When progress generic object detection is trained, reduce phenomenon high using convolutional neural networks rate of false alarm in the prior art, into
And while guaranteeing computational efficiency, reduce rate of false alarm, improves the detection accuracy to generic object.
Embodiment three：
The embodiment of the invention provides a kind of generic object detection methods.
According to embodiments of the present invention, a kind of more detailed embodiment of generic object detection method is provided, is needed
Bright, step shown in the flowchart of the accompanying drawings can be held in a computer system such as a set of computer executable instructions
Row, although also, logical order is shown in flow charts, and it in some cases, can be to be different from sequence herein
Execute shown or described step.
As shown in figure 5, the generic object detection method of the embodiment of the present invention includes the following steps：
S201：Process of convolution is carried out to input picture, obtains the characteristic pattern of input picture.
Specifically, inputting generic object piece when carrying out generic object detection in full convolutional neural networks and carrying out convolution
Processing, to obtain features described above figure.
S202：Process of convolution at least twice is carried out to characteristic pattern, using at least two classifiers, after each process of convolution
Intermediate features figure carry out convolution detection processing respectively, obtain at least two testing results.
Specifically, features described above figure is carried out continuous several times process of convolution, here, the number of progress process of convolution is limited extremely
Few is twice, it is, of course, also possible to include more than number twice.In addition, carrying out the number of process of convolution and the number of classifier
It is consistent.In this way, after to a characteristic pattern process of convolution of every progress, the intermediate features figure that will be obtained using a classifier
Convolution detection processing is carried out, a testing result is obtained, later, the step is repeated, to obtain more testing results.Alternatively,
Multiple convolution processing continuously is carried out to characteristic pattern, obtains multiple intermediate features figures, it is later, more by what is obtained using multiple classifiers
A intermediate features figure carries out convolution detection processing, obtains multiple testing results.
As shown in fig. 6, above-mentioned steps S202 carries out process of convolution at least twice to characteristic pattern, at least two classification are utilized
Device carries out convolution detection processing to the intermediate features figure after each process of convolution respectively, obtains at least two testing results, specifically
Including：
S2021：Characteristic pattern is divided into multiple blocks.
By taking 9 blocks that characteristic pattern is divided into 3*3 as an example, by above-mentioned block according to from top to bottom, from left to right suitable
Sequence is successively marked as 1,2,3 ... 8,9.
S2022：Process of convolution at least twice is carried out to characteristic pattern.
When it is implemented, be as unit of block, to characteristic pattern carry out process of convolution, i.e., respectively to each block carry out to
Few process of convolution twice.By taking the block marked as 1 as an example, cubic convolution processing is carried out to it, respectively obtain intermediate features Figure 11,
Intermediate features Figure 12 and intermediate features Figure 13.
S2023：Using at least two classifiers, convolution detection is carried out respectively to the intermediate features figure after each process of convolution
Processing, obtains at least two testing results of each block.
Since the block marked as 1 has only carried out cubic convolution processing, so three classifiers of setting, are denoted as point respectively
Class device 11, classifier 12 and classifier 13.When implementing, generic object inspection is carried out using 11 pairs of classifier intermediate characteristic patterns 11
It surveys, obtains testing result 11, carry out generic object detection using 12 pairs of classifier intermediate characteristic patterns 12, obtain testing result 12,
Generic object detection is carried out using 13 pairs of classifier intermediate characteristic patterns 13, obtains testing result 13.
S203：At least two testing results are merged, final detection result is obtained.
Specifically, the number of testing result is two or more by testing result obtained in step S202, it will be upper
It states multiple testing results to merge, obtains final detection result, and exported.Above-mentioned testing result 11, detection are tied
Fruit 12 and testing result 13 merge, and obtain final detection result 1.
As shown in fig. 7, above-mentioned steps S203 merges at least two testing results, final detection result, tool are obtained
Body includes：
S2031：For each block, if each testing result of the block is to be received, the most final inspection of the block
Surveying result is to be received.
It is illustrated by taking above-mentioned block 1 as an example, if above-mentioned testing result 11 is to be received, testing result 12 is to be connect
By testing result 13 is to be received, then the final detection result of the block 1 is to be received.
S2032：If at least one testing result of the block is to be rejected, the final detection result of the block is to be refused
Absolutely.
Be illustrated by taking above-mentioned block 1 as an example, if above-mentioned testing result 11 be rejected, meanwhile, 12 He of testing result
Testing result 13 be received, another situation is that testing result 12 be rejected, meanwhile, testing result 11 and testing result 13
To be received, another situation is that testing result 13 be rejected, meanwhile, testing result 11 and testing result 12 are to be received,
Another situation is that testing result 11 and testing result 12 be rejected, meanwhile, testing result 13 is to be received, another situation
It is testing result 11 and testing result 13 to be rejected, meanwhile, testing result 12 is to be received, another situation is that testing result
12 and testing result 13 be rejected, meanwhile, testing result 11 be received, another situation is that testing result 11, detection knot
Fruit 12 and testing result 13 are to be rejected, and the final detection result of the block 1 is to be rejected.
Generic object detection method provided in an embodiment of the present invention increases classifier during process of convolution, in this way
When carrying out generic object detection, be first to input picture carry out process of convolution, obtain characteristic pattern, later, to characteristic pattern into
Capable process of convolution at least twice, process of convolution here is usually continuous, also, utilizes at least two classifiers, to each
Intermediate features figure after process of convolution carries out convolution detection processing respectively, obtains at least two testing results, then, will at least two
A testing result merges, to obtain final detection result.Therefore, generic object detection side provided in an embodiment of the present invention
Method, when carrying out generic object detection, reduces in the prior art that is, by being combined convolutional neural networks and classifier
Using the high phenomenon of convolutional neural networks rate of false alarm, and then while guaranteeing computational efficiency, reduce rate of false alarm, improves pair
The detection accuracy of generic object.
Example IV：
The embodiment of the invention provides a kind of generic object detection devices.The generic object detection device is mainly used for executing
Generic object detection method provided by above content of the embodiment of the present invention, below to generic object provided in an embodiment of the present invention
Detection device does specific introduction.
As shown in fig. 7, a kind of generic object detection device mainly includes：Characteristic pattern obtains module 1,2 and of detection module
As a result merging module 3.
Characteristic pattern obtains module 1 and is used to carry out process of convolution to input picture, obtains characteristic pattern.
Detection module 2 is used to carry out process of convolution at least twice to characteristic pattern, using at least two classifiers, to every secondary volume
Product treated intermediate features figure carries out respectively generic object detection, obtains at least two testing results.
As a result merging module 3 obtains final detection result for merging at least two testing results.
Generic object detection device provided in an embodiment of the present invention, with generic object detection method provided by the above embodiment
Technical characteristic having the same reaches identical technical effect so also can solve identical technical problem.
Embodiment five：
The embodiment of the invention provides a kind of generic object detection systems.The generic object detection system is mainly used for executing
Generic object detection method provided by above content of the embodiment of the present invention, below to generic object provided in an embodiment of the present invention
Detection system does specific introduction.
As shown in figure 8, the generic object detection system mainly includes：Image collecting device, processor and storage device；
Image collecting device, for acquiring image information to be identified；
Computer program is stored on storage device, computer program executes in above-described embodiment when being run by processor
The method.
In addition, the present invention also provides a kind of computer storage medium, for being stored as method provided by the above embodiment
Computer software instructions used.
It is apparent to those skilled in the art that for convenience and simplicity of description, the system of foregoing description
Specific work process, can refer to corresponding processes in the foregoing method embodiment, details are not described herein.
In the description of the embodiment of the present invention, it should be noted that term " center ", "upper", "lower", "left", "right",
The orientation or positional relationship of the instructions such as "vertical", "horizontal", "inner", "outside" is to be based on the orientation or positional relationship shown in the drawings,
Be merely for convenience of description of the present invention and simplification of the description, rather than the device or element of indication or suggestion meaning must have it is specific
Orientation, be constructed and operated in a specific orientation, therefore be not considered as limiting the invention.
It is apparent to those skilled in the art that for convenience and simplicity of description, the device of foregoing description,
The specific work process of module and unit, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.
In several embodiments provided by the present invention, it should be understood that disclosed systems, devices and methods, it can be with
It realizes by another way.The apparatus embodiments described above are merely exemplary, for example, the division of the module,
Only a kind of logical function partition, there may be another division manner in actual implementation, in another example, multiple module or components can
To combine or be desirably integrated into another system, or some features can be ignored or not executed.Another point, it is shown or beg for
The mutual coupling, direct-coupling or communication connection of opinion can be through some communication interfaces, device or module it is indirect
Coupling or communication connection can be electrical property, mechanical or other forms.
The module as illustrated by the separation member may or may not be physically separated, aobvious as module
The component shown may or may not be physical module, it can and it is in one place, or may be distributed over multiple
On network module.Some or all of the modules therein can be selected to realize the mesh of this embodiment scheme according to the actual needs
's.
In addition, each functional module in each embodiment of the present invention can integrate in a first processing module,
It can be modules to physically exist alone, can also be integrated in two or more modules in a module.
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 the executable non-volatile computer-readable storage medium of a processor.Based on this understanding, of the invention
Technical solution substantially the part of the part that contributes to existing technology or the technical solution can be with software in other words
The form of product embodies, which is stored in a storage medium, including some instructions use so that
One computer equipment (can be personal computer, server or the network equipment etc.) executes each embodiment institute of the present invention
State all or part of the steps of method.And storage medium above-mentioned includes：USB flash disk, mobile hard disk, read-only memory (ROM,
Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic or disk etc. it is each
Kind can store the medium of program code.
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 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 (11)
1. a kind of neural network training method, which is characterized in that the neural network is detected for generic object；
The method includes：
Process of convolution is carried out to input picture, obtains the characteristic pattern of the input picture；
Process of convolution at least twice is carried out to the characteristic pattern, using at least two classifiers, in after each process of convolution
Between characteristic pattern carry out convolution detection processing respectively, obtain at least two testing results；
At least two testing result is merged, final detection result is obtained；
Anti-pass loss is carried out to one or more classifiers according to the final detection result, is updated described in one or more
Parameter in classifier.
2. the method according to claim 1, wherein described carry out at convolution at least twice the characteristic pattern
Reason carries out convolution detection processing to the intermediate features figure after each process of convolution respectively, obtains extremely using at least two classifiers
Few two testing results, specifically include：
The characteristic pattern is divided into multiple blocks；
Process of convolution at least twice is carried out to the characteristic pattern；
Using at least two classifiers, convolution detection processing is carried out to the intermediate features figure after each process of convolution respectively, obtained
At least two testing results of each block.
3. according to the method described in claim 2, it is characterized in that, described merge at least two testing result,
Final detection result is obtained, is specifically included：
For each block, if each testing result of the block is to be received, the final detection result of the block is quilt
Receive；
If at least one testing result of the block is to be rejected, the final detection result of the block is to be rejected.
4. according to the method described in claim 3, it is characterized in that, it is described according to the final detection result to one or more
The classifier carries out anti-pass loss, updates the parameter in one or more classifiers, specifically includes：
For each block, the final detection result of the block is compared with the legitimate reading of the block, obtains the block
Comparison result；
Anti-pass loss is carried out to one or more classifiers according to the comparison result of the block, is updated described in one or more
Parameter in classifier.
5. according to the method described in claim 4, it is characterized in that, the comparison result according to the block is to one or more
The classifier carries out anti-pass loss, updates the parameter in one or more classifiers, specifically includes：
If the legitimate reading of the block is that should be rejected, final detection result is to be rejected, then to point for refusing the block at first
Class device anti-pass loss updates the parameter in the classifier for refusing the block at first；
If the legitimate reading of the block is that should be rejected, final detection result is to be received, then to testing result closest to confidence
The classifier anti-pass of degree is lost, and updates testing result closest to the parameter in the classifier of confidence level；
If the legitimate reading of the block is that should be received, final detection result is to be received, then to testing result closest to confidence
The classifier anti-pass of degree is lost, and updates testing result closest to the parameter in the classifier of confidence level；
If the legitimate reading of the block is that should be received, final detection result is to be rejected, then is lower than confidence level to testing result
Classifier anti-pass loss, update testing result lower than confidence level classifier in parameter.
6. a kind of generic object detection method, which is characterized in that including：
Process of convolution is carried out to input picture, obtains the characteristic pattern of the input picture；
Process of convolution at least twice is carried out to the characteristic pattern, using at least two classifiers, in after each process of convolution
Between characteristic pattern carry out convolution detection processing respectively, obtain at least two testing results；
At least two testing result is merged, final detection result is obtained.
7. according to the method described in claim 6, it is characterized in that, described carry out at convolution at least twice the characteristic pattern
Reason carries out convolution detection processing to the intermediate features figure after each process of convolution respectively, obtains extremely using at least two classifiers
Few two testing results, specifically include：
The characteristic pattern is divided into multiple blocks；
Process of convolution at least twice is carried out to the characteristic pattern；
Using at least two classifiers, convolution detection processing is carried out to the intermediate features figure after each process of convolution respectively, obtained
At least two testing results of each block.
8. the method according to the description of claim 7 is characterized in that described merge at least two testing result,
Final detection result is obtained, is specifically included：
For each block, if each testing result of the block is to be received, the final detection result of the block is quilt
Receive；
If at least one testing result of the block is to be rejected, the final detection result of the block is to be rejected.
9. a kind of generic object detection device, which is characterized in that including：
Characteristic pattern obtains module, for carrying out process of convolution to input picture, obtains the characteristic pattern of the input picture；
Detection module, for carrying out process of convolution at least twice to the characteristic pattern, using at least two classifiers, to every secondary volume
Product treated intermediate features figure carries out respectively convolution detection processing, obtains at least two testing results；
As a result merging module obtains final detection result for merging at least two testing result.
10. a kind of generic object detection system, which is characterized in that the system comprises：Image collecting device, processor and storage
Device；
Described image acquisition device, for acquiring image information to be identified；
Computer program is stored on the storage device, the computer program is executed when being run by the processor as weighed
Benefit require any one of 6 to 8 described in method.
11. a kind of computer readable storage medium, computer program, feature are stored on the computer readable storage medium
The step of being, the described in any item methods of the claims 6 to 8 executed when the computer program is run by processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711161464.0A CN108875901B (en) | 2017-11-20 | 2017-11-20 | Neural network training method and universal object detection method, device and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711161464.0A CN108875901B (en) | 2017-11-20 | 2017-11-20 | Neural network training method and universal object detection method, device and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108875901A true CN108875901A (en) | 2018-11-23 |
| CN108875901B CN108875901B (en) | 2021-03-23 |
Family
ID=64325763
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711161464.0A Active CN108875901B (en) | 2017-11-20 | 2017-11-20 | Neural network training method and universal object detection method, device and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108875901B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109961006A (en) * | 2019-01-30 | 2019-07-02 | 东华大学 | A kind of low pixel multiple target Face datection and crucial independent positioning method and alignment schemes |
| CN110427802A (en) * | 2019-06-18 | 2019-11-08 | 平安科技（深圳）有限公司 | AU detection method, device, electronic equipment and storage medium |
| CN111191769A (en) * | 2019-12-25 | 2020-05-22 | 中国科学院苏州纳米技术与纳米仿生研究所 | Self-adaptive neural network training and reasoning device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120250983A1 (en) * | 2011-03-30 | 2012-10-04 | Sony Corporation | Object detecting apparatus and method |
| CN103679185A (en) * | 2012-08-31 | 2014-03-26 | 富士通株式会社 | Convolutional neural network classifier system as well as training method, classifying method and application thereof |
| CN104850845A (en) * | 2015-05-30 | 2015-08-19 | 大连理工大学 | Traffic sign recognition method based on asymmetric convolution neural network |
| CN106778867A (en) * | 2016-12-15 | 2017-05-31 | 北京旷视科技有限公司 | Object detection method and device, neural network training method and device |
| CN107229942A (en) * | 2017-04-16 | 2017-10-03 | 北京工业大学 | A kind of convolutional neural networks rapid classification method based on multiple graders |
-
2017
- 2017-11-20 CN CN201711161464.0A patent/CN108875901B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120250983A1 (en) * | 2011-03-30 | 2012-10-04 | Sony Corporation | Object detecting apparatus and method |
| CN103679185A (en) * | 2012-08-31 | 2014-03-26 | 富士通株式会社 | Convolutional neural network classifier system as well as training method, classifying method and application thereof |
| CN104850845A (en) * | 2015-05-30 | 2015-08-19 | 大连理工大学 | Traffic sign recognition method based on asymmetric convolution neural network |
| CN106778867A (en) * | 2016-12-15 | 2017-05-31 | 北京旷视科技有限公司 | Object detection method and device, neural network training method and device |
| CN107229942A (en) * | 2017-04-16 | 2017-10-03 | 北京工业大学 | A kind of convolutional neural networks rapid classification method based on multiple graders |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109961006A (en) * | 2019-01-30 | 2019-07-02 | 东华大学 | A kind of low pixel multiple target Face datection and crucial independent positioning method and alignment schemes |
| CN110427802A (en) * | 2019-06-18 | 2019-11-08 | 平安科技（深圳）有限公司 | AU detection method, device, electronic equipment and storage medium |
| CN111191769A (en) * | 2019-12-25 | 2020-05-22 | 中国科学院苏州纳米技术与纳米仿生研究所 | Self-adaptive neural network training and reasoning device |
| CN111191769B (en) * | 2019-12-25 | 2024-03-05 | 中国科学院苏州纳米技术与纳米仿生研究所 | Self-adaptive neural network training and reasoning device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108875901B (en) | 2021-03-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740534B (en) | Image processing method, device and processing equipment | |
| CN109145766B (en) | Model training method and device, recognition method, electronic device and storage medium | |
| CN108595585B (en) | Sample data classification method, model training method, electronic equipment and storage medium | |
| CN109313490B (en) | Eye gaze tracking using neural networks | |
| CN107688823B (en) | A kind of characteristics of image acquisition methods and device, electronic equipment | |
| CN106650662B (en) | Target object shielding detection method and device | |
| CN109255352A (en) | Object detection method, apparatus and system | |
| CN109447990A (en) | Image, semantic dividing method, device, electronic equipment and computer-readable medium | |
| CN108875537B (en) | Object detection method, device and system and storage medium | |
| CN108960114A (en) | Human body recognition method and device, computer readable storage medium and electronic equipment | |
| CN111597884A (en) | Facial action unit identification method and device, electronic equipment and storage medium | |
| CN108363998A (en) | A kind of detection method of object, device, system and electronic equipment | |
| CN109740415A (en) | Vehicle attribute recognition methods and Related product | |
| CN107918767B (en) | Object detection method, device, electronic equipment and computer-readable medium | |
| US20210166058A1 (en) | Image generation method and computing device | |
| CN108875901A (en) | Neural network training method and generic object detection method, device and system | |
| CN113095370A (en) | Image recognition method and device, electronic equipment and storage medium | |
| CN111061394B (en) | Touch force identification method, training method and device of model thereof and electronic system | |
| CN106326853A (en) | Human face tracking method and device | |
| CN108241853A (en) | A kind of video frequency monitoring method, system and terminal device | |
| CN108133169A (en) | A kind of embark on journey processing method and its device for text image | |
| CN106484614A (en) | A kind of method of verification picture processing effect, device and mobile terminal | |
| CN111126358B (en) | Face detection method, device, storage medium and equipment | |
| CN107844338B (en) | Application program management-control method, device, medium and electronic equipment | |
| CN111159481A (en) | Edge prediction method and device of graph data and terminal 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |