CN110210482A - Improve the object detection method of class imbalance - Google Patents
Improve the object detection method of class imbalance Download PDFInfo
- Publication number
- CN110210482A CN110210482A CN201910494701.8A CN201910494701A CN110210482A CN 110210482 A CN110210482 A CN 110210482A CN 201910494701 A CN201910494701 A CN 201910494701A CN 110210482 A CN110210482 A CN 110210482A
- Authority
- CN
- China
- Prior art keywords
- mask
- anchor frame
- case
- positive
- mask case
- 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
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of object detection methods for improving class imbalance, comprising: adds the branched structure for generating mask figure in traditional characteristic detection network, obtains the mask figure for having identical size with the characteristic pattern of given resolution；Mask figure is trained, is the positive and negative label of each mask case distribution in mask figure according to training result, for positive mask case, anchor frame thereon is trained, for negative mask case, ignores anchor frame on it, it is not involved in training, finally obtains trained target detection model；In test phase, the screening of anchor frame is carried out using trained target detection model, and then obtains object detection results.This method can be applied on the object detector based on deep learning well, preferably solve the problems, such as that positive and negative sample is unbalanced in depth targets detector, to promote the accuracy of algorithm of target detection.
Description
Technical field
The present invention relates to target detection technique field more particularly to a kind of object detection methods for improving class imbalance.
Background technique
Target detection is the basis in computer vision, its task is to orient all objects from an image,
And complete the Classification and Identification to the object.It is on major internal authority data set the result shows that, it is existing based on deep
The object detector of degree study is state-of-the-art detection method, these object detection methods use anchor frame to carry out detection block mostly
Prediction.Above-mentioned anchor frame (anchor) can be understood as preset box, the anchor frame that object detector passes through a variety of patterns of setting
Remove the object of covering different sizes.For the characteristic pattern of a W × H, if the anchor frame of 9 kinds of setting different patterns,
W × H × 9 anchor frame is just had, from the perspective of input picture, the center of these anchor frames is equidistantly to be evenly distributed on image
On.These object detectors of training surround how to train anchor frame and be unfolded, and such as will overlap bigger anchor with real-world object
Frame regards as positive sample (positive anchor frame), returns to the coordinate of this anchor frame, and effort allows it close to true object area, and
And use the classification of real-world object as supervision, classification task is carried out to anchor frame；Smaller anchor frame is overlapped with real-world object to regard
For negative sample (negative anchor frame), ignore the task of its coordinate recurrence, and uses the classification task of " background " this kind of supervision anchor frames.
However, the training of anchor frame will be influenced by serious class imbalance in the training process of object detector.This
It is the negligible amounts of positive anchor frame because the object in image is limited, and the quantity of negative anchor frame is made far more than the quantity of positive anchor frame
It is uneven at positive negative ratio.When the type setting of anchor frame is less, positive negative ratio is likely to be breached 1:1000 or so, and when anchor frame
When type setting is more, it is possible to reach 1:10000 or more.Such unbalanced ratio will greatly influence object detector
Reliability, so that " background " classification is more partial in prediction, recall rate is reduced.
What the various variants that the existing method for solving class imbalance mainly passes through bias weight method were realized.
It is such as sampled in training, it is specified that positive negative ratio is 1:3；Or when training network carries out backpropagation, only consider loss
Biggish " difficult sample " balances positive negative ratio；Either for one adaptive weight of all sample weightings.It in this way can band
Carry out two problems: (1) parameter introduced is difficult to illustrate how that setting is better, can only be chosen by didactic method,
Effect depends on data set；(2) method sampled carrys out power-assisted in model training simply by the means of mitigation, and there is no substantially
Change positive negative ratio of the anchor frame in training, training will be caused asynchronous with when test.
Summary of the invention
The object of the present invention is to provide a kind of object detection methods for improving class imbalance, promote the accurate of target detection
Degree.
The purpose of the present invention is what is be achieved through the following technical solutions:
A kind of object detection method improving class imbalance, comprising:
The branched structure for generating mask figure is added in traditional characteristic detection network, obtains the spy with given resolution
Sign figure has the mask figure of identical size；
Mask figure is trained, is the positive and negative label of each mask case distribution in mask figure according to training result, for
Positive mask case is trained anchor frame thereon, for negative mask case, ignores anchor frame on it, is not involved in training, final to obtain
To trained target detection model；
In test phase, the screening of anchor frame is carried out using trained target detection model, and then obtains target detection knot
Fruit.
As seen from the above technical solution provided by the invention, anchor frame is solved in object detector compared to traditional
The method of class imbalance, the present invention are more biased towards in fundamentally going to solve the problems, such as, reduction needs to consider in training and test
Anchor frame quantity, and devise a simple and efficient mask mechanism, can apply well based on deep learning
On object detector, preferably solve the problems, such as that positive and negative sample is unbalanced in depth targets detector, to promote target inspection
The accuracy of method of determining and calculating.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, required use in being described below to embodiment
Attached drawing be briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for this
For the those of ordinary skill in field, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing.
Fig. 1 is a kind of flow chart of object detection method for improving class imbalance provided in an embodiment of the present invention；
Fig. 2 is the structure chart of training stage mask mechanism provided in an embodiment of the present invention；
Fig. 3 is the schematic diagram provided in an embodiment of the present invention that anchor point frame training is instructed using mask figure.
Specific embodiment
With reference to the 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.Based on this
The embodiment of invention, every other implementation obtained by those of ordinary skill in the art without making creative efforts
Example, belongs to protection scope of the present invention.
The phenomenon that proportional imbalance positive and negative towards anchor point in the object detection method based on deep learning, is different from general inclined
Method of weighting is set, present invention introduces mask mechanism, propose a kind of object detection method for improving class imbalance, cardinal principle
Are as follows: to create a 2D mask figure, the width of mask figure and high width and Gao Xiangtong with characteristic pattern for the characteristic pattern of detection.Mask
Figure is made of grid one by one, referred to as mask case；Mask case can provide the score between prediction 0 to 1 accordingly, judge this
It whether include positive anchor frame in region, positive mask case is counted as the mask case that is positive in training, and the higher explanation of score is comprising just covering
The probability of code lattice is bigger；Only select that the anchor frame in positive mask case is trained in training；In test according to mask score
The classification score that provides with anchor frame obtains prediction block, to complete target detection.
As shown in Figure 1, a kind of object detection method for improving class imbalance provided in an embodiment of the present invention mainly includes
Following steps:
Step 1 detects the branched structure added in network for generating mask figure in traditional characteristic, obtains differentiating with specified
The characteristic pattern of rate has the mask figure of identical size.
In the embodiment of the present invention, traditional backbone network structure is improved, the backbone network is target detection
Convolutional neural networks in model, it will be appreciated that for the network for extracting detection feature.Spy of the backbone network in different resolution
It levies and is provided with different detection layers on figure, for detecting the different object of scale.It is done accordingly by the structure for backbone network
Improvement, can more agree with the present invention using mask figure Further aim and detect unbalanced method.
As shown in Fig. 2, the preferred embodiment for improving structure is as follows:
1) traditional characteristic detection network can extract the characteristic pattern of several different resolutions, by addition pond layer to each
The character figure of resolution ratio carries out down-sampling, and the characteristic pattern of each resolution ratio is narrowed down to the characteristic pattern with lowest resolution
Identical size, then all characteristic patterns are spliced.After carrying out pond-concatenation, so that it may obtain comprising different rulers
Object is spent, semantic information characteristic pattern abundant, on this characteristic pattern, the object information of different scale can all have corresponding table
It reaches, the characteristic pattern that Jiang Chihua-concatenation obtains is denoted as characteristic pattern A.
2) make size change over w × h × 1 of characteristic pattern A by adding convolutional layer, wherein w and h be respectively resolution ratio most
The width of low characteristic pattern and height；It adds convolutional layer and on the one hand reduces calculation amount for reducing dimension；It on the other hand is that addition is desired
Study mapping, allows it that can express object information on spatial position.
3) mask layer is added, regards the characteristic pattern A after size change over as mask figure.There is w × h grid in mask figure, often
One grid is trained and (will be introduced in subsequent step 2) using logistic regression.So far, backbone network structure is completed
Modification.It should be noted that above-mentioned modification can also suitably be repaired according to actual conditions.
Step 2 is trained mask figure, is the positive and negative mark of each mask case distribution in mask figure according to training result
Label, for positive mask case, are trained anchor frame thereon, for negative mask case, ignore anchor frame on it, be not involved in instruction
Practice, finally obtains trained target detection model.
This step preferred embodiment is as follows:
1) mask figure training.
In the embodiment of the present invention, for the mask figure that above-mentioned steps 2 obtain, using the training of logistic regression mode, logic is returned
The loss function returned is expressed as follows:
Lmask=∑ { yln σ (x)-(1-y) ln [1- σ (x)] }
Wherein, y is the label of mask case, and x is the output of mask case, and σ (x) expression becomes x using sigmoid function
It changes, makes its output between 0 to 1, the numerical value converted indicates that the center of the true box of object is fallen in inside corresponding mask case
Probability；
The distribution principle of the numerical value and mask case label that obtain according to transformation is that each mask case distributes positive and negative label；Its
In, the distribution principle of mask case label are as follows: if the center of the true box of object is fallen in mask case, distributing corresponding mask case is
Positive mask case；If this object, completely except mask case, this distributes corresponding mask case and is negative mask case.
Above-mentioned mask figure is the characteristic pattern of lowest resolution, in order to be applied to multiple detection layers, training process
In and subsequent introduction test phase, carrying out up-sampling operation to mask figure makes it extend to traditional characteristic detection network
All detection layers.
It will be understood by those skilled in the art that the effect of detection layers is the classification score for generating corresponding anchor frame using feature
(classification) adjusts the coordinate transform (coordinate recurrence) of anchor frame；Work is not involved in feature extraction phases detection layers.
In the embodiment of the present invention, it is desirable to generate a mask figure, quickly filter " negative sample ".Existing target detection side
" multilayer pyramid " prediction mode is mostly used in method, therefore is up-sampled herein, is applied on each layer, so as to
It plays a role in training and test.
Top sampling method is as follows: the filling of each mask case of the mask figure after expansion is that the preceding mask figure of expansion is corresponding
The duplication of positional value；In the training process, up-sampling is also done by for the label distribution of mask case；Therefore, mask figure
All detection layers can be extended to by way of up-sampling.
2) anchor frame training
After having mask figure, for anchor frame training also different from general method, in the embodiment of the present invention, according to covering
Code lattice positive and negative tag types carry out the training of anchor frame, as shown in figure 3, for positive mask case (in such as Fig. 3 three with compared with thick lines mark
The region of 2 × 2 sizes of note), corresponding anchor frame according to true box is overlapping is allocated than (IoU)；And for negative mask
Lattice then directly ignore anchor frame on it, are not involved in training.So, the negative anchor frame quantity for participating in training will greatly reduce,
The positive negative ratio of serious unbalance will greatly be mitigated.
The overall loss function of training anchor frame can be expressed as follows:
L=Lobj+Lmask+Lcls+αLloc
Wherein, LobjFor the loss of anchor frame object likelihood, LmaskFor the loss of mask figure, LclsFor Classification Loss, LlocFor
The loss that bounding box returns, α is weight factor, may be configured as 2, indicates that model is more prone to first guarantee the accuracy of positioning.
Step 3, in test phase, the screening of anchor frame is carried out using trained target detection model, and then obtain target
Testing result.
The preferred embodiment of this step is as follows:
Two threshold values are set, and first threshold value is for screening mask case, and second threshold value is for screening anchor frame；
Input picture is handled by trained target detection model, obtains corresponding mask figure；
By first threshold value, corresponding mask case is filtered out from mask figure；Pass through second threshold value covering from screening
Corresponding anchor frame is filtered out in code lattice, to obtain tentative prediction result；
According to the process (can be realized by routine techniques) of algorithm of target detection, for tentative prediction as a result, using non-pole
Big value restrainable algorithms, remove redundancy anchor frame, then select category score several prediction blocks in the top, as target detection knot
Fruit.
The present invention is described in detail below with reference to example；It is worth noting that, network type involved in following example,
The specific value etc. of threshold value is citing, is not construed as limiting.
In this example, for the structure for improving backbone network；In model training stage, mask figure is set；In model reasoning rank
Section combines three aspects of mask figure output test result to be introduced.
(1) structure of backbone network is improved:
1. the character network based on Darknet-53 network is improved, as shown in Fig. 2, selection is at 74 layers, 61 layers, 36 layers
Pond layer is added on the layer of three different scales, to find object using the feature on different scale.With input
For 512 × 512, the resolution ratio on these three scales is respectively 16 × 16,32 × 32,64 × 64.Then, we
Connecting window respectively for these three layers is 4, step-length 4, window 3, step-length 2, window 2, the maximum pond that step-length is 1
Layer, is all reduced to 16 × 16 for the resolution ratio of their characteristic patterns.Then, we again carry out these characteristic patterns on dimension direction
Splicing, it is lower to merge into a resolution ratio, but dimension is wider, the characteristic pattern rich in semantic information；
It should be noted that the Darknet-53 network that the present invention uses is only one of many backbone networks,
It can be carried out using other backbone networks, such as ResNet-101, VGG-16 according to the above-mentioned improvement thought of the present invention
Corresponding conversion.
2. 11 × 1 convolution sum 13 × 3 convolution is added after characteristic pattern, to reduce dimension described in 1.
Degree and expansion are semantic.Finally, the convolution using one 1 × 1 generates mask figure.For the structure of the network of script, do not carry out
Any change is only the branch for being added to a generation mask.
It should be noted that the convolutional layer added behind can suitably increase and can further lift scheme effect,
The present invention only discusses most simple way of realization；
3. continuing to add mask layer, having w × h grid in mask figure, each grid for the mask figure belonging in 2.
It is trained using logistic regression.So far, the modification of backbone network structure is completed.
(2) mask figure is set in model training stage:
As shown in Fig. 2, the dimension of mask figure is 1, each pixel above is referred to as grid, the value generation of grid
Table is fallen in inside this grid either with or without the center of object.Therefore, for the true box in original image, it is closed according to size
It is the wide high and corresponding coordinate of true box in available and mask figure equal resolution after scaled down.Calculate this
The central point of box, and centered on put the grid distribution fallen in and be positive label, for not comprising the grid of object, for its point
With negative label.Each grid uses logistic regression to be trained.
Using up-sampling, the mask figure of lowest resolution can be applied in the detection layers of all resolution ratio.Such as 16 ×
16 mask figure is up-sampled to 32 × 32 mask figure, and the grid in 16 × 16 mask figure upper left corners will correspond to 32 × 32 and cover originally
2 × 2 grids in code figure, numerical value are filled into 32 × 32 mask figure according to the method for duplication.According to this simple up-sampling
Mode, mask label can be extended in big resolution ratio, it is hereby achieved that different on big resolution ratio mask figure cover
Code lattice attribute.
By up-sampling, mask figure is assigned to the detection layers under each resolution ratio.Next it introduces and how to utilize mask
Lattice instruct the training of anchor frame:
As shown in figure 3, for positive mask case, anchor frame still according to true box overlap ratio number (IoU) divided
Match；And for negative mask case, then directly ignore anchor frame on it, is not involved in training.So, the negative anchor frame of training is participated in
Quantity will greatly reduce, and the positive negative ratio of serious unbalance will greatly be mitigated.According to statistics, it is having a size of 512 × 512 with input
Example, the positive negative ratio in a batch can reach 1:4000 originally, and will be maintained at after having used object likelihood mask
1:30 or so greatly alleviates extremely unbalanced positive negative ratio.
(3) mask figure is combined in the model reasoning stage
When model forward inference is used for test image, the grid that score is too small on mask figure is filtered out first.If covered
Code score is lower than 0.1 (first threshold value), then this grid is sifted out；Next, being screened in remaining mask case
Anchor frame.If this anchor frame is sifted out, is obtained preliminary less than 0.05 (second threshold value) by its object likelihood score
Prediction result.Next, for preliminary prediction result, being calculated using non-maxima suppression according to the process of algorithm of target detection
After method, sorts to it and select preceding 100 prediction blocks, obtain final testing result.
It should be noted that herein for screening mask threshold value and anchor frame threshold value can according to different data sets,
Leading to after too small amount of forward inference in test phase can determine.
By comparative experiments it is found that improved model is relative to former mould after alleviating imbalance using mask mechanism
Type and compare, higher score can be obtained on object likelihood；And it can show that higher recall rate, improve detection
Accuracy.
In order to make it easy to understand, doing necessary supplement below for the above process:
(1) foundation of basic network
Darknet-53 network is basic network used in YOLOv3, it includes 53 convolutional layers, a large amount of 3 × 3 and 1 ×
1 convolutional layer is used in network.Its accuracy rate on ImageNet data set is quite similar with ResNet-101, this says
The feature representation ability of the basic network of bright Darknet-53 and ResNet-101 is essentially identical.In the selection for carrying out backbone network
When, it can be foundation, the model foundation backbone network for selecting accuracy rate bigger, phase according to the accuracy rate on ImageNet data set
The testing result answered may also can be more preferable.
(2) training of model parameter
As shown in figure 3, being visualized to how object likelihood mask supervision figure is applied to detection layers in the training stage.
For positive mask case, the anchor frame being used only in the mask case is trained.It is described above using the value of mask case and is distributed
The positive and negative label of mask case, to filter out positive mask case, which can be understood as preliminary screening, and screening is cell.
After mask case screening, we only need to consider the anchor frame in remaining cell；It, can be in the following manner for anchor frame
Be allocated: when anchor frame is greater than threshold value with the overlapping ratio of corresponding object, we will distribute positive label, when overlapping than being less than threshold value,
We will distribute negative label.This threshold value generally may be configured as 0.5.
Further, it is also possible to be taken at random using the flip horizontal generallyd use in object detection method, colour switching is more
The data enhancement methods such as scale training help to be promoted the robustness of detector.
(3) model forward inference
The mask threshold value (i.e. first threshold value) that can be set is 0.1, and object likelihood threshold value (i.e. second threshold value) is
0.05 helps to filter out useless prediction block.Meanwhile we using 0.6 non-maxima suppression threshold value (i.e. NMS threshold value) come
Carry out the removal of redundancy frame.In implementation process, it can also help to mention using the skill of fusion convolutional layer and batch normalization layer
Lifting speed.
As specified above, three parameters for needing to adjust are introduced in the model of addition altogether, these parameters are only
It needs to be adjusted in test, and is not necessarily to be adjusted in training, be listed as follows:
| Parameter | Mask threshold value | Anchor frame threshold value | NMS threshold value |
| Default value | 0.1 | 0.05 | 0.6 |
1 parameter value of table
Corresponding threshold value can be step-length according to 0.05, obtain optimal group by a small number of forward inferences on verifying collection
It closes.Therefore, various introducing parameters involved in this implementation, can do according to the actual conditions of data set used with tune appropriate
It is whole.
Improved scheme provided by the invention can adapt to the more difficult target detection scenes such as intensive, wisp.In reality
In testing, about 8% or so mean accuracy can be promoted compared with the correlation technique of bias weight.
It is uneven can preferably to solve the positive and negative sample in depth targets detector for scheme provided in an embodiment of the present invention
The problem of, to promote the accuracy of algorithm of target detection.
Through the above description of the embodiments, those skilled in the art can be understood that above-described embodiment can
The mode of necessary general hardware platform can also be added to realize by software by software realization.Based on this understanding,
The technical solution of above-described embodiment can be embodied in the form of software products, which can store non-easy at one
In the property lost storage medium (can be CD-ROM, USB flash disk, mobile hard disk etc.), including some instructions are with so that a computer is set
Standby (can be personal computer, server or the network equipment etc.) executes method described in each embodiment of the present invention.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
Within the technical scope of the present disclosure, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with the protection model of claims
Subject to enclosing.
Claims (6)
1. a kind of object detection method for improving class imbalance characterized by comprising
The branched structure for generating mask figure is added in traditional characteristic detection network, obtains the characteristic pattern with given resolution
Mask figure with identical size；
Mask figure is trained, is the positive and negative label of each mask case distribution in mask figure according to training result, for just covering
Code lattice, are trained anchor frame thereon, for negative mask case, ignore anchor frame on it, be not involved in training, finally obtain instruction
The target detection model perfected；
In test phase, the screening of anchor frame is carried out using trained target detection model, and then obtains object detection results.
2. a kind of object detection method for improving class imbalance according to claim 1, which is characterized in that described to pass
The branched structure for generating mask figure is added in system feature detection network, obtains having with the characteristic pattern of given resolution identical
The mask figure of size includes:
The traditional characteristic detection network can extract the characteristic pattern of several different resolutions, by addition pond layer to each point
The character figure of resolution carries out down-sampling, and the characteristic pattern of each resolution ratio is narrowed down to the characteristic pattern phase with lowest resolution
Same size, then all characteristic patterns are spliced, obtain the characteristic pattern A comprising different scale object and semantic information；
Make size change over w × h × 1 of characteristic pattern A by adding convolutional layer, wherein w and h is respectively the minimum feature of resolution ratio
The width of figure and height；
Mask layer is added, regards the characteristic pattern A after size change over as mask figure.
3. a kind of object detection method for improving class imbalance according to claim 1, which is characterized in that described pair is covered
Code figure is trained, and is distributed positive and negative label according to training result for each mask case in mask figure and is included:
Using logistic regression mode training mask figure, the loss function of logistic regression is expressed as follows:
Lmask=∑ { y ln σ (x)-(1-y) ln [1- σ (x)] }
Wherein, y is the label of mask case, and x is the output of mask case, and σ (x) expression converts x using sigmoid function,
Make its output between 0 to 1, it is general inside corresponding mask case that the numerical value converted indicates that the center of the true box of object is fallen in
Rate；
The distribution principle of the numerical value and mask case label that obtain according to transformation is that each mask case distributes positive and negative label；Wherein, it covers
The distribution principle of code case marker label are as follows: if the center of the true box of object is fallen in mask case, distribute corresponding mask case and be positive and cover
Code lattice；If this object, completely except mask case, this distributes corresponding mask case and is negative mask case.
4. a kind of object detection method for improving class imbalance according to claim 1 or 3, which is characterized in that the party
Method further include: carrying out up-sampling operation to mask figure makes it extend to all detection layers of traditional characteristic detection network.
5. a kind of object detection method for improving class imbalance according to claim 1, which is characterized in that according to mask
The positive and negative tag types of lattice carry out the training of anchor frame, and for positive mask case, corresponding anchor frame is according to overlapping than carrying out with true box
Distribution；
The overall loss function of training anchor frame can be expressed as follows:
L=Lobj+Lmask+Lcls+αLloc
Wherein, LobjFor the loss of anchor frame object likelihood, LmaskFor the loss of mask figure, LclsFor Classification Loss, LlocFor boundary
The loss that frame returns, α is weight factor.
6. a kind of object detection method for improving class imbalance according to claim 1, which is characterized in that described to survey
In the examination stage, the screening of anchor frame is carried out using trained target detection model, and then obtain object detection results and include:
Two threshold values are set, and first threshold value is for screening mask case, and second threshold value is for screening anchor frame；
Input picture is handled by trained target detection model, obtains corresponding mask figure；
By first threshold value, corresponding mask case is filtered out from mask figure；By second threshold value from the mask case of screening
In filter out corresponding anchor frame, to obtain tentative prediction result；
According to the process of algorithm of target detection, for tentative prediction as a result, using non-maxima suppression algorithm, redundancy anchor is removed
Frame, then category score several prediction blocks in the top are selected, as object detection results.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910494701.8A CN110210482B (en) | 2019-06-05 | 2019-06-05 | Target detection method for improving class imbalance |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910494701.8A CN110210482B (en) | 2019-06-05 | 2019-06-05 | Target detection method for improving class imbalance |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110210482A true CN110210482A (en) | 2019-09-06 |
| CN110210482B CN110210482B (en) | 2022-09-06 |
Family
ID=67791559
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910494701.8A Active CN110210482B (en) | 2019-06-05 | 2019-06-05 | Target detection method for improving class imbalance |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110210482B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111209888A (en) * | 2020-01-15 | 2020-05-29 | 武夷学院 | Human-computer interface visual recognition system and method |
| CN111563502A (en) * | 2020-05-09 | 2020-08-21 | 腾讯科技（深圳）有限公司 | Image text recognition method and device, electronic equipment and computer storage medium |
| CN111914727A (en) * | 2020-07-28 | 2020-11-10 | 联芯智能(南京)科技有限公司 | Small target human body detection method based on balance sampling and nonlinear feature fusion |
| CN115187950A (en) * | 2022-09-13 | 2022-10-14 | 安徽中科星驰自动驾驶技术有限责任公司 | Novel balance mask secondary sampling method for deep learning training data enhancement |
| CN116563665A (en) * | 2023-04-25 | 2023-08-08 | 北京百度网讯科技有限公司 | Training method of target detection model, target detection method, device and equipment |
| CN113496139B (en) * | 2020-03-18 | 2024-02-13 | 北京京东乾石科技有限公司 | Method and apparatus for detecting objects from images and training object detection models |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107169421A (en) * | 2017-04-20 | 2017-09-15 | 华南理工大学 | A kind of car steering scene objects detection method based on depth convolutional neural networks |
| US20180268292A1 (en) * | 2017-03-17 | 2018-09-20 | Nec Laboratories America, Inc. | Learning efficient object detection models with knowledge distillation |
| CN108875732A (en) * | 2018-01-11 | 2018-11-23 | 北京旷视科技有限公司 | Model training and example dividing method, device and system and storage medium |
| CN109784349A (en) * | 2018-12-25 | 2019-05-21 | 东软集团股份有限公司 | Image object detection model method for building up, device, storage medium and program product |
-
2019
- 2019-06-05 CN CN201910494701.8A patent/CN110210482B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180268292A1 (en) * | 2017-03-17 | 2018-09-20 | Nec Laboratories America, Inc. | Learning efficient object detection models with knowledge distillation |
| CN107169421A (en) * | 2017-04-20 | 2017-09-15 | 华南理工大学 | A kind of car steering scene objects detection method based on depth convolutional neural networks |
| CN108875732A (en) * | 2018-01-11 | 2018-11-23 | 北京旷视科技有限公司 | Model training and example dividing method, device and system and storage medium |
| CN109784349A (en) * | 2018-12-25 | 2019-05-21 | 东软集团股份有限公司 | Image object detection model method for building up, device, storage medium and program product |
Non-Patent Citations (4)
| Title |
|---|
| FEIMO LI ET AL: "Class-imbalance aware CNN extension for high resolution aerial image based vehicle localization and categorization", 《2017 2ND INTERNATIONAL CONFERENCE ON IMAGE, VISION AND COMPUTING (ICIVC)》 * |
| MATEUSZBUDA ET AL: "A systematic study of the class imbalance problem in convolutional neural networks", 《NEURAL NETWORKS》 * |
| 吴止锾等: "类别非均衡遥感图像语义分割的全卷积网络方法", 《光学学报》 * |
| 姚群力等: "深度卷积神经网络在目标检测中的研究进展", 《计算机工程与应用》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111209888A (en) * | 2020-01-15 | 2020-05-29 | 武夷学院 | Human-computer interface visual recognition system and method |
| CN113496139B (en) * | 2020-03-18 | 2024-02-13 | 北京京东乾石科技有限公司 | Method and apparatus for detecting objects from images and training object detection models |
| CN111563502A (en) * | 2020-05-09 | 2020-08-21 | 腾讯科技（深圳）有限公司 | Image text recognition method and device, electronic equipment and computer storage medium |
| CN111563502B (en) * | 2020-05-09 | 2023-12-15 | 腾讯科技（深圳）有限公司 | Image text recognition method and device, electronic equipment and computer storage medium |
| CN111914727A (en) * | 2020-07-28 | 2020-11-10 | 联芯智能(南京)科技有限公司 | Small target human body detection method based on balance sampling and nonlinear feature fusion |
| CN111914727B (en) * | 2020-07-28 | 2024-04-26 | 联芯智能(南京)科技有限公司 | Small target human body detection method based on balance sampling and nonlinear feature fusion |
| CN115187950A (en) * | 2022-09-13 | 2022-10-14 | 安徽中科星驰自动驾驶技术有限责任公司 | Novel balance mask secondary sampling method for deep learning training data enhancement |
| CN115187950B (en) * | 2022-09-13 | 2022-11-22 | 安徽中科星驰自动驾驶技术有限责任公司 | Novel balance mask secondary sampling method for deep learning image data enhancement |
| CN116563665A (en) * | 2023-04-25 | 2023-08-08 | 北京百度网讯科技有限公司 | Training method of target detection model, target detection method, device and equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110210482B (en) | 2022-09-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110210482A (en) | Improve the object detection method of class imbalance | |
| CN110503112B (en) | Small target detection and identification method for enhancing feature learning | |
| CN110287932B (en) | Road blocking information extraction method based on deep learning image semantic segmentation | |
| CN108416307B (en) | Method, device and equipment for detecting pavement cracks of aerial images | |
| CN109816012A (en) | A kind of multiscale target detection method of integrating context information | |
| CN109859190A (en) | A kind of target area detection method based on deep learning | |
| CN109711474A (en) | A kind of aluminium material surface defects detection algorithm based on deep learning | |
| CN113076871B (en) | Fish shoal automatic detection method based on target shielding compensation | |
| CN109711288A (en) | Remote sensing ship detecting method based on feature pyramid and distance restraint FCN | |
| CN108345911A (en) | Surface Defects in Steel Plate detection method based on convolutional neural networks multi-stage characteristics | |
| CN111079739B (en) | Multi-scale attention feature detection method | |
| CN106845383A (en) | People's head inspecting method and device | |
| CN109325504A (en) | A kind of underwater sea cucumber recognition methods and system | |
| CN109241871A (en) | A kind of public domain stream of people's tracking based on video data | |
| CN106846344A (en) | A kind of image segmentation optimal identification method based on the complete degree in edge | |
| CN112767357A (en) | Yolov 4-based concrete structure disease detection method | |
| CN111914795A (en) | Method for detecting rotating target in aerial image | |
| CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
| CN109671071A (en) | A kind of underground piping defect location and grade determination method based on deep learning | |
| CN115222946B (en) | Single-stage instance image segmentation method and device and computer equipment | |
| CN110458201A (en) | A kind of remote sensing image object-oriented classification method and sorter | |
| CN107944403A (en) | Pedestrian's attribute detection method and device in a kind of image | |
| Zheng et al. | A review of remote sensing image object detection algorithms based on deep learning | |
| CN114565824A (en) | Single-stage rotating ship detection method based on full convolution network | |
| CN116503389B (en) | Automatic detection method for external absorption of tooth root |
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 |