CN109241982A - Object detection method based on depth layer convolutional neural networks - Google Patents
Object detection method based on depth layer convolutional neural networks Download PDFInfo
- Publication number
- CN109241982A CN109241982A CN201811035114.4A CN201811035114A CN109241982A CN 109241982 A CN109241982 A CN 109241982A CN 201811035114 A CN201811035114 A CN 201811035114A CN 109241982 A CN109241982 A CN 109241982A
- Authority
- CN
- China
- Prior art keywords
- region
- frame
- feature
- target detection
- convolutional neural
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
The present invention discloses a kind of object detection method based on depth layer convolutional neural networks, carries out feature extraction network, including deep-neural-network and shallow-layer neural network to training image first, obtains union feature figure；Secondly using recommending net RPN further to extract joint characteristic pattern, region recommended characteristics figure is obtained；Then Feature Dimension Reduction is carried out to region recommended characteristics figure；Model is classified and returned using the region recommended characteristics figure after dimensionality reduction later, obtains target detection model；Finally test image is treated using the target detection model to be detected.The present invention can be obviously improved the accuracy rate to small target deteection, and target detection speed is not substantially reduced while keeping big target detection accuracy rate constant.
Description
Technical field
The present invention relates to target detection technique fields, and in particular to a kind of target inspection based on depth layer convolutional neural networks
Survey method.
Background technique
Target detection technique extensive application in terms of intelligent transportation, Road Detection, military target.With depth
The appearance of learning art and large-scale visual identity data set is spent, depth targets detection technique is developed rapidly, wherein with base
In R-CNN (Region-based Convolutional Neural Networks) two stages target detection frame and be based on
The single phase target detection frame directly returned is most representative.Target detection frame based on R-CNN is mainly by image convolution
Feature generates, object candidate area is recommended, candidate region is classified and recurrence three parts composition.With traditional object detection method
It compares, R-CNN series methods eliminate the artificial subjective one-sided for extracting feature, while realizing target's feature-extraction and classification
Process it is two-in-one.Candidate region recommendation is eliminated based on the single phase target detection frame object detection method directly returned
Link, by directly returning the classification of target in multiple positions of image and frame completes target detection.With based on direct
The method of recurrence is compared, although the detection speed of R-CNN series methods is slow, detection accuracy generally wants high, due to big
Most object detection tasks is more to precise requirements, thus the application based on R-CNN series methods is more extensive.
Paper " Rich feature hierarchies for accurate object detection and
Semantic segmentation Tech report is (using function gradation structure abundant in accurate target detection and language
The technical report of justice segmentation) " (it is published in " Conference on Computer Vision and Pattern
Recognition (computer vision and pattern-recognition meeting) ") propose a kind of R- for recommending to combine with CNN based on region
CNN object detection method opens new era of depth targets detection.Paper " Fast R-CNN (the convolution based on fast area
Network method) " (it is published in " (the computer vision world International Conference on Computer Vision
Meeting) ") region of interest pond layer is embedded on the basis of R-CNN, and multitask loss function is added in a network and carries out frame
It returns, detects the candidate frame of target more acurrate.Paper " Faster R-CNN:Towards Real-Time Object
Detection with Region Proposal Networks (faster R-CNN: suggests network implementations target using region
Real-time detection) " (it is published in " International Conference on Neural Information Processing
Systems (international neural information processing systems meeting) ") region recommendation network is embedded on the basis of Fast R-CNN
(Region Proposal Network, RPN), significantly improves the speed of target detection, and truly realize end
Training and test pattern to end.However, the object detection method of current the type is all it can be seen from this series of method
The structure of improved model or new detection means is added around the promotion of detection accuracy and speed, but most methods all cause
Power is easy to ignore the extracting method of target signature, this causes target detection to wisp in the improvement to object detection method
Detectability is insufficient, it is therefore necessary to which inventing a kind of can improve wisp detectability and be able to maintain big object detection ability again
Method.
Summary of the invention
To be solved by this invention is that existing object detection method real-time is poor and not high to small target deteection accuracy rate
Problem provides a kind of object detection method based on depth layer convolutional neural networks, is keeping big target detection accuracy rate not
While change, it can be obviously improved the accuracy rate to small target deteection, and target detection speed is not substantially reduced.
To solve the above problems, the present invention is achieved by the following technical solutions:
Based on the object detection method of depth layer convolutional neural networks, comprise the following steps that
Step 1, based on the target detection model of ImageNet data set pre-training, target detection model is joined
Number initialization；
Step 2 carries out feature extraction to training image, it may be assumed that
Training image is carried out the convolution feature that convolution algorithm extracts image by step 2.1；
The obtained convolution feature of step 2.1 is respectively fed to shallow-layer convolutional neural networks and deep layer convolution by step 2.2
Neural network carries out feature extraction；
Step 2.3, will be by the extracted feature of the obtained shallow-layer convolutional neural networks of step 2.2 and deep layer convolutional Neural
The extracted feature of network carries out organic joint, and is compressed into a unified space, obtains union feature figure；
Step 3 carries out traversal and convolution, and using area using sliding window to the obtained union feature figure of step 2
The anchor mechanisms of recommendation network generate a certain number of regions on union feature figure and recommend frame, and recommend frame according to these regions
Region recommended characteristics figure is extracted from union feature figure；
Step 4 carries out dimensionality reduction to the obtained region recommended characteristics figure of step 3 using Feature Dimension Reduction device；
Step 5 is sent in target detection model the obtained dimensionality reduction rear region recommended characteristics figure of step 4, to target
Detection model carries out classification based training and regression training, obtains final target detection model；
Image to be tested is sent in the obtained final target detection model of step 5, and obtains to be measured by step 6
Attempt classification and the regression result of picture.
As an improvement, the step 3 still further comprises the mistake for recommending frame to be screened in the region in joint characteristic pattern
Journey: recommend frame to use non-maximum suppression method in obtained region, retain the region for being more than or equal to α with real estate Duplication
Frame is recommended to recommend frame as positive sample region, and region of the Duplication less than β recommends frame as negative sample region and recommends frame, finally
The score height for recommending frame according to region recommends frame and negative sample region recommendation frame middle selection score most from positive sample region
δ high region recommends frame to recommend frame as the region ultimately remained in union feature figure, and is ultimately remained according to these
Frame is recommended to extract region recommended characteristics figure from union feature figure in region in union feature figure；Wherein α, β and δ are setting
Value, 0 < α < 1,0 < β < 1, δ > 1.
In above scheme, α > β.
Compared with prior art, the present invention has a characteristic that
(1) in feature extraction network, deep layer network is mainly used for capturing the high-level semantic of big target, and shallow-layer network master
It is used to retain the underlying image feature of Small object, two kinds of networks are combined the spy that can make full use of different convolutional layers by the present invention
Sign is to improve the purpose to target detection capabilities；
(2) RPN used directly generates candidate region on convolution characteristic pattern by using " anchor " mechanism.Although it is essential
On be still by the way of window sliding, but due to its region recommend, classification and return together share convolutional layer feature, from
And the detection speed of whole network is increased dramatically；
(3) Feature Dimension Reduction device can not only make more compact structure, moreover it is possible to carry out dimensionality reduction to characteristic pattern, furthermore Feature Dimension Reduction device
It instead of one layer of full articulamentum, increased speed；
(4) SoftmaxWithLoss function and SmoothL1Loss function are that most popular target detection is damaged at present
Function is lost, therefore can preferably complete object detection task.
Detailed description of the invention
Fig. 1 is the schematic diagram of the object detection method based on depth layer convolutional neural networks.
Fig. 2 is the schematic diagram for expanding convolution, and (a) is common convolution characteristic pattern；It (b) is that the expansion that flare factor is 2 is rolled up
Product characteristic pattern.
Specific embodiment
Technical solution for a better understanding of the present invention with reference to the accompanying drawing makees specifically the embodiment of the present invention
It is bright.The present embodiment is implemented based on the technical solution of the present invention, gives detailed embodiment and specific behaviour
Make process, but protection scope of the present invention is not limited to following embodiments.
Based on the target detection model of depth layer convolutional neural networks, be roughly divided into four parts: first part is feature
Extract network, including deep-neural-network and shallow-layer neural network；Second part is that net RPN is recommended in region；Part III is special
Levy dimensionality reduction network；4th is that full articulamentum and classification return layer.
In shallow-layer neural network, we no longer need to capture the high-level semantics features of image, and are desirable to obtain bottom
Characteristics of image, therefore do not need too deep network, that is, it does not need using a large amount of convolutional layer.It is more excellent in order to allow parallel-connection structure to obtain
Effect, since conv2-1, we only use 4 convolutional layers, and it is 5 × 5 that each layer, which has 24 kernel sizes, filling system
The filter that number is 3.Deep layer network and shallow-layer network to make possess identical spatial resolution, herein also in shallow-layer network
It is 4 × 4 that kernel size is designed after each convolutional layer, the average pond layer that step-length is 2, in this structure using be averaged Chi Huake with
Ensure that excessive image information will not be lost because of maximum pond.
In deep-neural-network, the parameter from conv1-1 to pool4 is identical as VGG16, we arrive conv5-1
It is 2 that tri- layers of conv5-3, which are all improved to fill factor, and kernel size is 3 × 3, step-length 1, the expansion convolution that flare factor is 2
Layer.Expansion convolution is the common method of image segmentation field, can increase receptive field in the case where not changing characteristic pattern size,
It include more global informations, realization principle is as shown in Figure 2, wherein (a) is common convolution characteristic pattern, is (b) expansion
The expansion convolution characteristic pattern that coefficient is 2.For 7 × 7 characteristic area, practical convolution kernels size is 3x3, and cavity value is 1,
Other weights are 0 i.e. in addition to 9 stains.Although not changing relative to common convolution characteristic pattern kernel size, its
Receptive field has had increased to 7x7, this allows each convolution output to contain more global informations.
Model needs to carry out joint training, the entire specific training process of model are as follows: first using alternative optimization coaching method
Step carrys out initialization feature using the model of ImageNet data set pre-training and extracts network, and with PASCAL VOC data set list
Solely training RPN obtains candidate frame.Second step reinitializes feature extraction net using the model of ImageNet data set pre-training
Network, and the candidate frame of first step generation is added, train a list on PASCAL VOC data set using DS-CNN model at this time
Only detection network, it is therefore an objective to obtain convolution layer parameter by the penalty values of full articulamentum and the candidate frame of RPN.Third step,
Re -training DS-CNN model, the model initialization obtained using second step and fixed convolution layer parameter, are not involved in convolutional layer
Backpropagation recycles the parameter of RPN in RPN model initialization individually trained in the first step and fixed DS-CNN, makes RPN not
Participate in backpropagation.Trained purpose is that feature extraction network is made to be connected with RPN.4th step, the model obtained using third step
Convolution layer parameter and RPN parameter reinitialize and fix DS-CNN model, so that convolutional layer and RPN is not involved in reversed biography
It broadcasts, the purpose of this training is the full articulamentum of fine tuning, the result optimized.
Referring to Fig. 1, a kind of object detection method based on depth layer convolutional neural networks specifically comprises the following steps:
Step (1) joins the training pattern of target detection based on the model of ImageNet data set pre-training
Number initialization.
Step (2) carries out feature extraction to training image.
Step (2.1) by the image of input by two layers with 64 kernel sizes be 3 × 3 filter (filter) into
Row convolution algorithm extracts the convolution feature of image, this two layers of convolutional layer is fixed using the model of ImageNet data set pre-training
Parameter is not involved in backpropagation.
The convolutional layer feature that step (2.2) utilizes step (2.1) to obtain continues to be fed into shallow-layer convolutional neural networks and carries out spy
Sign is extracted.Shallow-layer convolutional neural networks include 4 layers of convolutional layer and 4 layers of average pond layer.
For step (2.3) while step (2.1) carry out, the convolutional layer feature in step (2.1) is also fed into deep layer convolution
Neural network carries out feature extraction.Deep layer convolutional neural networks include 11 layers of convolutional layer and 4 layers of maximum pond layer.
Feature in (2.2) and (2.3) is carried out organic joint by step (2.4) Concat feature combiner, and is compressed into
One unified space, the dimension after joint is 536 dimensions.
Step (3) obtains object candidate area recommended characteristics figure.
Traversal and convolution, using area recommendation network are carried out using 3 × 3 sliding window to the characteristic pattern after joint
" anchor " mechanism of (Region Proposal Network, RPN), every sliding is primary, and the center position of window can generate 12
Kind " anchor " generates 12 regions and recommends frames, and recommends frame to extract certain amount from union feature figure according to these regions
Recommended characteristics figure, the corresponding recommended characteristics figure of frame is recommended in one of region.
In the present embodiment, RPN carries out traversal and convolution using 3 × 3 sliding window to the characteristic pattern after joint, 3
The center position of × 3 sliding windows, 4 kinds of scales (64,128,256,512) of corresponding input picture and 3 kinds of length-width ratios (1:1,
1:2,2:1), it can produce 12 kinds of different regions altogether and recommend frame, that is, generate 12 kinds of anchors.Therefore for 14 × 14 feature of input
Figure shares a region in about 2300 (14 × 14 × 12) and recommends frame.
Recommend frame to screen effective region, RPN partial parameters are adjusted, are retained using non-maximum suppression method
It is more than or equal to 0.5 conduct positive sample with real estate Duplication, and Duplication is used as negative sample, last basis less than 0.3
Region recommendation scores height, 500 regions for choosing highest scoring recommend frame as the area ultimately remained on union feature figure
Recommend frame in domain.
After obtaining region and recommending frame, RPN can recommend region frame and corresponding characteristic pattern to take out as new characteristic pattern
That is provincial characteristics figure.Then one trained picture of every input eventually exports 500 region recommended characteristics figures.
Step (4) carries out dimensionality reduction to region recommended characteristics figure using Feature Dimension Reduction device.
Feature Dimension Reduction device is made of region of interest pond layer and monokaryon convolutional layer.Region of interest pond layer can be after RPN
The characteristic pattern of fixed size is exported, plays compressive features figure herein.Monokaryon convolutional layer be kernel size be 1 × 1 and
The convolutional layer that step-length is 1, using can not only make more compact structure after the layer of region of interest pond, moreover it is possible to region recommended characteristics
Figure carries out dimensionality reduction.Feature Mapping size can be fixed as to 7 × 7 by using dimensionality reduction device, and by dimension from 536 be reduced to 512 after again
Sign is sent into full articulamentum.
Step (5) carries out classification and regression training to model using region recommended characteristics figure, obtains target detection training mould
Type.
Region recommended characteristics figure after dimensionality reduction obtains 4096 dimensions by full articulamentum (fully connected, FC) processing
Characteristic pattern, then carry out classification based training and regression training.
Layer of classifying includes 2 elements, for differentiating target and non-targeted estimated probability.Classification based training is the spy of 4096 dimensions
Sign figure obtains the characteristic pattern of 21 dimensions by full articulamentum cls_score layers, and cls_score layers, for classifying, export K+1 dimension group
P indicates the probability for belonging to K class and background.Because used training dataset PASCAL VOC has K=20 class, and background
Belong to 1 class, then the output of full articulamentum is 21.
Loss_cls layers use SoftmaxWithLoss function as the loss function of classification.It is corresponding by the u that really classifies
Probability determine that calculation formula is
Lcls=-log pu (1)
Returning layer includes 4 coordinate elements (x, y, w, h), for determining target position.Regression training is the spy of 4096 dimensions
Sign figure obtains the characteristic pattern of 84 dimensions by full articulamentum bbox_prdict layers, and bbox_prdict layers for adjusting candidate region position
It sets, exports 4*K dimension group t, indicate the parameter that should translate scaling when being belonging respectively to K class.
The loss_bbox layers of loss function for using SmoothL1Loss function to position as detection block.It is used to compare back
The parameter t for the translation scaling predicted when returninguDifference between true parameter v, calculation formula are
Wherein function g is Smooth L1 error, and value formula is
SoftmaxWithLoss function and SmoothL1Loss function is used to be iterated and ask as loss function respectively
The minimum value of Loss.Final training pattern is obtained after the completion of iteration.The result of loss function is classification results and regression result
Weighted sum, do not consider if being classified as background return loss, its calculation formula is
Step (6) fixes all parameters in test network using the model that training obtains as initialization model, uses
Softmax classifier and object candidate area recommended method obtain classification and the regression result of image to be tested.
It should be noted that although the above embodiment of the present invention be it is illustrative, this be not be to the present invention
Limitation, therefore the invention is not limited in above-mentioned specific embodiment.Without departing from the principles of the present invention, all
The other embodiment that those skilled in the art obtain under the inspiration of the present invention is accordingly to be regarded as within protection of the invention.
Claims (3)
1. the object detection method based on depth layer convolutional neural networks, characterized in that comprise the following steps that
Step 1, based on the target detection model of ImageNet data set pre-training, to target detection model carry out parameter at the beginning of
Beginningization；
Step 2 carries out feature extraction to training image, it may be assumed that
Training image is carried out the convolution feature that convolution algorithm extracts image by step 2.1；
The obtained convolution feature of step 2.1 is respectively fed to shallow-layer convolutional neural networks and deep layer convolutional Neural by step 2.2
Network carries out feature extraction；
Step 2.3, will be by the extracted feature of the obtained shallow-layer convolutional neural networks of step 2.2 and deep layer convolutional neural networks
Extracted feature carries out organic joint, and is compressed into a unified space, obtains union feature figure；
Step 3 carries out traversal and convolution using sliding window to the obtained union feature figure of step 2, and using area is recommended
The anchor mechanisms of network generate a certain number of regions on union feature figure and recommend frame, and recommend frame from connection according to these regions
It closes in characteristic pattern and extracts region recommended characteristics figure；
Step 4 carries out dimensionality reduction to the obtained region recommended characteristics figure of step 3 using Feature Dimension Reduction device；
Step 5 is sent in target detection model the obtained dimensionality reduction rear region recommended characteristics figure of step 4, to target detection
Model carries out classification based training and regression training, obtains final target detection model；
Image to be tested is sent in the obtained final target detection model of step 5 by step 6, and is obtained and to be measured attempted
The classification of picture and regression result.
2. the object detection method according to claim 1 based on depth layer convolutional neural networks, characterized in that the step
Rapid 3 still further comprise the process for recommending frame to be screened in the region in joint characteristic pattern: recommending frame to obtained region
Using non-maximum suppression method, retains the region recommendation frame with real estate Duplication more than or equal to α and pushed away as positive sample region
Frame is recommended, and region of the Duplication less than β recommends frame as negative sample region and recommends frame, finally recommends the score of frame high according to region
It is low, recommend δ region of frame and negative sample region recommendation frame middle selection highest scoring to recommend frame conduct from positive sample region
It ultimately remains in the region in union feature figure and recommends frame, and ultimately remain in the region in union feature figure according to these and recommend
Frame extracts region recommended characteristics figure from union feature figure；Wherein α, β and δ are setting value, 0 < α < 1,0 < β < 1, δ > 1.
3. the object detection method according to claim 1 based on depth layer convolutional neural networks, characterized in that α > β.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811035114.4A CN109241982B (en) | 2018-09-06 | 2018-09-06 | Target detection method based on deep and shallow layer convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811035114.4A CN109241982B (en) | 2018-09-06 | 2018-09-06 | Target detection method based on deep and shallow layer convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109241982A true CN109241982A (en) | 2019-01-18 |
| CN109241982B CN109241982B (en) | 2021-01-29 |
Family
ID=65060758
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811035114.4A Expired - Fee Related CN109241982B (en) | 2018-09-06 | 2018-09-06 | Target detection method based on deep and shallow layer convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109241982B (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109902732A (en) * | 2019-02-22 | 2019-06-18 | 哈尔滨工业大学（深圳） | Automobile automatic recognition method and relevant apparatus |
| CN109934115A (en) * | 2019-02-18 | 2019-06-25 | 苏州市科远软件技术开发有限公司 | Construction method, face identification method and the electronic equipment of human face recognition model |
| CN110032975A (en) * | 2019-04-15 | 2019-07-19 | 禁核试北京国家数据中心 | A kind of pick-up method of seismic phase |
| CN110276289A (en) * | 2019-06-17 | 2019-09-24 | 厦门美图之家科技有限公司 | Generate the method and human face characteristic point method for tracing of Matching Model |
| CN110516670A (en) * | 2019-08-26 | 2019-11-29 | 广西师范大学 | Suggested based on scene grade and region from the object detection method for paying attention to module |
| CN111209962A (en) * | 2020-01-06 | 2020-05-29 | 电子科技大学 | Combined image classification method based on CNN (CNN) feature extraction network) and combined heat map feature regression |
| CN111310760A (en) * | 2020-02-13 | 2020-06-19 | 辽宁师范大学 | Method for detecting onychomycosis characters by combining local prior characteristics and depth convolution characteristics |
| CN111461161A (en) * | 2019-01-22 | 2020-07-28 | 斯特拉德视觉公司 | Object detection method and device based on CNN and strong fluctuation resistance |
| CN111538916A (en) * | 2020-04-20 | 2020-08-14 | 重庆大学 | Interest point recommendation method based on neural network and geographic influence |
| CN111583655A (en) * | 2020-05-29 | 2020-08-25 | 苏州大学 | Traffic flow detection method, device, equipment and medium |
| CN111783942A (en) * | 2020-06-08 | 2020-10-16 | 北京航天自动控制研究所 | Brain cognition process simulation method based on convolution cyclic neural network |
| EP3690731A3 (en) * | 2019-01-31 | 2020-10-28 | StradVision, Inc. | Method and device for attention-driven resource allocation by using reinforcement learning and v2x communication to thereby achieve safety of autonomous driving |
| CN113449406A (en) * | 2020-03-27 | 2021-09-28 | 华晨宝马汽车有限公司 | Tightening tool scheme recommendation method and device and storage medium |
| CN113516040A (en) * | 2021-05-12 | 2021-10-19 | 山东浪潮科学研究院有限公司 | Method for improving two-stage target detection |
| CN113657933A (en) * | 2021-08-16 | 2021-11-16 | 浙江新再灵科技股份有限公司 | Preparation method of elevator advertisement recommendation data |
| CN114882596A (en) * | 2022-07-08 | 2022-08-09 | 深圳市信润富联数字科技有限公司 | Behavior early warning method and device, electronic equipment and storage medium |
| CN116910371A (en) * | 2023-09-07 | 2023-10-20 | 南京大数据集团有限公司 | Recommendation method and system based on deep relation |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105512661A (en) * | 2015-11-25 | 2016-04-20 | 中国人民解放军信息工程大学 | Multi-mode-characteristic-fusion-based remote-sensing image classification method |
| CN106599939A (en) * | 2016-12-30 | 2017-04-26 | 深圳市唯特视科技有限公司 | Real-time target detection method based on region convolutional neural network |
| WO2017079522A1 (en) * | 2015-11-04 | 2017-05-11 | Nec Laboratories America, Inc. | Subcategory-aware convolutional neural networks for object detection |
| CN107240066A (en) * | 2017-04-28 | 2017-10-10 | 天津大学 | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks |
| CN107239731A (en) * | 2017-04-17 | 2017-10-10 | 浙江工业大学 | A kind of gestures detection and recognition methods based on Faster R CNN |
| CN107316001A (en) * | 2017-05-31 | 2017-11-03 | 天津大学 | Small and intensive method for traffic sign detection in a kind of automatic Pilot scene |
| CN107437245A (en) * | 2017-06-26 | 2017-12-05 | 西南交通大学 | High-speed railway touching net method for diagnosing faults based on depth convolutional neural networks |
| CN107742099A (en) * | 2017-09-30 | 2018-02-27 | 四川云图睿视科技有限公司 | A kind of crowd density estimation based on full convolutional network, the method for demographics |
| CN107862287A (en) * | 2017-11-08 | 2018-03-30 | 吉林大学 | A kind of front zonule object identification and vehicle early warning method |
| CN108230269A (en) * | 2017-12-28 | 2018-06-29 | 北京智慧眼科技股份有限公司 | Grid method, device, equipment and storage medium are gone based on depth residual error network |
-
2018
- 2018-09-06 CN CN201811035114.4A patent/CN109241982B/en not_active Expired - Fee Related
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017079522A1 (en) * | 2015-11-04 | 2017-05-11 | Nec Laboratories America, Inc. | Subcategory-aware convolutional neural networks for object detection |
| CN105512661A (en) * | 2015-11-25 | 2016-04-20 | 中国人民解放军信息工程大学 | Multi-mode-characteristic-fusion-based remote-sensing image classification method |
| CN106599939A (en) * | 2016-12-30 | 2017-04-26 | 深圳市唯特视科技有限公司 | Real-time target detection method based on region convolutional neural network |
| CN107239731A (en) * | 2017-04-17 | 2017-10-10 | 浙江工业大学 | A kind of gestures detection and recognition methods based on Faster R CNN |
| CN107240066A (en) * | 2017-04-28 | 2017-10-10 | 天津大学 | Image super-resolution rebuilding algorithm based on shallow-layer and deep layer convolutional neural networks |
| CN107316001A (en) * | 2017-05-31 | 2017-11-03 | 天津大学 | Small and intensive method for traffic sign detection in a kind of automatic Pilot scene |
| CN107437245A (en) * | 2017-06-26 | 2017-12-05 | 西南交通大学 | High-speed railway touching net method for diagnosing faults based on depth convolutional neural networks |
| CN107742099A (en) * | 2017-09-30 | 2018-02-27 | 四川云图睿视科技有限公司 | A kind of crowd density estimation based on full convolutional network, the method for demographics |
| CN107862287A (en) * | 2017-11-08 | 2018-03-30 | 吉林大学 | A kind of front zonule object identification and vehicle early warning method |
| CN108230269A (en) * | 2017-12-28 | 2018-06-29 | 北京智慧眼科技股份有限公司 | Grid method, device, equipment and storage medium are gone based on depth residual error network |
Non-Patent Citations (4)
| Title |
|---|
| FISHER YU 等: "Multi-scale context aggregation by dilated convolutions", 《ARXIV》 * |
| SHAOQING REN 等: "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| 张飞云: "基于深度学习的车辆定位及车型识别研究", 《中国优秀硕士学位论文全文数据库 工程科技II辑》 * |
| 曹诗雨 等: "基于Fast R-CNN的车辆目标检测", 《中国图象图形学报》 * |
Cited By (33)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111461161A (en) * | 2019-01-22 | 2020-07-28 | 斯特拉德视觉公司 | Object detection method and device based on CNN and strong fluctuation resistance |
| CN111461161B (en) * | 2019-01-22 | 2024-03-15 | 斯特拉德视觉公司 | CNN-based object detection method and device with strong fluctuation resistance |
| US11010668B2 (en) | 2019-01-31 | 2021-05-18 | StradVision, Inc. | Method and device for attention-driven resource allocation by using reinforcement learning and V2X communication to thereby achieve safety of autonomous driving |
| EP3690731A3 (en) * | 2019-01-31 | 2020-10-28 | StradVision, Inc. | Method and device for attention-driven resource allocation by using reinforcement learning and v2x communication to thereby achieve safety of autonomous driving |
| CN109934115A (en) * | 2019-02-18 | 2019-06-25 | 苏州市科远软件技术开发有限公司 | Construction method, face identification method and the electronic equipment of human face recognition model |
| CN109934115B (en) * | 2019-02-18 | 2021-11-02 | 苏州市科远软件技术开发有限公司 | Face recognition model construction method, face recognition method and electronic equipment |
| CN109902732B (en) * | 2019-02-22 | 2021-08-27 | 哈尔滨工业大学（深圳） | Automatic vehicle classification method and related device |
| CN109902732A (en) * | 2019-02-22 | 2019-06-18 | 哈尔滨工业大学（深圳） | Automobile automatic recognition method and relevant apparatus |
| CN110032975B (en) * | 2019-04-15 | 2021-09-07 | 禁核试北京国家数据中心 | Seismic facies picking method |
| CN110032975A (en) * | 2019-04-15 | 2019-07-19 | 禁核试北京国家数据中心 | A kind of pick-up method of seismic phase |
| CN110276289A (en) * | 2019-06-17 | 2019-09-24 | 厦门美图之家科技有限公司 | Generate the method and human face characteristic point method for tracing of Matching Model |
| CN110276289B (en) * | 2019-06-17 | 2021-09-07 | 厦门美图之家科技有限公司 | Method for generating matching model and face characteristic point tracking method |
| CN110516670B (en) * | 2019-08-26 | 2022-04-22 | 广西师范大学 | Target detection method based on scene level and area suggestion self-attention module |
| CN110516670A (en) * | 2019-08-26 | 2019-11-29 | 广西师范大学 | Suggested based on scene grade and region from the object detection method for paying attention to module |
| CN111209962B (en) * | 2020-01-06 | 2023-02-03 | 电子科技大学 | Combined image classification method based on CNN (CNN) feature extraction network and combined heat map feature regression |
| CN111209962A (en) * | 2020-01-06 | 2020-05-29 | 电子科技大学 | Combined image classification method based on CNN (CNN) feature extraction network) and combined heat map feature regression |
| CN111310760A (en) * | 2020-02-13 | 2020-06-19 | 辽宁师范大学 | Method for detecting onychomycosis characters by combining local prior characteristics and depth convolution characteristics |
| CN111310760B (en) * | 2020-02-13 | 2023-05-26 | 辽宁师范大学 | Method for detecting alpha bone inscription characters by combining local priori features and depth convolution features |
| CN113449406A (en) * | 2020-03-27 | 2021-09-28 | 华晨宝马汽车有限公司 | Tightening tool scheme recommendation method and device and storage medium |
| CN113449406B (en) * | 2020-03-27 | 2024-01-23 | 华晨宝马汽车有限公司 | Tightening tool scheme recommendation method and device and storage medium |
| CN111538916B (en) * | 2020-04-20 | 2023-04-18 | 重庆大学 | Interest point recommendation method based on neural network and geographic influence |
| CN111538916A (en) * | 2020-04-20 | 2020-08-14 | 重庆大学 | Interest point recommendation method based on neural network and geographic influence |
| CN111583655B (en) * | 2020-05-29 | 2021-12-24 | 苏州大学 | Traffic flow detection method, device, equipment and medium |
| CN111583655A (en) * | 2020-05-29 | 2020-08-25 | 苏州大学 | Traffic flow detection method, device, equipment and medium |
| CN111783942B (en) * | 2020-06-08 | 2023-08-01 | 北京航天自动控制研究所 | Brain cognitive process simulation method based on convolutional recurrent neural network |
| CN111783942A (en) * | 2020-06-08 | 2020-10-16 | 北京航天自动控制研究所 | Brain cognition process simulation method based on convolution cyclic neural network |
| CN113516040B (en) * | 2021-05-12 | 2023-06-20 | 山东浪潮科学研究院有限公司 | Method for improving two-stage target detection |
| CN113516040A (en) * | 2021-05-12 | 2021-10-19 | 山东浪潮科学研究院有限公司 | Method for improving two-stage target detection |
| CN113657933A (en) * | 2021-08-16 | 2021-11-16 | 浙江新再灵科技股份有限公司 | Preparation method of elevator advertisement recommendation data |
| CN114882596B (en) * | 2022-07-08 | 2022-11-15 | 深圳市信润富联数字科技有限公司 | Behavior early warning method and device, electronic equipment and storage medium |
| CN114882596A (en) * | 2022-07-08 | 2022-08-09 | 深圳市信润富联数字科技有限公司 | Behavior early warning method and device, electronic equipment and storage medium |
| CN116910371A (en) * | 2023-09-07 | 2023-10-20 | 南京大数据集团有限公司 | Recommendation method and system based on deep relation |
| CN116910371B (en) * | 2023-09-07 | 2024-01-23 | 南京大数据集团有限公司 | Recommendation method and system based on deep relation |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109241982B (en) | 2021-01-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109241982A (en) | Object detection method based on depth layer convolutional neural networks | |
| CN111797716B (en) | Single target tracking method based on Siamese network | |
| CN110472627B (en) | End-to-end SAR image recognition method, device and storage medium | |
| CN111640125B (en) | Aerial photography graph building detection and segmentation method and device based on Mask R-CNN | |
| CN108416266B (en) | Method for rapidly identifying video behaviors by extracting moving object through optical flow | |
| CN108537824B (en) | Feature map enhanced network structure optimization method based on alternating deconvolution and convolution | |
| CN109034210A (en) | Object detection method based on super Fusion Features Yu multi-Scale Pyramid network | |
| CN110033473B (en) | Moving target tracking method based on template matching and depth classification network | |
| CN108009509A (en) | Vehicle target detection method | |
| CN109961034A (en) | Video object detection method based on convolution gating cycle neural unit | |
| CN110516539A (en) | Remote sensing image building extracting method, system, storage medium and equipment based on confrontation network | |
| CN110443763B (en) | Convolutional neural network-based image shadow removing method | |
| CN112784736B (en) | Character interaction behavior recognition method based on multi-modal feature fusion | |
| Wang et al. | Object instance detection with pruned Alexnet and extended training data | |
| CN111401293B (en) | Gesture recognition method based on Head lightweight Mask scanning R-CNN | |
| CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
| CN106023257A (en) | Target tracking method based on rotor UAV platform | |
| Pavel et al. | Recurrent convolutional neural networks for object-class segmentation of RGB-D video | |
| CN111833322B (en) | Garbage multi-target detection method based on improved YOLOv3 | |
| CN111460980A (en) | Multi-scale detection method for small-target pedestrian based on multi-semantic feature fusion | |
| CN112613350A (en) | High-resolution optical remote sensing image airplane target detection method based on deep neural network | |
| CN110334656A (en) | Multi-source Remote Sensing Images Clean water withdraw method and device based on information source probability weight | |
| CN110852199A (en) | Foreground extraction method based on double-frame coding and decoding model | |
| Zhou et al. | Building segmentation from airborne VHR images using Mask R-CNN | |
| CN110363218A (en) | A kind of embryo's noninvasively estimating method and device |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210129 Termination date: 20210906 |