CN108898060A - Based on the model recognizing method of convolutional neural networks under vehicle environment - Google Patents
Based on the model recognizing method of convolutional neural networks under vehicle environment Download PDFInfo
- Publication number
- CN108898060A CN108898060A CN201810540260.6A CN201810540260A CN108898060A CN 108898060 A CN108898060 A CN 108898060A CN 201810540260 A CN201810540260 A CN 201810540260A CN 108898060 A CN108898060 A CN 108898060A
- Authority
- CN
- China
- Prior art keywords
- loss function
- branch
- model
- cnn
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Abstract
The invention discloses the model recognizing methods based on convolutional neural networks under a kind of vehicle environment, propose the semantic compact bilinearity pond method for vehicle cab recognition, layered label tree and compact bilinearity pond are combined together by it, and have shown superior performance on CompCars data set and Stamford car data collection.In this way, compact bilinearity pond method is connected between the semanteme of automobile different stage using semantic, and reinforces them mutually during the training period.The present invention softmax loss function is generalized to be intended to make full use of priori knowledge evade loss function.Experiment shows that the present invention improves the accuracy rate of CompCars data set and the vehicle cab recognition task on the car data collection of Stamford.
Description
Technical field
The present invention relates to the objective classification method based on computer vision field, be mainly based upon caffe depth frame and
The improved model recognizing method of convolutional neural networks.
Background technique
Deep learning and convolutional neural networks（CNN）It is achieved in recent years in public safety field and makes us sipping the achievement of tongue.
In public security system, task relevant to automobile accounts for a big chunk of all Computer Vision Tasks.Car license recognition has been at present
Through being widely used in traffic safety system, vehicle cab recognition also has become the task of increased popularity in computer vision field.
2013, Krause et al. issued the data set of vehicle cab recognition（Stamford car data collection）.And it is led in computer vision
Various research work have been carried out in domain to carry out vehicle identification.With other general objects（As recognition of face and ImageNet are classified）
Identification or classification compare, vehicle cab recognition represents typical challenging fine granularity identification mission.Due between vehicle
Existing vision difference is very small, therefore the variation in classification is very small.In addition, the different shape of vehicle, the different views of observation
Point and appearance make automatic system（Even human eye）Be difficult to differentiate between the subclass of vehicle, for example, different brands different model vehicle.
Currently, there are many fine granularity recognition methods for being used for vehicle, and have on the car data collection of Stamford
Fairly good performance.For example, document 1（Xiao Liu, Tian Xia, Jiang Wang, and Yuanqing Lin.
2016. Fully Convolutional Attention Localization Networks: Efficient
Attention Localization for Fine-Grained Recognition. CoRR abs/1603.06765
(2016).）In FCAN can reach 89.1% accuracy rate, document 2（Tsung-Yu Lin, Aruni Roy Chowdhury,
and Subhransu Maji. 2015. Bilinear CNN Models for Fine-Grained Visual
Recognition. In 2015 IEEE International Conference on Computer Vision, ICCV
2015, Santiago, Chile, December 7-13, 2015. 1449–1457.）In BCNN can reach 91.3%
Accuracy.How to position differentiated region however, they only focus on and subtle vision difference shows.With it
His fine object is different, and vehicle has unique tree structure：Brand, model and time.Although existing largely more about being layered
The work sutdy of label study, but they use traditional basic CNN model, rather than apply fine-grained method.But it is
The accuracy and robustness of vehicle cab recognition are improved, the method for hierarchical classification is highly studied.
Summary of the invention
The invention proposes a kind of new deep neural network frames that can learn to be layered multi-tag, wherein mainly including
Two o'clock innovation：1）Propose a kind of new neural network framework, i.e., semantic compact bilinearity pond, by document 3（Yang Gao,
Oscar Beijbom, Ning Zhang, and Trevor Darrell. 2016. Compact Bilinear
Pooling. In 2016 IEEE Conference on Computer Vision and Pattern Recognition,
CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. 317–326.）In compact bilinearity pond
（CBP）Method is combined with the semantic structure of vehicle, as shown in Figure 2.CNN branch A and branch B in original BCNN are extended to
The branch in Three Represents brand, model and time；2）The present invention devises the loss letter that can make full use of priori knowledge
Number, allows the result of brand branch to instruct the prediction of model branch.
The technical solution of invention is：
（1）Fine granularity identification
As described above, two main tasks of fine grit classification research are how to position differentiated region and to subtle vision
Difference is indicated.In order to solve these problems, many methods are proposed in recent years.Substantially, pervious model can be divided into
Two classes：1）Disaggregated model based on supervised learning（It is known as S model in the present invention）With 2）Classification mould based on unsupervised learning
Type（It is known as U model in the present invention）.
S model passes through the bounding box that application additionally artificially marks and part annotates（Such as car light, logo and automobile interior decoration）
To solve the problems, such as.2014, et al..The R-CNN based on part is proposed, it can train three kinds of detection models：The first
For detecting the fine part of automobile, second for detecting headstock, the third is for detecting vehicle body.In the R- based on part
After CNN, S.Branson et al. proposes posture normalization CNN, which carries out posture alignment behaviour to the regional area of image
Make, and extracts the feature of different levels according to the image-region of different levels.2016, there is scholar to propose Mask- using FCN
CNN reduces the influence of noise from background to position differentiated part and select useful convolution descriptor.Obviously, above-mentioned
Method cannot generally be generally applicable to various situations, because their working depth is dependent on a large amount of mark.For this original
Cause, U model can become mainstream naturally.Two-stage pays close attention to model and applies visual attention in task, and uses cluster filter from sense
The positioning of part is realized in the region of interest.Many research work all use attention mechanism in existing method.FCAN
Pay attention to positioning network using the full convolution based on intensified learning, it positions multiple portions using attention mechanism simultaneously.Document 4
（Jianlong Fu, Heliang Zheng, and Tao Mei. 2017. Look Closer to See Better:
Recurrent Attention Convolutional Neural Network for Fine-Grained Image
Recognition. In 2017 IEEE Conference on Computer Vision and Pattern
Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. 4476–4484.）In
RA-CNN recursively learns accurate region jointly and pays attention to and the character representation based on region.RA-CNN uses attention interested
Sub-network in a recursive manner pay close attention to from coarse to fine by generating region.Equally, document 5（Heliang Zheng, Jianlong Fu,
Tao Mei, and Jiebo Luo. 2017. Learning Multi-attention Convolutional Neural
Network for FineGrained Image Recognition. In IEEE International Conference
on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017. 5219–
5227.）In MA-CNN also use identical attention mechanism, but MA-CNN uses the pipeline being made of three subnets：Volume
Product network, channel packet network and sorter network.Although having used attention mechanism, Bilinear Structure has more attracted researcher,
It is two feature extractors that its main thought, which is by two CNN drop applications, and uses external knot in each distinguishing part
Structure outputs it multiplication.Then, it obtains image descriptor after pondization operation.However, due to original bilinearity pond
Higher-dimension problem has scholar to come significantly under the premise of not sacrificing precision using the projection of tensor sketch and the projection of random Maclaurin
Reduce dimension.In the present invention, compact bilinearity pond method is used as our basic fine granularity technologies.
（2）Layered Learning
Traditional CNN disaggregated model is typically designed to be predicted as the softmax layer of output with top n.Most of them be it is continuous,
Minority is parallel, but is intended merely to reduce gradient disappearance, such as GoogLeNet.However, being distinguished different classes of for us
Degree of difficulty be different.Some classifications may be more difficult to predict than other classifications, if we are by CNN disaggregated model with flat
Structure is trained, then this classification and imbalance.In addition to this, CNN of the CNN model of lower level compared to higher level
The feature that model generates is less.So naturally enough, some papers propose the CNN model of layering, such as HD-CNN.HD-CNN's
Main contributions are to prove that the hierarchical attribute of CNN model can be combined with layered label structure.The use of HD-CNN model and CNN mould
Type lower level is corresponding to share thick category classifier to distinguish simple classification, and uses actually CNN model higher level
Secondary fine classification classifier distinguishes complicated classification.It is trained end to end in addition, B-CNN is completed, rather than as HD-
The process of CNN two such stage-training.More fortunately, weight distribution is lost in the overall structure of B-CNN and modification
BT strategy is that the present invention provides good guidances.
Although HD-CNN and B-CNN use Layered Learning, they are special just with the level of CNN model different levels
Sign, rather than the semantic connection between label.Some work attempt to find out this problem, document 6（Jia Deng,
Alexander C. Berg, and Fei-Fei Li. 2011. Hierarchical semantic indexing for
large scale image retrieval. In The 24th IEEE Conference on Computer Vision
and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June
2011.785–792.）In method be intended to ask by using the predefined comparison function constructed according to known structure to solve this
Topic.Document 7（Xiaofan Zhang, Feng Zhou, Yuanqing Lin, and Shaoting Zhang. 2016.
Embedding Label Structures for Fine-Grained Feature Representation. In 2016
IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las
Vegas, NV, USA, June 27-30, 2016. 1114–1123.）In method triple loss functions are generalized to and can be answered
Four-tuple for the training stage constrains.
（3）Semantic compact bilinearity pond
Semantic compact bilinearity pond frame in the present invention combines the semantic structure of CBP method and vehicle.With mark
Sign the layered label structure of tree representation vehicle.In tag tree, coarse level is brand（Such as Audi）, finer grade is
Model（Such as Audi A8）, most thin grade is the time（Such as Audi A8 2011）.
The CNN branch B of original BCNN is extended to three branches by the present invention, extracts feature as traditional CNN model；It will
CNN branch A is extended to three branches, and main task is the differentiated region of positioning.Between brand branch and model branch
Corresponding relationship for, we are by brand-A（Indicate the brand branch of branch A）With brand-B（Indicate the brand branch of branch B）
Output be multiplied, generate brand branch bilinearity vector.Softmax loss layer later plays the role of loss function, removes
This softmax loss layer has also connect the combination of one softmax layers He argmax layers after bilinearity vector.Model branch
Similar to brand branch, but loss function is evaded in use, and input is two argmax layers of output of Liang Ge branch.Finally,
Softmax is lost and evades loss portfolio into final loss.
Since CBP method achieves good performance in finegrained tasks, and different branches have different losses
Function will reinforce the iteration of parameter in Liang Ge branch, it is possible to make full use of CBP method and the label using semanteme CBP structure
Tree construction identifies vehicle.
（4）Evade loss function
Shown in the definition such as formula (1) and formula (2) of traditional softmax loss function and gradient：
Formula (1)
Formula (2)
Wherein s indicates softmax loss function, and I is the quantity of batch processing, and y is the label of i-th of image.Refer to
The intersection entropy loss of i-th of image.
In order to make full use of the semantic information of vehicle, the present invention devises the new neural network of one kind and loses to make full use of
Priori knowledge.The loss function and its gradient definition used in the present invention evaded is as shown in formula (3) and formula (4)：
Formula (3)
Formula (4)
Loss function is evaded in wherein c expression.According to the decision-making mechanism being described below,Equal to 0 or 1.
Obviously, the element in formula (1) and formula (2) is continuous in integer range.In contrast, in formula (3) and formula (4)
With discrete element, because the loss of certain images in every batch of is set as zero by the present invention.In order to determine the loss of which image
It will be ignored, the present invention is based on the interdependencies between the label of two levels to construct label matrix.
Decide whether to retain the image based on the adaptation function of label matrix in addition, the present invention devises one.Also
It is to say, if the mismatch that interdepends of the model prediction of this image and corresponding brand label, the present invention will not calculate
The loss of this image.In this way, it is an object of the present invention to make full use of priori knowledge, so that the result of brand branch
It can be with the prediction of guidance model branch.
Compared to the prior art possessed beneficial effect is the present invention：
For compared with the existing technology, the present invention is made that both sides is improved：First, by the layer of CBP method and car category
Secondary structure is combined together, CNN（Branch A）And BCNN（Branch B）It combines and expands Liang Tiao branch vehicle major class for identification
（Brand）With vehicle subclass（Concrete model）；Second, a new loss function is devised, guarantees the prediction result of model level
Belong to the prediction result of correct brand levels.Since the above both sides is improved, the present invention improves vehicle cab recognition model
Accuracy and robustness are embodied in following three aspects：
1）The layering vehicle cab recognition CBP method for combining grade labelling tree and bilinearity pond shows on CompCars data set
Performance out shows：It can use bilinearity pond method to realize the semantic connection between different levels, in the training process
They can mutually reinforce；
2）The mode of learning of CBP model is different from traditional mode of learning of CNN model, in CBP model, returns in loss weight
After biography, learning rate can be substantially improved；
3）Softmax loss function has been generalized to by this model to be evaded in loss function, it is intended to ensure that the priority predicted is vapour
The brand of vehicle.Experiment shows that our method improves CompCars data set and gets on the car the accuracy of type identification.
Detailed description of the invention
Fig. 1 is the explanatory diagram of automobile semantic structure in the present invention；
Fig. 2 is semantic compact bilinearity pond convolutional neural networks proposed by the invention（SCBPCNN）Structural schematic diagram；
Fig. 3 is the structural schematic diagram of vehicle layered label tree；
Fig. 4 is the present invention without using schematic network structure when evading loss function；
Fig. 5 is the schematic network structure of the invention used when evading loss function；
Fig. 6 is the two automobile samples and their label for test that the present invention uses.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing to of the invention real
The specific embodiment of the recognition methods and system of applying the vehicle of example is illustrated.It should be appreciated that specific reality described herein
Example is applied only to explain the present invention, is not intended to limit the present invention.
（1）The data set that the present invention uses
Experiment of the invention carries out on the car data collection of Stamford first.Stamford car data collection is comprising 16,185,196
Kind automobile image.Wherein there are 8,144 training images and 8,041 test image, each classification substantially 5-5 is divided into training set
And test set.In experiment of the invention, the distribution of training set and test set is identical as official's distribution.
The present invention is also tested on CompCars data set.CompCars data set is a more fully automobile
Data set, wherein the data comprising two kinds of scenes, the respectively image of network property and monitoring property.Network property data
78,126 images are divided into three subsets, we use first part in an experiment.First part includes 431 sections of vehicles, always
Share 30955 vehicle images.Wherein we are by 16,016 images as training set, remaining 14,939 image is as survey
Examination collection.
According to the model label formally provided, the present invention constructs one for CompCars data set and includes 74 brand marks
Label, the tag tree of 431 model labels and 1343 time labels, and construct for Stamford car data collection comprising 49 product
The tag tree of board label and 196 model labels.
（2）Experiment description
In experiment of the invention, we construct one based on Stamford car data collection and Compcars first part data
The tag tree of collection.Tag tree will not used by brand label, signal label and time label composition in laboratory of the invention
Time label encounters convergence problem when because CNN model is in the training process to time label.Then, we use former
The CBP method of beginning trains two CNN models, wherein only brand branch or model branch, changes result and be considered as of the invention
Benchmark.After obtaining benchmark, We conducted the first combination experiment of CBP method and layered label tree, but Liang Ge branch
Softmax loss function is all used as usual.Then, in order to prove the practicability for evading loss function, we are by second
The substitution of softmax loss function is evaded in loss function, to make full use of priori knowledge, the result of brand branch is referred to
The prediction of guided mode type branch.
In all experiments, present invention uses two D networks as partial descriptor extractor, i.e. CNN branch A and B.
Input picture is dimensioned to 512x512 by us, the image for being then 448x448 at 10 sizes by an image cropping.
In the training process, we select stochastic gradient descent（SGD）As our optimization method, momentum 0.9.We use two
A step trains network as CBP.Firstly, the method that we use tensor sketch, tape symbol square root layer, L2 normalization
Layer, be also fully connected layer to network it is last it is several layers of be finely adjusted, including compact bilinearity pond layer.Then, we instruct again
Practice entire SCBP-CNN network.All depth models of the present invention use Caffe deep learning frame in two NVIDIA
It is trained on TITAN Xp GPU, all images for training of the present invention both are from CompCars data set and Stamford vapour
Car data collection enhances and is pre-processed without any data.
（3）Experimental result
Containing there are two the SCBP-CNN models of softmax loss function by the recognition accuracy of CompCars data set model level
The accuracy rate for improving 0.3%, Stanford Cars data set improves 0.5%.In addition, have softmax loss function and
The SCBP-CNN model for evading loss function makes the recognition accuracy of CompCars data set model level improve 1.2%,
The accuracy rate of Stanford Cars data set improves 1.1%.
For superiority of the invention is explained further, we list two samples：7 system of Benz GL series and BWM
Column.When using CBP method, which is identified as Audi A8 for left side, and right side is identified as Great Wall M1.However, our two
A SCBP-CNN model all uses the priori knowledge of brand level to carry out correct prediction to the automobile in left side.For right side
Automobile, there are two the SCBP-CNN models of softmax loss to make correct prediction to its brand for tool.Evade in application
After loss function, the model can the automobile model to right side correctly predicted.
It can to sum up obtain, the present invention improves the accuracy rate of vehicle cab recognition really.
Claims (4)
1. based on the model recognizing method of convolutional neural networks under a kind of vehicle environment, which is characterized in that for vehicle cab recognition
Semantic compact bilinearity pond method, it is intended to which make full use of priori knowledge evades loss function；
The semantic compact bilinearity pond method for vehicle cab recognition, layered label tree and compact bilinearity pondization are combined
Together, the semanteme of automobile different stage is attached；
It is described to evade loss function, make full use of priori knowledge to promote softmax loss function, not based on automobile semanteme
Interdependency between the label of same level constructs label matrix.
2. the method according to claim 1, wherein the CNN branch B of original BCNN is extended to three branches,
Feature is extracted as tradition CNN model；CNN branch A is extended to three branches, main task is the differentiated area of positioning
Domain.
3. the method according to claim 1, wherein the definition for evading loss function and gradient are expressed as：
Loss function is evaded in wherein c expression, and I is the quantity of batch processing, and y is the label of i-th of image,Refer to
The intersection entropy loss of i-th of image,Equal to 0 or 1.
4. according to the method described in claim 2, it is characterized in that, generating the bilinearity vector of brand branch, later
Softmax loss layer plays the role of loss function, in addition to this softmax loss layer, has also connect one after bilinearity vector
Softmax layers and argmax layers of combination, model branch is similar to brand branch, but loss function is evaded in use, and input is
Two argmax layers of output of Liang Ge branch, finally, softmax is lost and evades loss portfolio into final loss.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810540260.6A CN108898060A (en) | 2018-05-30 | 2018-05-30 | Based on the model recognizing method of convolutional neural networks under vehicle environment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810540260.6A CN108898060A (en) | 2018-05-30 | 2018-05-30 | Based on the model recognizing method of convolutional neural networks under vehicle environment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108898060A true CN108898060A (en) | 2018-11-27 |
Family
ID=64343627
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810540260.6A Pending CN108898060A (en) | 2018-05-30 | 2018-05-30 | Based on the model recognizing method of convolutional neural networks under vehicle environment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108898060A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109636846A (en) * | 2018-12-06 | 2019-04-16 | 重庆邮电大学 | Object localization method based on circulation attention convolutional neural networks |
| CN109766928A (en) * | 2018-12-21 | 2019-05-17 | 创新奇智(重庆)科技有限公司 | A kind of object classification method based on image, system and electronic equipment |
| CN109815973A (en) * | 2018-12-07 | 2019-05-28 | 天津大学 | A kind of deep learning method suitable for the identification of fish fine granularity |
| CN109948610A (en) * | 2019-03-14 | 2019-06-28 | 西南交通大学 | A kind of vehicle fine grit classification method in the video based on deep learning |
| CN110209857A (en) * | 2019-05-23 | 2019-09-06 | 厦门市美亚柏科信息股份有限公司 | The more attribute recognition approaches of vehicle, device, medium based on neural network structure search |
| CN111079851A (en) * | 2019-12-27 | 2020-04-28 | 常熟理工学院 | Vehicle type identification method based on reinforcement learning and bilinear convolution network |
| CN111340543A (en) * | 2020-02-25 | 2020-06-26 | 上海大学 | Fine-grained taxi taking demand prediction method based on attention mechanism |
| CN111882525A (en) * | 2020-07-01 | 2020-11-03 | 上海品览数据科技有限公司 | Image reproduction detection method based on LBP watermark characteristics and fine-grained identification |
| CN113011514A (en) * | 2021-03-29 | 2021-06-22 | 吉林大学 | Intracranial hemorrhage sub-type classification algorithm applied to CT image based on bilinear pooling |
| CN113743814A (en) * | 2021-09-13 | 2021-12-03 | 上海联影医疗科技股份有限公司 | Medical image quality evaluation and processing method, device, equipment and medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107886073A (en) * | 2017-11-10 | 2018-04-06 | 重庆邮电大学 | A kind of more attribute recognition approaches of fine granularity vehicle based on convolutional neural networks |
| CN108027899A (en) * | 2015-08-25 | 2018-05-11 | 高通股份有限公司 | Method for the performance for improving housebroken machine learning model |
-
2018
- 2018-05-30 CN CN201810540260.6A patent/CN108898060A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108027899A (en) * | 2015-08-25 | 2018-05-11 | 高通股份有限公司 | Method for the performance for improving housebroken machine learning model |
| CN107886073A (en) * | 2017-11-10 | 2018-04-06 | 重庆邮电大学 | A kind of more attribute recognition approaches of fine granularity vehicle based on convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| TSUNG-YU LIN ET AL.: "Improved Bilinear Pooling with CNNs", 《ARXIV》 * |
| YANG GAO ET AL.: "Compact Bilinear Pooling", 《ARXIV》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109636846A (en) * | 2018-12-06 | 2019-04-16 | 重庆邮电大学 | Object localization method based on circulation attention convolutional neural networks |
| CN109636846B (en) * | 2018-12-06 | 2022-10-11 | 重庆邮电大学 | Target positioning method based on cyclic attention convolution neural network |
| CN109815973A (en) * | 2018-12-07 | 2019-05-28 | 天津大学 | A kind of deep learning method suitable for the identification of fish fine granularity |
| CN109766928A (en) * | 2018-12-21 | 2019-05-17 | 创新奇智(重庆)科技有限公司 | A kind of object classification method based on image, system and electronic equipment |
| CN109766928B (en) * | 2018-12-21 | 2020-04-17 | 创新奇智(重庆)科技有限公司 | Object classification method and system based on image and electronic equipment |
| CN109948610A (en) * | 2019-03-14 | 2019-06-28 | 西南交通大学 | A kind of vehicle fine grit classification method in the video based on deep learning |
| CN110209857B (en) * | 2019-05-23 | 2022-04-22 | 厦门市美亚柏科信息股份有限公司 | Vehicle multi-attribute identification method, device and medium based on neural network structure search |
| CN110209857A (en) * | 2019-05-23 | 2019-09-06 | 厦门市美亚柏科信息股份有限公司 | The more attribute recognition approaches of vehicle, device, medium based on neural network structure search |
| CN111079851A (en) * | 2019-12-27 | 2020-04-28 | 常熟理工学院 | Vehicle type identification method based on reinforcement learning and bilinear convolution network |
| CN111340543A (en) * | 2020-02-25 | 2020-06-26 | 上海大学 | Fine-grained taxi taking demand prediction method based on attention mechanism |
| CN111340543B (en) * | 2020-02-25 | 2024-03-22 | 上海大学 | Fine granularity taxi taking demand prediction method based on attention mechanism |
| CN111882525A (en) * | 2020-07-01 | 2020-11-03 | 上海品览数据科技有限公司 | Image reproduction detection method based on LBP watermark characteristics and fine-grained identification |
| CN113011514A (en) * | 2021-03-29 | 2021-06-22 | 吉林大学 | Intracranial hemorrhage sub-type classification algorithm applied to CT image based on bilinear pooling |
| CN113011514B (en) * | 2021-03-29 | 2022-01-14 | 吉林大学 | Intracranial hemorrhage sub-type classification algorithm applied to CT image based on bilinear pooling |
| CN113743814A (en) * | 2021-09-13 | 2021-12-03 | 上海联影医疗科技股份有限公司 | Medical image quality evaluation and processing method, device, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108898060A (en) | Based on the model recognizing method of convolutional neural networks under vehicle environment | |
| Zhu et al. | Traffic sign detection and recognition using fully convolutional network guided proposals | |
| CN107134144B (en) | A kind of vehicle checking method for traffic monitoring | |
| CN109190479A (en) | A kind of video sequence expression recognition method based on interacting depth study | |
| CN105574550B (en) | A kind of vehicle identification method and device | |
| CN108764164A (en) | A kind of method for detecting human face and system based on deformable convolutional network | |
| CN106815604B (en) | Method for viewing points detecting based on fusion of multi-layer information | |
| CN109815867A (en) | A kind of crowd density estimation and people flow rate statistical method | |
| CN107862261A (en) | Image people counting method based on multiple dimensioned convolutional neural networks | |
| CN106529477A (en) | Video human behavior recognition method based on significant trajectory and time-space evolution information | |
| CN110263786A (en) | A kind of road multi-targets recognition system and method based on characteristic dimension fusion | |
| CN109086668A (en) | Based on the multiple dimensioned unmanned aerial vehicle remote sensing images road information extracting method for generating confrontation network | |
| CN109492583A (en) | A kind of recognition methods again of the vehicle based on deep learning | |
| CN105467449B (en) | Deep layer Thin oil sandwich quantitatively characterizing method based on earthquake classification Sensitive Attributes fusion | |
| CN109214336A (en) | A kind of vehicle window marker detection method and device | |
| CN112633071B (en) | Pedestrian re-identification data domain adaptation method based on data style decoupling content migration | |
| Ye et al. | Real-time object detection network in UAV-vision based on CNN and transformer | |
| CN110348396A (en) | A kind of road top text traffic sign and device based on deep learning | |
| CN106960176A (en) | A kind of pedestrian's gender identification method based on transfinite learning machine and color characteristic fusion | |
| CN108647595A (en) | Vehicle recognition methods again based on more attribute depth characteristics | |
| CN110956158A (en) | Pedestrian shielding re-identification method based on teacher and student learning frame | |
| CN109086659A (en) | A kind of Human bodys' response method and apparatus based on multimode road Fusion Features | |
| CN109993061A (en) | A kind of human face detection and tracing method, system and terminal device | |
| CN112966736B (en) | Vehicle re-identification method based on multi-view matching and local feature fusion | |
| CN107967480A (en) | A kind of notable object extraction method based on label semanteme |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181127 |
|
| WD01 | Invention patent application deemed withdrawn after publication |