CN109800778A - A kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible - Google Patents
A kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible Download PDFInfo
- Publication number
- CN109800778A CN109800778A CN201811463226.XA CN201811463226A CN109800778A CN 109800778 A CN109800778 A CN 109800778A CN 201811463226 A CN201811463226 A CN 201811463226A CN 109800778 A CN109800778 A CN 109800778A
- Authority
- CN
- China
- Prior art keywords
- sample
- frame
- threshold value
- iou
- loss
- 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
Landscapes
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
A kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible, it include: step 1, image object detection based on deep learning: step 2, based on online sample method for digging, the key parameter of use adjusts setting are as follows: step 3, negative difficulty divide sample excavate be hardly possible divide on sample basis by adjusting in training RPN formed the positive and negative sample proportion of mini-batch be 1:3, be trained；Step 4, redundancy frame is rejected, the multiple calculating of loss is avoided.The Suggestion box that RPN layer network generates reasonably is removed into redundancy using improved non-maxima suppression algorithm.The present invention relaxes the definition of negative sample in the case where not exptended sample, goes out more difficult training samples by sample online mining itself；Provided with positive and negative sample proportion, the largest loss, difficult sample train, rare rationally and are easily calculated；The loss returned to classification, frame has carried out equilibrating processing, can satisfy the lasting reduction of training loss.
Description
Technical field
Sample is divided to excavate skills such as (HNEM) based on online sample digging technology (OHEM) and negative difficulty the present invention relates to a kind of
The method that art combines.
Technical background
In recent years with the rapid development of computer science and technology, image procossing, image object based on computer technology
Detection etc. also obtains unprecedented fast development, and wherein deep learning is extracted crucial by the digital picture feature of study magnanimity
Target signature has been more than the mankind in target detection, is brought to industry one and another pleasantly surprised.The two of computer vision
Big task is image classification and target location detection, and the Image Feature Detection of conventional method is by exper ienced algorithm
Engineers design corresponds to matching template of target, such as deformable model (DPM), HOG feature extraction etc., the former target inspection
Survey is to be positioned by the window of sliding to target, is then classified using the template matching of fixed character to it.It is deposited
The problem of be that time-consuming for detection, characteristic matching precision is low, to specific target effective, leads to model generalization ability weak equal one
Series of problems.And the deep learning method risen recently can be directed to the target of complex scene, carrying out effectively feature identifies,
Recognition effect is much better than conventional method, but there is also shortcoming: (1) training set of the huge deep learning of data set amount needs
Want thousands of a samples, can effective learning data feature, and data are more, and effect is better, gives data collection task band
Carry out certain difficulty；(2) hardware requirement is higher, training big data sample, need the video memory of at least four G, to hardware propose compared with
High request；(3) training skills is strong, and when parameter setting is unreasonable, sample training effect is poor, the situation of difficult training occurs.
Summary of the invention
In order to overcome the above-mentioned deficiency of the prior art, the present invention provides one kind for sample problem and divides sample to dig based on difficulty
The Faster RCNN object detection method of pick, the present invention are a kind of online sample digging technologies when not increasing sample
Divide sample to excavate the technology combined with negative difficulty, model is enable to divide sample for existing difficulty, targetedly learns its spy
Sign, achievees the effect that model further promotes generalization, robustness.
To achieve the above object, the invention adopts the following technical scheme:
A kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible, includes the following steps:
Step 1, the image object detection based on deep learning；
Currently, the image object detection model based on deep learning is mostly based on convolutional neural networks, so master of the present invention
It to be analyzed based on Faster RCNN, and propose a kind of reasonable improved method.
Faster RCNN uses Softmax Loss and Smooth L1 Loss and returns joint to class probability and frame
Training, loss function formula are as follows:
Wherein, Ncls=256 indicate the number of prospect, and frame returns number Nreg=2400 be last characteristic pattern sliding
Maximum number of dimensions 40 × 60；I indicates some Suggestion box；piExpression predicts the probability of corresponding classification, including prospect, back
Scape；The value of expression prospect and background remembers prospectBackgroundThe loss of target frame recurrence is calculated with this；
tiIndicate the coordinate information of the Suggestion box, i.e. center point coordinate and the suggestion high t of frame widthi=(xi,yi,wi,hi), andIndicate true
Just corresponding information on target object, similarlyAnd λ balances the power of frame recurrence and Classification Loss
Weight；LclsFor softmax Classification Loss function,Frame returns loss using smooth L1's
Method.
As above method training to a certain extent, can be generated to the difficult sample data deficiency trained, shortage trains it etc.
Problem.But sample excavation is able to solve difficulty and divides sample training problem.So being carried out to the method in step 1 following several
The improvement of step.
Step 2 is based on online sample method for digging, and the key parameter that the present invention uses adjusts setting are as follows:
(1) setting hardly possible divides the Filtering system of sample.The difficulty of each iteration screening divides sample to be according to current total losses L
({pi},{ti) descending arrangement, and preceding B/N sample size is screened, wherein B=64, N are picture number N=trained every time
1, backpropagation speed trained so faster, because only that the gradient for needing to adjust on a small quantity.
(2) arithmetic speed is improved.In back transfer, the difficulty of screening divide sample to be got by forward direction costing bio disturbance, is passed through
It blames a point sample losses and sets 0 operation, not can be reduced model video memory amount.So back transfer filters out B/N difficulty and sample is divided to carry out
Gradient transmitting, trained video memory capacity are reduced to 3057M from 3527M.
(3) at RPN layers and finally, full articulamentum increases OHEM module simultaneously.Since good classification results depend on accurately
Target positioning, i.e. the generation of Suggestion box, thus to RPN layer also increase OHEM module, be conducive to frame return out most accurately
Position, even more improve the validity of the feature extraction of classification.
(4) adjustment Classification Loss and frame appropriate return the weight of loss.In loss function formula, Classification Loss
LclsLoss L is returned with frameregIt is unbalanced, Ncls=256 be the number of classification, and Nreg=2400 be last characteristic pattern
The maximum dimension of sliding, inverted about 10 times of the difference of the two, therefore λ=10 are taken, be conducive to returning the same of preferable frame in this way
When, targetedly learn corresponding target signature.
(5) non-maxima suppression (NMS) algorithm is adjusted.Classical non-maxima suppression algorithm cannot retain adjacent well
Or the multi-target detection frame of overlapping, cause the recall rate of target detection to reduce.Therefore, the improved non-maxima suppression of the present invention
Algorithm uses the confidence level penalty mechanism of classification, has also been effectively maintained Suggestion box while removing the Suggestion box of redundancy,
And further improve mAP.Specific operation is talked about in step 3.
(6) data enhancing work improves the generalization ability of model.Pass through random left and right mirror image switch and illumination when training
Saturation degree is adjusted, and is increased sample diversity, is prevented over-fitting.To improve model to the detectability of different sized images, instructing
Multiple dimensioned training is used when practicing, by the random ruler of the short side setting { 224,416,480,512,600,672,900 } of image
It is very little, another side equal proportion scaling.The enhancing work of sample also further increases mAP.
(7) strategy of learning rate is had adjusted.40k is set by the drop point of learning rate, is equivalent to 8 epoch, and
Hereafter decline a learning rate again every 20k the number of iterations, be conducive to the global scope search capability for improving early period in this way,
It avoids falling into local minizing point；Meanwhile the smaller learning rate that the later period uses, the fining adjustment of minimum is carried out, favorably
In dropping many times for loss.
Step 3, it is to divide the mini- formed on sample basis by adjusting RPN in training in hardly possible that negative difficulty, which divides sample to excavate,
The positive and negative sample proportion of batch is 1:3, is trained.Because being found by many experiments, when ratio is 1:3, in the table of model
Existing power is all best, specific strategy are as follows:
(1) cancel the threshold restriction of negative sample.In original Faster R-CNN, the mechanism for being determined as background is basis
Friendship between the candidate Suggestion box that RPN is generated and true frame and size than IoU is set, when IoU ratio [0.1,
0.5) when, it is considered to be the Suggestion box of background.This ratio setting be disadvantageous in that have ignored it is below lower than 0.1 it is rare,
It loses larger, important difficulty and divides negative sample, these features could not be learnt well.Therefore, the present invention, which is arranged, determines background
[0,0.5) threshold value is.Threshold value provided with positive sample is the maximum one or IoU threshold value of IoU value in [0.7,1.0] model
It encloses.
(2) the ratio 1:3 of difficult point of positive negative sample of setting.As previously mentioned, positive negative sample combines, and learn more about
Background information can be improved the ability of model detection and localization target under specific background environment.And it is provided with each target
Picture batch size is 64, then the quantity of positive sample is 16, and the quantity of negative sample is 48.
(3) it uses non-maximum in view of the positive sample and label predicted is there are multiple the case where repeating simultaneously and presses down
Method processed is then deleted using 0.7 threshold value when the IoU of target and label is below 0.7.
To the negative difficulty in step 3 divide sample excavate parameter setting summarize it is as follows:
(1) parameter name: FG_THRESH；The meaning of representative: positive sample IoU threshold value；Parameter value: [0.7,1.0]；
(2) parameter name: BG_THRESH_LO；The meaning of representative: negative sample IoU threshold value；Parameter value: [0,0.5)；
(3) parameter name: HNEM_NMS_THRESH；The meaning of representative: non-maxima suppression threshold value；Parameter value:
0.7；
(4) parameter name: HNEM_BATCHSIZE；The meaning of representative: picture object lot size；Parameter value: 64；
(5) parameter name: RPN_FG_FRACTION；The meaning of representative: positive sample ratio；Parameter value: 0.25；
(6) parameter name: RPN_BG_FRACTION；The meaning of representative: negative sample ratio；Parameter value: 0.75.
Step 4 rejects redundancy frame, avoids the multiple calculating of loss.Using improved non-maxima suppression algorithm
The Suggestion box that RPN layer network generates reasonably is removed into redundancy.Specific operation are as follows:
It avoids pushily deleting the Suggestion box that IoU is greater than threshold value, the improved procedure taken is to reduce its confidence level.It takes
The linear weighting of calculation method, Gauss weighting method, exponential weighted method.
Wherein, the calculation method of linear weighted function is that IoU is greater than to its confidence level of the reduction of threshold value, that is, introduces the think of of penalty function
Think:
In formula, siIndicate that the confidence score of current generic, a indicate weight coefficient, 0 < a≤1, bmIndicate confidence level
Frame corresponding to highest scoring, biIndicate current frame, IoU (bm,bi) indicate both friendship and ratio, NtExpression gives
Threshold value.
And use the calculation formula of Gauss weighting method are as follows:
In formula, the punishment dynamics of the value Different Effects penalty function of σ, gaussian weighing function has smooth compared to linear weighted function
The features such as transition.
Exponential weighted method equally uses certain threshold value, when its IoU is greater than the threshold value, punishes it.It compares
Weigthed sums approach has the characteristics that smooth transition at threshold value；Compared to Gauss weighting method, can retain in threshold value previous stage
Compared with multiple weighing value.Calculation formula:
In formula, NtIt is same to indicate IoU threshold value.
By multiple experiment, for each improved non-maxima suppression algorithm, corresponding detailed process and respectively
Adaptation situation be summarized as follows:
(1) linear method of weighting screens the highest frame of confidence level in certain one kind as most by confidence level descending
Excellent frame, the IoU and given threshold N of more next framet, when being less than the threshold value, confidence level score value is constant；Otherwise it sets
Confidence threshold is reduced to the ratio of a (1-IoU), i.e. IoU is bigger, and punishment dynamics are bigger.Circulate operation, until all to deckle
The confidence level of frame is less than given confidence threshold value threshold, then gives up, log history optimal value.This method is suitable for weight
Multiple frame is more, time complexity is low, the scene of quick the selection result, that is, the dynamics punished is larger, and screening finishes quickly, but
Obtained effect is not too much ideal.
(2) it is directed to Gauss weighting method, equally filters out optimal confidence level frame, but without setting NtThreshold value, pass through finger
Number functions successively decrease, i.e., IoU is bigger, confidence level score value successively decrease it is bigger, and set σ value control the degree successively decreased.Recycle ratio
Compared with confidence score rejects these redundancy frames when being less than set confidence threshold value threshold；Otherwise it is high to retain confidence level
Optimal frame.The frame quantity that this method is suitable for predicting to generate is medium, time requirement is not high, the accurate feelings of quantity statistics
Condition, and the screening seamlessly transitted helps to obtain preferably recurrence frame and obtains although time complexity is relatively linear poor
Result it is more excellent.
(3) it is directed to exponential weighted method, the IoU of same relatively suboptimum confidence level frame and optimal frame sets certain threshold
Value Nt.When being less than the threshold value, confidence level score value is constant；Otherwise confidence threshold value is reduced toThe index is presented
The effect successively decreased.It is finished until all frames all detect, retains the Suggestion box that confidence level is greater than threshold value.This method is suitable for protecting
It stays more recurrence frame, delete the scene for selecting partial redundance frame, can be applied to only to detect target whether there is or not be not required to the feelings of statistical magnitude
Condition, therefore effect is poorer than linear and Gauss weighting method.
So can choose different calculation methods with applicable situation according to different conditions to reduce its confidence level.
The invention has the advantages that
Excavated based on online sample divides sample to excavate the technology combined with negative difficulty, and most prominent feature is not expand sample
In the case where this, relax the definition of negative sample, more difficult training samples are gone out by sample online mining itself；Provided with just
Negative sample ratio rationally and easily calculates the largest loss, difficult sample train, rare；The damage that classification, frame are returned
Mistake has carried out equilibrating processing, can satisfy the lasting reduction of training loss.Secondly it is overlapped missing inspection problem for multiple target, led to
Improved non-maxima suppression is crossed, penalty function thought is introduced using three kinds of different modes and lowers its confidence level；Through Experimental comparison,
It obtains and uses Gauss weighting method best to modelling effect, improve recall rate, solve the problems, such as multiple target missing inspection.The improvement is calculated
Method may also extend into the context of detection of other field, such as product defects context of detection (chain is overlapped missing inspection problem), Hang Renjian
Survey problem, transport truck statistical problem etc..
Detailed description of the invention
Fig. 1 is the process that different size Anchor of the invention generate Suggestion box；
Fig. 2 is that the difficulty of screening of the invention divides sample back transfer process；
Fig. 3 is the OHEM module added of the invention；
Fig. 4 is improved non-maxima suppression algorithm flow of the invention；
Fig. 5 a~Fig. 5 f is the loss convergence curve and learning rate image that NMS Gauss of the invention improves front and back, wherein
Fig. 5 a is RPN layers of Classification Loss, and Fig. 5 b is that RPN layers of frame return loss, and Fig. 5 c is full articulamentum Classification Loss, and Fig. 5 d is complete
Articulamentum frame returns loss, and Fig. 5 e is total losses, and Fig. 5 f is the learning rate strategy of adjustment；
Fig. 6 is improved Gauss NMS detection effect of the invention.
Specific embodiment
The present invention provides a kind of online sample digging technology when not increasing sample and negative for sample problem
Difficulty divides sample to excavate the technology combined, so that model is divided sample for existing difficulty, targetedly learns its feature,
Achieve the effect that model further promotes generalization, robustness.
To achieve the above object, the invention adopts the following technical scheme:
A kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible, includes the following steps:
Step 1, the image object detection based on deep learning；
Currently, the image object detection model based on deep learning is mostly based on convolutional neural networks, so master of the present invention
It to be analyzed based on Faster RCNN, and propose a kind of reasonable improved method.
Faster RCNN uses Softmax Loss and Smooth L1 Loss and returns joint to class probability and frame
Training, loss function formula are as follows:
Wherein, Ncls=256 indicate the number of prospect, and frame returns number Nreg=2400 be last characteristic pattern sliding
Maximum number of dimensions 40 × 60；I indicates some Suggestion box；piExpression predicts the probability of corresponding classification, including prospect, back
Scape；The value of expression prospect and background remembers prospectBackgroundThe loss of target frame recurrence is calculated with this；
tiIndicate the coordinate information of the Suggestion box, i.e. center point coordinate and the suggestion high t of frame widthi=(xi,yi,wi,hi), andIndicate true
Just corresponding information on target object, similarlyAnd λ balances the power of frame recurrence and Classification Loss
Weight；LclsFor softmax Classification Loss function,Frame returns loss using smooth L1's
Method.
As above method training to a certain extent, can be generated to the difficult sample data deficiency trained, shortage trains it etc.
Problem.But sample excavation is able to solve difficulty and divides sample training problem.So being carried out to the method in step 1 following several
The improvement of step.
Step 2 is based on online sample method for digging, and the key parameter that the present invention uses adjusts setting are as follows:
(1) setting hardly possible divides the Filtering system of sample.The difficulty of each iteration screening divides sample to be according to current total losses L
({pi},{ti) descending arrangement, and preceding B/N sample size is screened, wherein B=64, N are picture number N=trained every time
1, backpropagation speed trained so faster, because only that the gradient for needing to adjust on a small quantity.
(2) arithmetic speed is improved.In back transfer, the difficulty of screening divide sample to be got by forward direction costing bio disturbance, is passed through
It blames a point sample losses and sets 0 operation, not can be reduced model video memory amount.So back transfer filters out B/N difficulty and sample is divided to carry out
Gradient transmitting, trained video memory capacity are reduced to 3057M from 3527M.
(3) at RPN layers and finally, full articulamentum increases OHEM module simultaneously.Since good classification results depend on accurately
Target positioning, i.e. the generation of Suggestion box, thus to RPN layer also increase OHEM module, be conducive to frame return out most accurately
Position, even more improve the validity of the feature extraction of classification.
(4) adjustment Classification Loss and frame appropriate return the weight of loss.In loss function formula, Classification Loss
LclsLoss L is returned with frameregIt is unbalanced, Ncls=256 be the number of classification, and Nreg=2400 be last characteristic pattern
The maximum dimension of sliding, inverted about 10 times of the difference of the two, therefore λ=10 are taken, be conducive to returning the same of preferable frame in this way
When, targetedly learn corresponding target signature.
(5) non-maxima suppression (NMS) algorithm is adjusted.Classical non-maxima suppression algorithm cannot retain adjacent well
Or the multi-target detection frame of overlapping, cause the recall rate of target detection to reduce.Therefore, the improved non-maxima suppression of the present invention
Algorithm uses the confidence level penalty mechanism of classification, has also been effectively maintained Suggestion box while removing the Suggestion box of redundancy,
And further improve mAP.Specific operation is talked about in step 3.
(6) data enhancing work improves the generalization ability of model.Pass through random left and right mirror image switch and illumination when training
Saturation degree is adjusted, and is increased sample diversity, is prevented over-fitting.To improve model to the detectability of different sized images, instructing
Multiple dimensioned training is used when practicing, by the random ruler of the short side setting { 224,416,480,512,600,672,900 } of image
It is very little, another side equal proportion scaling.The enhancing work of sample also further increases mAP.
(7) strategy of learning rate is had adjusted.40k is set by the drop point of learning rate, is equivalent to 8 epoch, and
Hereafter decline a learning rate again every 20k the number of iterations, be conducive to the global scope search capability for improving early period in this way,
It avoids falling into local minizing point；Meanwhile the smaller learning rate that the later period uses, the fining adjustment of minimum is carried out, favorably
In dropping many times for loss.
Step 3, it is to divide the mini- formed on sample basis by adjusting RPN in training in hardly possible that negative difficulty, which divides sample to excavate,
The positive and negative sample proportion of batch is 1:3, is trained.Because being found by many experiments, when ratio is 1:3, in the table of model
Existing power is all best, specific strategy are as follows:
(1) cancel the threshold restriction of negative sample.In original Faster R-CNN, the mechanism for being determined as background is basis
Friendship between the candidate Suggestion box that RPN is generated and true frame and size than IoU is set, when IoU ratio [0.1,
0.5) when, it is considered to be the Suggestion box of background.This ratio setting be disadvantageous in that have ignored it is below lower than 0.1 it is rare,
It loses larger, important difficulty and divides negative sample, these features could not be learnt well.Therefore, the present invention, which is arranged, determines background
[0,0.5) threshold value is.Threshold value provided with positive sample is the maximum one or IoU threshold value of IoU value in [0.7,1.0] model
It encloses.
(2) the ratio 1:3 of difficult point of positive negative sample of setting.As previously mentioned, positive negative sample combines, and learn more about
Background information can be improved the ability of model detection and localization target under specific background environment.And it is provided with each target
Picture batch size is 64, then the quantity of positive sample is 16, and the quantity of negative sample is 48.
(3) it uses non-maximum in view of the positive sample and label predicted is there are multiple the case where repeating simultaneously and presses down
Method processed is then deleted using 0.7 threshold value when the IoU of target and label is below 0.7.
To the negative difficulty in step 3 divide sample excavate parameter setting summarize it is as follows:
(7) parameter name: FG_THRESH；The meaning of representative: positive sample IoU threshold value；Parameter value: [0.7,1.0]；
(8) parameter name: BG_THRESH_LO；The meaning of representative: negative sample IoU threshold value；Parameter value: [0,0.5)；
(9) parameter name: HNEM_NMS_THRESH；The meaning of representative: non-maxima suppression threshold value；Parameter value:
0.7；
(10) parameter name: HNEM_BATCHSIZE；The meaning of representative: picture object lot size；Parameter value: 64；
(11) parameter name: RPN_FG_FRACTION；The meaning of representative: positive sample ratio；Parameter value: 0.25；
(12) parameter name: RPN_BG_FRACTION；The meaning of representative: negative sample ratio；Parameter value: 0.75.
Step 4 rejects redundancy frame, avoids the multiple calculating of loss.Using improved non-maxima suppression algorithm
The Suggestion box that RPN layer network generates reasonably is removed into redundancy.Specific operation are as follows:
It avoids pushily deleting the Suggestion box that IoU is greater than threshold value, the improved procedure taken is to reduce its confidence level.It takes
The linear weighting of calculation method, Gauss weighting method, exponential weighted method.
Wherein, the calculation method of linear weighted function is that IoU is greater than to its confidence level of the reduction of threshold value, that is, introduces the think of of penalty function
Think:
In formula, siIndicate that the confidence score of current generic, a indicate weight coefficient, 0 < a≤1, bmIndicate confidence level
Frame corresponding to highest scoring, biIndicate current frame, IoU (bm,bi) indicate both friendship and ratio, NtExpression gives
Threshold value.
And use the calculation formula of Gauss weighting method are as follows:
In formula, the punishment dynamics of the value Different Effects penalty function of σ, gaussian weighing function has smooth compared to linear weighted function
The features such as transition.
Exponential weighted method equally uses certain threshold value, when its IoU is greater than the threshold value, punishes it.It compares
Weigthed sums approach has the characteristics that smooth transition at threshold value；Compared to Gauss weighting method, can retain in threshold value previous stage
Compared with multiple weighing value.Calculation formula:
In formula, NtIt is same to indicate IoU threshold value.
By multiple experiment, for each improved non-maxima suppression algorithm, corresponding detailed process and respectively
Adaptation situation be summarized as follows:
(1) linear method of weighting screens the highest frame of confidence level in certain one kind as most by confidence level descending
Excellent frame, the IoU and given threshold N of more next framet, when being less than the threshold value, confidence level score value is constant；Otherwise it sets
Confidence threshold is reduced to the ratio of a (1-IoU), i.e. IoU is bigger, and punishment dynamics are bigger.Circulate operation, until all to deckle
The confidence level of frame is less than given confidence threshold value threshold, then gives up, log history optimal value.This method is suitable for weight
Multiple frame is more, time complexity is low, the scene of quick the selection result, that is, the dynamics punished is larger, and screening finishes quickly, but
Obtained effect is not too much ideal.
(2) it is directed to Gauss weighting method, equally filters out optimal confidence level frame, but without setting NtThreshold value, pass through finger
Number functions successively decrease, i.e., IoU is bigger, confidence level score value successively decrease it is bigger, and set σ value control the degree successively decreased.Recycle ratio
Compared with confidence score rejects these redundancy frames when being less than set confidence threshold value threshold；Otherwise it is high to retain confidence level
Optimal frame.The frame quantity that this method is suitable for predicting to generate is medium, time requirement is not high, the accurate feelings of quantity statistics
Condition, and the screening seamlessly transitted helps to obtain preferably recurrence frame and obtains although time complexity is relatively linear poor
Result it is more excellent.
(3) it is directed to exponential weighted method, the IoU of same relatively suboptimum confidence level frame and optimal frame sets certain threshold
Value Nt.When being less than the threshold value, confidence level score value is constant；Otherwise confidence threshold value is reduced toThe index is presented
The effect successively decreased.It is finished until all frames all detect, retains the Suggestion box that confidence level is greater than threshold value.This method is suitable for protecting
It stays more recurrence frame, delete the scene for selecting partial redundance frame, can be applied to only to detect target whether there is or not be not required to the feelings of statistical magnitude
Condition, therefore effect is poorer than linear and Gauss weighting method.
So can choose different calculation methods with applicable situation according to different conditions to reduce its confidence level.
In order to confirm the feasibility of the method for the present invention, according to above parameter designing, tested as follows:
By being tested in the common data sets of VOC2007 and VOC2012, wherein having 20 classification+backgrounds etc.
In 21 classification problems, training set 5,000 multiple, test set 5,000 multiple.It compared taking the online sample under different condition to dig
The strategy combination that pick divides sample to excavate with negative difficulty, is finally analyzed as follows experimental result:
Using classical Faster RCNN object detection method as reference:
(1) it only adds OHEM module in last articulamentum entirely or only when adding OHEM module for RPN layers, effect does not all have
It gets a promotion；
(2) by RPN layers and finally, full articulamentum all adds OHEM module, and using the online sample of random positive and negative sample proportion
When this method for digging, loss weight=1:1 that frame returns, effect promoting；
(3) RPN layers are classified on the basis of (2): the loss weight that frame returns is set as 1:10, the i.e. side of λ=10
Method further improves 1.3% after two classes loss balance, improves 1.8% compared to original, it is seen that both balanced damage
Lose the training for being conducive to model；
(4) the random ratio of positive negative sample is removed, is set as the ratio of 1:3, i.e., negative difficulty divides sample digging technology, and λ
=10, effect improves 0.4% again, and other ratios are not obviously improved, and illustrates that negative difficulty divides sample to excavate in model training
It plays a role；
(5) use improved NMS linear weighted function algorithm, by many experiments select optimal parameter be set as a=1,
Nt=0.2, threshold=0.001；Further improve 1.3%；
(6) it is weighted using Gauss, by comparative analysis, σ=0.3, threshold=0.003, the target inspection of promotion is set
Survey the object that object is aggregated pattern mostly, such as dog, people, bird, sheep, plant, improve 3.1% respectively, 7.6%, 9.1%,
1.3%, 24.8%, cause whole mAP to promote 6.3%, promotes 4.4 percentage points compared to original；
(7) exponential weighted method is used, N is sett=0.1, threshold=0.0001 is promoted effect ratio (5) and (6) two
Kind method is small；
(8) learning rate strategy is had adjusted on the basis of best NMS Gauss improved method, loss can be carried out again
Decline (attached drawing 5), mAP improves 4.5 compared to original；
(9) training sample for using VOC2012, is tested on 07 test set, is equivalent to and is increased more training samples,
Obtain the promotion of 3.1 percentage points；
(10) it is tested on the basis of (9) using the method for Gauss NMS, mAP improves 1.8 points.It is restrained by loss
Tracing analysis, it can be found that learning rate strategy adjusted can make, frame is returned and the loss of classification constantly declines；
(11) because declining when iteration 30k times several primary originally according to the learning rate adjustable strategies in Fater RCNN
Learning rate, analysis finds that the strategy can not seek global minimum well in experiment.Therefore, it increases and searches early period
The drop point of learning rate is set 40k by the duration of rope, is conducive to avoid falling into local minizing point.
By the above analysis of experimental results it is found that using OHEM and the HNEM method combined and improved NMS and instruction
Practice strategy preferable model inspection effect available in the case where possessing original data set, improves 4.5 percentage points.Wherein, using changing
Into Gaussian function non-maxima suppression algorithm, on the basis of master mould promoted 1.8 percentage points, the recall rate of model can be made
It is significantly improved, there is preferable applicability to target object detection in groups.
The locating effect of improved NMS algorithm is as shown in Fig. 6, analyzes it is found that the innovatory algorithm was overlapped multiple target
The more traditional NMS algorithm of detection effect has biggish promotion in target detection, reduces the risk of target object missing inspection.So that
Recall rate index gets a promotion, and single P-R area under the curve (AP) more previously had the advantage further increased, leads to entirety
MAP obtain preferable result.
Content described in this specification embodiment is only enumerating to the way of realization of inventive concept, protection of the invention
Range should not be construed as being limited to the specific forms stated in the embodiments, and protection scope of the present invention is also and in this field skill
Art personnel conceive according to the present invention it is conceivable that equivalent technologies mean.
Claims (1)
1. a kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible, includes the following steps:
Step 1, the image object detection based on deep learning；
Analyzed based on Faster RCNN, Faster RCNN use Softmax Loss and Smooth L1 Loss to point
Class probability and frame return joint training, loss function formula are as follows:
Wherein, Ncls=256 indicate the number of prospect, and frame returns number Nreg=2400 be the maximum of last characteristic pattern sliding
Number of dimensions 40 × 60；I indicates some Suggestion box；piExpression predicts the probability of corresponding classification, including prospect, background；
The value of expression prospect and background remembers prospectBackgroundThe loss of target frame recurrence is calculated with this；tiIt indicates
The coordinate information of the Suggestion box, i.e. center point coordinate and the suggestion high t of frame widthi=(xi,yi,wi,hi), andIndicate really corresponding
Information on target object, similarlyAnd λ balances the weight of frame recurrence and Classification Loss；LclsFor
Softmax Classification Loss function,Frame returns the method that loss uses smooth L1；
Step 2, it is based on online sample method for digging, the key parameter of use adjusts setting are as follows:
(21) setting hardly possible divides the Filtering system of sample；The difficulty of each iteration screening divides sample to be according to current total losses L ({ pi},
{ti) descending arrangement, and preceding B/N sample size is screened, wherein B=64, N are picture number N=1 trained every time, in this way
Trained backpropagation speed faster, because only that the gradient for needing to adjust on a small quantity；
(22) arithmetic speed is improved；In back transfer, the difficulty of screening divide sample to be got by forward direction costing bio disturbance, by blaming
Point sample losses set 0 operation, not can be reduced model video memory amount；So back transfer filters out B/N difficulty and sample is divided to carry out gradient
Transmitting, trained video memory capacity are reduced to 3057M from 3527M；
(23) at RPN layers and finally, full articulamentum increases OHEM module simultaneously；Since good classification results depend on accurately
Target positioning, the i.e. generation of Suggestion box are conducive to frame and return out most accurate position so also increasing OHEM module to RPN layers
It sets, even more improves the validity of the feature extraction of classification；
(24) adjustment Classification Loss and frame appropriate return the weight of loss；In loss function formula, Classification Loss LclsWith
Frame returns loss LregIt is unbalanced, Ncls=256 be the number of classification, and Nreg=2400 be last characteristic pattern sliding
Maximum dimension, inverted about 10 times of the difference of the two, therefore λ=10 are taken, be conducive to have needle while returning preferable frame in this way
Learn corresponding target signature to property；
(25) non-maxima suppression (NMS) algorithm is adjusted；Classical non-maxima suppression algorithm cannot retain well it is adjacent or
The multi-target detection frame of overlapping causes the recall rate of target detection to reduce；Therefore, the improved non-maxima suppression algorithm of the present invention
Using the confidence level penalty mechanism of classification, it is also effectively maintained Suggestion box while removing the Suggestion box of redundancy, has been gone forward side by side
One step improves mAP；Specific operation is talked about in step 3；
(26) data enhancing work improves the generalization ability of model；It is saturated when training by random left and right mirror image switch and illumination
Degree is adjusted, and is increased sample diversity, is prevented over-fitting；To improve model to the detectability of different sized images, in training
Multiple dimensioned training has been used, it is another by the random size of the short side setting { 224,416,480,512,600,672,900 } of image
Side equal proportion scaling；The enhancing work of sample also further increases mAP；
(27) strategy of learning rate is had adjusted；40k is set by the drop point of learning rate, is equivalent to 8 epoch, and this
Decline a learning rate again every 20k the number of iterations afterwards, is conducive to the global scope search capability for improving early period in this way, avoids
Fall into local minizing point；Meanwhile the smaller learning rate that the later period uses, the fining adjustment of minimum is carried out, is conducive to damage
That loses drops many times；
Step 3, negative difficulty divide sample excavate be hardly possible divide on sample basis by adjusting in training RPN formed mini-batch just
Negative sample ratio is 1:3, is trained, specifically includes:
(31) cancel the threshold restriction of negative sample；In original Faster R-CNN, the mechanism for being determined as background is according to RPN
Friendship between the candidate Suggestion box and true frame of the generation and size than IoU is set, when IoU ratio [0.1,0.5)
When, it is considered to be the Suggestion box of background；This ratio setting, which is disadvantageous in that, has ignored rare, loss below lower than 0.1
Larger, important difficulty divides negative sample, these features could not be learnt well；Therefore, the threshold value for determining background is arranged in the present invention
For [0,0.5)；Threshold value provided with positive sample is the maximum one or IoU threshold value of IoU value in [0.7,1.0] range；
(32) the ratio 1:3 of difficult point of positive negative sample of setting；As previously mentioned, positive negative sample combines, and learn more about back
Scape information can be improved the ability of model detection and localization target under specific background environment；And it is provided with each target figure
Piece batch size is 64, then the quantity of positive sample is 16, and the quantity of negative sample is 48；
(33) non-maxima suppression side is used in view of the positive sample and label predicted is there are multiple the case where repeating simultaneously
Method is then deleted using 0.7 threshold value when the IoU of target and label is below 0.7；
Sample is divided to excavate being provided that for parameter the negative difficulty in step 3
The meaning that FG_THRESH is represented: positive sample IoU threshold value；Parameter value: [0.7,1.0]；
The meaning that BG_THRESH_LO is represented: negative sample IoU threshold value；Parameter value: [0,0.5)；
The meaning that HNEM_NMS_THRESH is represented: non-maxima suppression threshold value；Parameter value: 0.7；
The meaning that HNEM_BATCHSIZE is represented: picture object lot size；Parameter value: 64；
The meaning that RPN_FG_FRACTION is represented: positive sample ratio；Parameter value: 0.25；
The meaning that RPN_BG_FRACTION is represented: negative sample ratio；Parameter value: 0.75；
Step 4, redundancy frame is rejected, the multiple calculating of loss is avoided；Using improved non-maxima suppression algorithm by RPN
The Suggestion box that layer network generates reasonably removes redundancy；Specific operation are as follows:
It avoids pushily deleting the Suggestion box that IoU is greater than threshold value, the improved procedure taken is to reduce its confidence level；The calculating taken
The linear weighting of method, Gauss weighting method, exponential weighted method；
Wherein, the calculation method of linear weighted function is that IoU is greater than to its confidence level of the reduction of threshold value, that is, introduces the thought of penalty function:
In formula, siIndicate that the confidence score of current generic, a indicate weight coefficient, 0 < a≤1, bmIndicate confidence score
Frame corresponding to highest, biIndicate current frame, IoU (bm,bi) indicate both friendship and ratio, NtIndicate given threshold
Value；
And use the calculation formula of Gauss weighting method are as follows:
In formula, the punishment dynamics of the value Different Effects penalty function of σ, gaussian weighing function has smooth transition compared to linear weighted function
The features such as；
Exponential weighted method equally uses certain threshold value, when its IoU is greater than the threshold value, punishes it；Compared to linear
Weighting method has the characteristics that smooth transition at threshold value；Compared to Gauss weighting method, more power can be retained in threshold value previous stage
Value；Calculation formula:
In formula, NtIt is same to indicate IoU threshold value；
By multiple experiment, for each improved non-maxima suppression algorithm, corresponding detailed process and respective suitable
Situation is answered to be summarized as follows:
(41) linear method of weighting screens in certain one kind the highest frame of confidence level as optimal by confidence level descending
Frame, the IoU and given threshold N of more next framet, when being less than the threshold value, confidence level score value is constant；Otherwise confidence level
Threshold value is reduced to the ratio of a (1-IoU), i.e. IoU is bigger, and punishment dynamics are bigger；Circulate operation, until setting for all frames undetermined
Reliability is less than given confidence threshold value threshold, then gives up, log history optimal value；This method is suitable for repeating frame
It is more, time complexity is low, the scene of quick the selection result, that is, the dynamics punished is larger, screens the effect that finishes, but obtain quickly
Fruit is not too much ideal；
(42) it is directed to Gauss weighting method, equally filters out optimal confidence level frame, but without setting NtThreshold value, pass through index letter
Number successively decrease, i.e., IoU is bigger, confidence level score value successively decrease it is bigger, and set σ value control the degree successively decreased；Circulation compares, and sets
Confidence score rejects these redundancy frames when being less than set confidence threshold value threshold；Otherwise retain high optimal of confidence level
Frame；The situation that the frame quantity that this method is suitable for predicting to generate is medium, time requirement is not high, quantity statistics are accurate, and
The screening of smooth transition help to obtain it is preferable return frame, although time complexity is relatively linear poor, obtained result compared with
It is excellent；
(43) it is directed to exponential weighted method, the IoU of same relatively suboptimum confidence level frame and optimal frame sets certain threshold value Nt；
When being less than the threshold value, confidence level score value is constant；Otherwise confidence threshold value is reduced toIndex presentation is successively decreased
Effect；It is finished until all frames all detect, retains the Suggestion box that confidence level is greater than threshold value；This method is suitable for retaining more time
Return frame, delete the scene for selecting partial redundance frame, can be applied to need to only to detect target, whether there is or not, the case where being not required to statistical magnitude, therefore imitate
Fruit is poorer than linear and Gauss weighting method；
Different calculation methods is selected to reduce its confidence level with applicable situation according to different conditions.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811463226.XA CN109800778B (en) | 2018-12-03 | 2018-12-03 | Faster RCNN target detection method based on difficultly-divided sample mining |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811463226.XA CN109800778B (en) | 2018-12-03 | 2018-12-03 | Faster RCNN target detection method based on difficultly-divided sample mining |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109800778A true CN109800778A (en) | 2019-05-24 |
| CN109800778B CN109800778B (en) | 2020-10-09 |
Family
ID=66556406
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811463226.XA Active CN109800778B (en) | 2018-12-03 | 2018-12-03 | Faster RCNN target detection method based on difficultly-divided sample mining |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109800778B (en) |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110728316A (en) * | 2019-09-30 | 2020-01-24 | 广州海昇计算机科技有限公司 | Classroom behavior detection method, system, device and storage medium |
| CN110736709A (en) * | 2019-10-26 | 2020-01-31 | 苏州大学 | blueberry maturity nondestructive testing method based on deep convolutional neural network |
| CN110826634A (en) * | 2019-11-11 | 2020-02-21 | 北京百度网讯科技有限公司 | Training method and device of target detection model, electronic equipment and storage medium |
| CN110909660A (en) * | 2019-11-19 | 2020-03-24 | 佛山市南海区广工大数控装备协同创新研究院 | Plastic bottle detection and positioning method based on target detection |
| CN110969627A (en) * | 2019-11-29 | 2020-04-07 | 北京达佳互联信息技术有限公司 | Image processing method and device |
| CN111080601A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Method for identifying fault image of pull ring grinding shaft of derailment brake device of railway wagon |
| CN111091164A (en) * | 2020-03-25 | 2020-05-01 | 魔视智能科技(上海)有限公司 | Semantic segmentation model training method and system based on oem |
| CN111275011A (en) * | 2020-02-25 | 2020-06-12 | 北京百度网讯科技有限公司 | Mobile traffic light detection method and device, electronic equipment and storage medium |
| CN111291796A (en) * | 2020-01-21 | 2020-06-16 | 中国科学技术大学 | Sampling-free method used in target detector model training process |
| CN111597941A (en) * | 2020-05-08 | 2020-08-28 | 河海大学 | Target detection method for dam defect image |
| CN111666839A (en) * | 2020-05-25 | 2020-09-15 | 东华大学 | Road pedestrian detection system based on improved Faster RCNN |
| CN111723764A (en) * | 2020-06-29 | 2020-09-29 | 中国农业大学 | Improved fast RCNN hydroponic vegetable seedling state detection method |
| CN111723852A (en) * | 2020-05-30 | 2020-09-29 | 杭州迪英加科技有限公司 | Robust training method for target detection network |
| CN111986252A (en) * | 2020-07-16 | 2020-11-24 | 浙江工业大学 | Method for accurately positioning candidate bounding box in target segmentation network |
| CN112200187A (en) * | 2020-10-16 | 2021-01-08 | 广州云从凯风科技有限公司 | Target detection method, device, machine readable medium and equipment |
| CN112232450A (en) * | 2020-12-15 | 2021-01-15 | 浙江啄云智能科技有限公司 | Multi-stage comprehensive difficult sample mining method and target detection method |
| WO2021027135A1 (en) * | 2019-08-15 | 2021-02-18 | 平安科技（深圳）有限公司 | Cell detection model training method and apparatus, computer device and storage medium |
| CN112464743A (en) * | 2020-11-09 | 2021-03-09 | 西北工业大学 | Small sample target detection method based on multi-scale feature weighting |
| CN112801182A (en) * | 2021-01-27 | 2021-05-14 | 安徽大学 | RGBT target tracking method based on difficult sample perception |
| CN113160175A (en) * | 2021-04-23 | 2021-07-23 | 杭州迪英加科技有限公司 | Tumor lymphatic vessel infiltration detection method based on cascade network |
| CN113160267A (en) * | 2021-05-21 | 2021-07-23 | 上海通途半导体科技有限公司 | Motion vector calculation method and device based on deep learning |
| CN113344086A (en) * | 2021-06-16 | 2021-09-03 | 深圳市商汤科技有限公司 | Man-machine loop method, device, system, electronic equipment and storage medium |
| CN113780277A (en) * | 2021-09-08 | 2021-12-10 | 浙江啄云智能科技有限公司 | Training method and device of target detection model, electronic equipment and storage medium |
| CN114998749A (en) * | 2022-07-28 | 2022-09-02 | 北京卫星信息工程研究所 | SAR data amplification method for target detection |
| CN116168259A (en) * | 2023-04-26 | 2023-05-26 | 厦门微图软件科技有限公司 | Automatic defect classification algorithm applied to OLED lighting system |
| CN117197592A (en) * | 2023-11-06 | 2023-12-08 | 英特灵达信息技术(深圳)有限公司 | Target detection model training method and device, electronic equipment and medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN106960195A (en) * | 2017-03-27 | 2017-07-18 | 深圳市丰巨泰科电子有限公司 | A kind of people counting method and device based on deep learning |
| 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 |
| CN108596053A (en) * | 2018-04-09 | 2018-09-28 | 华中科技大学 | A kind of vehicle checking method and system based on SSD and vehicle attitude classification |
-
2018
- 2018-12-03 CN CN201811463226.XA patent/CN109800778B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN106960195A (en) * | 2017-03-27 | 2017-07-18 | 深圳市丰巨泰科电子有限公司 | A kind of people counting method and device based on deep learning |
| 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 |
| CN108596053A (en) * | 2018-04-09 | 2018-09-28 | 华中科技大学 | A kind of vehicle checking method and system based on SSD and vehicle attitude classification |
Non-Patent Citations (2)
| Title |
|---|
| REN SHAOQING: ""Faster R-CNN Towards Real-Time Object Detection with Region Proposal Network"", 《IEEE TRANSACTION ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| 冯小雨: ""基于改进Faster R-CNN的空中目标检测"", 《光学学报》 * |
Cited By (39)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021027135A1 (en) * | 2019-08-15 | 2021-02-18 | 平安科技（深圳）有限公司 | Cell detection model training method and apparatus, computer device and storage medium |
| CN110728316A (en) * | 2019-09-30 | 2020-01-24 | 广州海昇计算机科技有限公司 | Classroom behavior detection method, system, device and storage medium |
| CN110736709A (en) * | 2019-10-26 | 2020-01-31 | 苏州大学 | blueberry maturity nondestructive testing method based on deep convolutional neural network |
| CN110826634A (en) * | 2019-11-11 | 2020-02-21 | 北京百度网讯科技有限公司 | Training method and device of target detection model, electronic equipment and storage medium |
| CN110909660A (en) * | 2019-11-19 | 2020-03-24 | 佛山市南海区广工大数控装备协同创新研究院 | Plastic bottle detection and positioning method based on target detection |
| CN110969627B (en) * | 2019-11-29 | 2024-03-29 | 北京达佳互联信息技术有限公司 | Image segmentation model training method, image processing method and device |
| CN110969627A (en) * | 2019-11-29 | 2020-04-07 | 北京达佳互联信息技术有限公司 | Image processing method and device |
| CN111080601A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Method for identifying fault image of pull ring grinding shaft of derailment brake device of railway wagon |
| CN111291796A (en) * | 2020-01-21 | 2020-06-16 | 中国科学技术大学 | Sampling-free method used in target detector model training process |
| CN111275011A (en) * | 2020-02-25 | 2020-06-12 | 北京百度网讯科技有限公司 | Mobile traffic light detection method and device, electronic equipment and storage medium |
| CN111275011B (en) * | 2020-02-25 | 2023-12-19 | 阿波罗智能技术(北京)有限公司 | Mobile traffic light detection method and device, electronic equipment and storage medium |
| CN111091164A (en) * | 2020-03-25 | 2020-05-01 | 魔视智能科技(上海)有限公司 | Semantic segmentation model training method and system based on oem |
| CN111597941A (en) * | 2020-05-08 | 2020-08-28 | 河海大学 | Target detection method for dam defect image |
| CN111597941B (en) * | 2020-05-08 | 2021-02-09 | 河海大学 | Target detection method for dam defect image |
| CN111666839A (en) * | 2020-05-25 | 2020-09-15 | 东华大学 | Road pedestrian detection system based on improved Faster RCNN |
| CN111723852A (en) * | 2020-05-30 | 2020-09-29 | 杭州迪英加科技有限公司 | Robust training method for target detection network |
| CN111723852B (en) * | 2020-05-30 | 2022-07-22 | 杭州迪英加科技有限公司 | Robust training method for target detection network |
| CN111723764A (en) * | 2020-06-29 | 2020-09-29 | 中国农业大学 | Improved fast RCNN hydroponic vegetable seedling state detection method |
| CN111986252B (en) * | 2020-07-16 | 2024-03-29 | 浙江工业大学 | Method for accurately positioning candidate bounding boxes in target segmentation network |
| CN111986252A (en) * | 2020-07-16 | 2020-11-24 | 浙江工业大学 | Method for accurately positioning candidate bounding box in target segmentation network |
| CN112200187A (en) * | 2020-10-16 | 2021-01-08 | 广州云从凯风科技有限公司 | Target detection method, device, machine readable medium and equipment |
| CN112464743A (en) * | 2020-11-09 | 2021-03-09 | 西北工业大学 | Small sample target detection method based on multi-scale feature weighting |
| CN112464743B (en) * | 2020-11-09 | 2023-06-02 | 西北工业大学 | Small sample target detection method based on multi-scale feature weighting |
| CN112232450A (en) * | 2020-12-15 | 2021-01-15 | 浙江啄云智能科技有限公司 | Multi-stage comprehensive difficult sample mining method and target detection method |
| CN112232450B (en) * | 2020-12-15 | 2021-03-16 | 浙江啄云智能科技有限公司 | Multi-stage comprehensive difficult sample mining method and target detection method |
| CN112801182A (en) * | 2021-01-27 | 2021-05-14 | 安徽大学 | RGBT target tracking method based on difficult sample perception |
| CN112801182B (en) * | 2021-01-27 | 2022-11-04 | 安徽大学 | RGBT target tracking method based on difficult sample perception |
| CN113160175A (en) * | 2021-04-23 | 2021-07-23 | 杭州迪英加科技有限公司 | Tumor lymphatic vessel infiltration detection method based on cascade network |
| CN113160267B (en) * | 2021-05-21 | 2024-04-16 | 上海通途半导体科技有限公司 | Motion vector calculation method and device based on deep learning |
| CN113160267A (en) * | 2021-05-21 | 2021-07-23 | 上海通途半导体科技有限公司 | Motion vector calculation method and device based on deep learning |
| CN113344086B (en) * | 2021-06-16 | 2022-07-01 | 深圳市商汤科技有限公司 | Man-machine loop method, device, system, electronic equipment and storage medium |
| CN113344086A (en) * | 2021-06-16 | 2021-09-03 | 深圳市商汤科技有限公司 | Man-machine loop method, device, system, electronic equipment and storage medium |
| CN113780277A (en) * | 2021-09-08 | 2021-12-10 | 浙江啄云智能科技有限公司 | Training method and device of target detection model, electronic equipment and storage medium |
| CN113780277B (en) * | 2021-09-08 | 2023-06-30 | 浙江啄云智能科技有限公司 | Training method and device of target detection model, electronic equipment and storage medium |
| CN114998749A (en) * | 2022-07-28 | 2022-09-02 | 北京卫星信息工程研究所 | SAR data amplification method for target detection |
| CN116168259B (en) * | 2023-04-26 | 2023-08-08 | 厦门微图软件科技有限公司 | Automatic defect classification method applied to OLED lighting system |
| CN116168259A (en) * | 2023-04-26 | 2023-05-26 | 厦门微图软件科技有限公司 | Automatic defect classification algorithm applied to OLED lighting system |
| CN117197592B (en) * | 2023-11-06 | 2024-03-01 | 英特灵达信息技术(深圳)有限公司 | Target detection model training method and device, electronic equipment and medium |
| CN117197592A (en) * | 2023-11-06 | 2023-12-08 | 英特灵达信息技术(深圳)有限公司 | Target detection model training method and device, electronic equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109800778B (en) | 2020-10-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109800778A (en) | A kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible | |
| CN109977943A (en) | A kind of images steganalysis method, system and storage medium based on YOLO | |
| CN109671102A (en) | A kind of composite type method for tracking target based on depth characteristic fusion convolutional neural networks | |
| CN109002845A (en) | Fine granularity image classification method based on depth convolutional neural networks | |
| CN109886998A (en) | Multi-object tracking method, device, computer installation and computer storage medium | |
| CN107944369A (en) | A kind of pedestrian detection method based on tandem zones generation network and enhancing random forest | |
| CN107229904A (en) | A kind of object detection and recognition method based on deep learning | |
| CN107609525A (en) | Remote Sensing Target detection method based on Pruning strategy structure convolutional neural networks | |
| CN107818302A (en) | Non-rigid multiple dimensioned object detecting method based on convolutional neural networks | |
| CN103473786B (en) | Gray level image segmentation method based on multi-objective fuzzy clustering | |
| CN113807188B (en) | Unmanned aerial vehicle target tracking method based on anchor frame matching and Siamese network | |
| CN110991362A (en) | Pedestrian detection model based on attention mechanism | |
| CN110427807A (en) | A kind of temporal events motion detection method | |
| CN110751121B (en) | Unsupervised radar signal sorting method based on clustering and SOFM | |
| CN111767847B (en) | Pedestrian multi-target tracking method integrating target detection and association | |
| CN107688786A (en) | A kind of method for detecting human face based on concatenated convolutional neutral net | |
| CN108460790A (en) | A kind of visual tracking method based on consistency fallout predictor model | |
| CN109886128A (en) | A kind of method for detecting human face under low resolution | |
| CN111833322B (en) | Garbage multi-target detection method based on improved YOLOv3 | |
| CN113763424B (en) | Real-time intelligent target detection method and system based on embedded platform | |
| CN111145145B (en) | Image surface defect detection method based on MobileNet | |
| CN110008899B (en) | Method for extracting and classifying candidate targets of visible light remote sensing image | |
| CN110298865A (en) | The space-based Celestial Background small point target tracking of cluster device is separated based on threshold value | |
| CN110033012A (en) | A kind of production method for tracking target based on channel characteristics weighted convolution neural network | |
| CN110032952A (en) | A kind of road boundary point detecting method based on deep learning |
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 |