CN106599827A - Small target rapid detection method based on deep convolution neural network - Google Patents
Small target rapid detection method based on deep convolution neural network Download PDFInfo
- Publication number
- CN106599827A CN106599827A CN201611129536.9A CN201611129536A CN106599827A CN 106599827 A CN106599827 A CN 106599827A CN 201611129536 A CN201611129536 A CN 201611129536A CN 106599827 A CN106599827 A CN 106599827A
- Authority
- CN
- China
- Prior art keywords
- small object
- network
- training
- convolution
- vgg16
- 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
- G06V20/588—Recognition of the road, e.g. of lane markings; Recognition of the vehicle driving pattern in relation to the road
-
- 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
-
- 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
-
- 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
- G06V20/582—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of traffic signs
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a small target rapid detection method based on a deep convolution neural network. The deep convolution neural network is improved by the following steps: selecting the sliding windows on the convolution feature map of the last shared convolution layer of a VGG16 network as candidate boxes, wherein the sliding windows adopted are half-pixel precision sliding window; deleting a fifth pooling layer, and retaining other convolution layers and pooling layers; adding a convolution layer with a 3*3 convolution kernel; and using two convolution layers with 1*1 convolution kernels to replace all full-connection layers in the network to get the network adopted in the invention, training the network using collected data to get a small target classification model, and using the model to detect small targets. By using the method, the computational complexity is reduced, and the detection rate of small targets is improved.
Description
Technical field
The invention belongs to autonomous driving vehicle drives field and senior drive assist system field, and in particular to a kind of to be based on depth
The Small object method for quick of convolutional neural networks.
Background technology
Autonomous driving vehicle (Autonomous vehicles driving, AVD), drives also known as pilotless automobile, computer
Automobile or wheeled mobile robot are sailed, is that one kind realizes unpiloted intelligent automobile by computer system.Autonomous driving vehicle
By artificial intelligence, vision calculating, radar, supervising device and global positioning system cooperative cooperating, computer can be appointed no
Operation motor vehicles in automatic safe ground during who generic operation.Senior drive assist system technical field is derived from and serves automatically
There is car in Driving Skills field, senior drive assist system (Advanced Driver Assistant System, ADAS)
Road keeps aid system, automatic parking aid system, BAS, reversing aid system and driving aid system.It is not difficult
It was found that, each subsystem be unable to do without the detection of Small object (lane line, pedestrian, the small-scale obstacle thing, traffic mark symbol etc.), senior
Drive assist system by the environmental data inside and outside vehicle-mounted imageing sensor collecting cart, then carries out object first
Detection, recognition and tracking, give warning in advance so as to predict danger, it is ensured that driver safety drives.The accurate detection of lane line at present
It is the bottleneck of autonomous driving vehicle driving and senior drive assist system technical field with positioning, using the data of imageing sensor acquisition
Judge lane line whether there is, and lane line occur accurate location be lane detection technology key, and the present invention
Problem to be solved.
Autonomous driving vehicle drives and senior drive assist system is developed so far, and various lane detection technologies emerge in an endless stream, mesh
Front existing method for detecting lane lines mainly has two classes, and a class is the method for feature based, and a class is the side based on model in addition
Method.
The method for detecting lane lines of feature based mainly using lane line feature identification lane line, for example color, texture,
The features such as the haar-like of shape, geometry and hand-designed, this kind of method speed are fast, but accuracy is low.With nerve net
The development of network, is gradually emerged based on the method for detecting lane lines of model, and is quickly grown, because of its higher robustness and accurately
Property becomes current main stream approach.This kind of method includes training and testing two parts, gathers depth of the mass data to design first
Degree neural network model model is trained, and the size of data set and diversification comprehensively determine training deep neural network model
Quality.After the deep neural network track line model for obtaining training, in new cycle testss, known with track line model
Other lane line simultaneously provides the confidence level of identification.Experiment proves that the method effect based on model is preferable, but as neutral net has
Substantial amounts of convolution operation, in the training, test and use in model, amount of calculation and EMS memory occupation amount are all very big, to hardware
Requirement is very high, and live effect is poor, is not suitable for the development and application of traffic scene embedded product, at present most of experts
It is devoted to lifting the speed and precision of the method.The present invention is also based on inspection of such method for Small objects such as lane lines
Survey is made that improvement.
The content of the invention
The present invention proposes a kind of Small object method for quick based on depth convolutional neural networks, and the method can be reduced
Computation complexity, improves the recall rate of Small object.
Small object method for quick of the present invention based on depth convolutional neural networks, including training and test two parts,
Concretely comprise the following steps:
(1) gather training data and pretreatment is carried out to training data;
Step (1) concretely comprises the following steps:
(1-1) imageing sensor is set up, acquisition need to extract the coloured image of Small object;
(1-2) training data is cleaned, removes dimmer appearance and the invalid targets being blocked, to improve training number
According to multiformity and training process efficiency;
(1-3) the Small object delineator in coloured image is fixed using own calibration tool;
(1-4) training data is extended, obtains enough training datas, to overcome the training data number of manual demarcation
The limited shortcoming of amount.
(2) network is trained using the training data of pretreatment, obtains Small object sorter model;
Step (2) concretely comprise the following steps:
(2-1) pre-training (Pre-train) stage:VGG16 networks are obtained using ImageNet database trainings;
(2-2) finely tune (Fine-tuning) stage:Based on VGG16 networks, carried out using the training data of pretreatment
Network training, obtains Small object sorter model;
Step (2-2's) concretely comprises the following steps:
(2-2-1) select sliding window on the convolution characteristic pattern of last shared convolutional layer (5-3 convolutional layers) of VGG16 networks
Mouth is used as candidate frame, and adopts half-pixel accuracy sliding window;
(2-2-2) on the basis of VGG16 network structures, the 5th pond layer is deleted, retains other convolutional layers and pond layer;
Increase convolutional layer of the convolution kernel for 3x3;It is all of full articulamentum in 1x1 convolutional layer alternate networks with two convolution kernels,
Obtain improved VGG16 network structures;
(2-2-3) network training is carried out using improved VGG16 network structures using the candidate frame selected, obtains Small object
Sorter model.
In step (2-2-1), select sliding window on convolution characteristic pattern make as candidate frame every on characteristic pattern
The candidate frame of the different sizes of individual position generation, and the ratio of width to height of candidate frame is limited, only have specific target tested to limit
Survey and position;Can solve the problem that output characteristic figure precision is too low to be accurately positioned target using half-pixel accuracy sliding window
Problem, and increased candidate frame corresponding to true frame probability.
In step (2-2-2), by the improvement to VGG16 network structures, existing net can be make use of to the full extent
Network resource, adjusts the visible range of network, to realize across passage interaction and information integration, decrease or increase dimension (convolution kernel passage
Quantity) and process arbitrary dimension input.
In step (2), it is speed-up computation, is calculated using GPU.
(3) collecting test data pretreatment is carried out to test data, obtains the coloured image of Small object;
(4) test image is given, the network that training is obtained performs a forward direction transmission, and each sliding window is obtained through network
To position and the dimensional information of the classification confidence and candidate frame of target (Small object/non-Small object), by classification confidence with it is solid
Determine thresholding to be compared, obtain the small target deteection result of high-class confidence level;The Small object of high-class confidence level herein is
Classification confidence exceedes the Small object of fixed threshold;
(5) Small object equation is obtained based on optimization method using the small target deteection result of all high confidence levels.
Small object method for quick of the present invention uses a kind of depth convolutional neural networks end to end, and the network expands
Region candidate network (Region Proposal Network, RPN) has been opened up, while space is added in candidate frame generting machanism
The translation transformation in domain can be used for detection and the positioning of little scaled target so as to improve the number of effective positive sample.
Small object method for quick proposed by the invention has relatively low computation complexity, the position of all Small objects
Only need to be obtained with by a forward direction transmission.
Compared with the method using manual designs feature, depth convolutional neural networks can be extracted with more high-level semantic
Feature;
Compared with the method based on adaboost sliding windows, depth convolutional neural networks have more preferable little mesh end to end
Mark Detection results;
Compared with region candidate network method, the method for the present invention can significantly improve the recall rate of Small object.
To after Small object, deviate early warning for Small object with Small object equation being obtained based on the method for optimization.
Description of the drawings
Fig. 1 is the FB(flow block) of training stage in Small object method for quick of the present invention;
Fig. 2 is the comparison diagram of region candidate network method and the candidate frame generting machanism of the inventive method;Wherein, a is area
The candidate frame generting machanism of domain candidate network;B is candidate frame generting machanism of the present invention using network;
Fig. 3 is the comparison diagram of VGG16 networks and the inventive method using network;Wherein, a is VGG16 networks;B is this
The network of bright employing;
Fig. 4 is the FB(flow block) of test phase in Small object method for quick of the present invention;
Fig. 5 is that the testing result of the embodiment of the present invention compares exemplary plot;Wherein, testing results of a for adaboost methods
Figure;B is the testing result figure of common convolutional neural networks;Testing result figures of the c for the inventive method.
Specific embodiment
In order to more specifically describe the present invention, below in conjunction with the accompanying drawings and specific embodiment is to technical scheme
It is described in detail.
Using the present invention based on depth convolutional neural networks Small object method for quick detect Small object can be
Lane line, pedestrian, the small-scale obstacle thing, traffic mark symbol etc., in the present embodiment, Small object selects lane line, using base of the present invention
The process detected to lane line in the Small object method for quick of depth convolutional neural networks includes training and tests two
Individual part.
As shown in figure 1, being specifically included using the training part detected to lane line by the inventive method:
Step 1:Imageing sensor is set up, acquisition need to extract the coloured image of lane line;For improve algorithm robustness and
Adaptability, the coloured image of the lane line of collection include various road conditions, weather, block and the situation under illumination.
Step 2:Training data is cleaned, dimmer appearance and the invalid lane line being blocked is removed, to improve instruction
Practice the efficiency of the multiformity and training process of data.
Step 3:The lane line delineator in coloured image is fixed using own calibration tool.
Step 4:Training data is extended, enough training datas are obtained, to overcome the training data of manual demarcation
A limited number of shortcomings, specially:Instrument is intercepted using own sample, according to the track line profile that own calibration tool is given,
Complete lane line is intercepted into into blockage one by one from top to bottom, each blockage is the tiny area centered on lane line
Domain, the ratio of width to height of blockage arbitrarily can be chosen.Positive sample of the blockage on these lane lines as training, in coloured image
Negative sample of the blockage in other regions as training.
Step 5:Network is trained using the training data of pretreatment, obtains lane line sorter model;To accelerate
Calculate, using GPU parallel computations, training process is divided into pre-training stage and fine setting stage, specially:
The pre-training stage:VGG16 networks are obtained using ImageNet database trainings;
The fine setting stage:Based on VGG16 networks, network training is carried out using the training data of pretreatment, obtain track
Line sorter model:
First, select sliding window on the convolution characteristic pattern of last shared convolutional layer (5-3 convolutional layers) of VGG16 networks
As candidate frame, can so make each position on characteristic pattern produce the candidate frame of different sizes, and limit candidate frame
The ratio of width to height, with limit only have specific target be detected and position;Using using half-pixel accuracy sliding window, can so solve
Certainly output characteristic figure precision is too low to the problem for being accurately positioned target, and increased candidate frame corresponding to true frame
Probability.The method is as shown in Figure 2 with the contrast of the candidate frame generting machanism of region candidate network method.
Then, for small target deteection the characteristics of, adjusts network structure, as lane line is small in coloured image
Target, the perception domain of neutral net should reduce, and otherwise excessive environmental information will be transfused to network, and this environmental information is unfavorable
In the detection of localized mass (lane line).
To reduce the perception domain of neutral net, on the basis of VGG16 network structures, through procedure below, this is obtained
Network configuration employed in bright method:
The pond layer 5 in VGG16 network structures is deleted, retains other convolutional layers and pond layer, to utilize to the full extent
Conventional network resources;
Increase convolutional layer 6-1 of the convolution kernel for 3x3, to adjust the visible range of network;
It is that 1x1 convolutional layers replace all of full articulamentum using two convolution kernels, i.e., using convolutional layer 6-2 and convolutional layer 6-
3 replace full articulamentum 1, full articulamentum 2 and full articulamentum 3, to realize across passage interaction and information integration, decrease or increase dimension
The input of degree (quantity of convolution kernel passage) and process arbitrary dimension.This method network structure is right with VGG16 network structures
It is more as shown in Figure 3 than scheming.
Finally, network training is carried out using improved VGG16 network structures using selected candidate frame, obtains Small object point
Class device model.Training parameter can be by following selection:
Stochastic gradient descent method (Stochastic Gradient Descent, the SGD) batch processing that iteration is used every time
(mini-batch) it is made up of 256 samples that 2 coloured images are randomly selected.If target candidate frame and the true frame of target
Target candidate frame is considered as positive sample, is otherwise considered as negative more than 0.5 by overlapping rate (Intersection-over-Union, IoU)
Sample.
As the sample major part included in data base is all to be easy to the sample for detecting, the more difficult inspection of only a small amount of sample
Survey, use the method that (Online Hard Example Mining, OHEM) is excavated based on online hardly possible example in the process of the present invention
To improve the effectiveness of training.For each SGD iteration, convolution characteristic pattern is given, read-only network performs a forward direction and transmits simultaneously
The loss cost of all input samples is calculated, the sample of more difficult detection can be selected by losing cost value sequence, then only
The forward direction of network is carried out with these more difficult samples and is transmitted backward, and accumulation gradient value is sent into network and is trained.
The index of a candidate frame in one batch processing of hypothesis is i, piIt is classification confidence that network class layer is obtained,
If this candidate frame is positive sample (lane line block), classification actual value ciFor 1, otherwise ciFor 0, Pi=(Px,Py,Pw,Ph) be
The position of candidate frame and scale parameter, (Px,Py) be candidate frame center, (Pw,Ph) it is the wide and high of candidate frame, Gi=(Gx,
Gy,Gw,Gh) be true frame position and scale parameter, (Gx,Gy) be true frame center, (Gw,Gh) be true frame width and
It is high;Lcls(pi,ci) it is target classification error, Lreg(Pi,Gi) be target predicted position error, whole loss function can be with table
It is shown as:
Wherein, Ncls、NregIt is normalization coefficient, α is scale factor, balances two kinds of error functions to obtain two kinds of error letters
The suitable ratio of number.
Target classification error LclsIt is the log error (log loss) of two class targets (lane line block/non-lane line block):
It is assumed that candidate frame is represented as Pi, the corresponding character representation on convolution characteristic pattern is Φ5(Pi).Similarly, mesh
Mark true frame and be expressed as Gi.The present invention needs to find a kind of mapping f so thatAnd meetObtain target
The estimation of true frameIf candidate frame is close to true frame, then f is represented by a kind of linear regression model (LRM).Hypothesis needs to learn
The parameter of habit is W*, it means that if candidate frame is converted to d (Pi)=[dx(Pi),dy(Pi),dw(Pi),dh(Pi)], i.e., it is right
Translation transformation (the d that candidate frame is carried outx(Pi),dy(Pi)) and change of scale (dw(Pi),dh(Pi)) meet
Meanwhile, actual translational movement (tx,ty) and scaling (tw,th) be expressed as:
tx=(Gx-Px)/Pw
ty=(Gy-Py)/Ph
tw=log (Gw/Pw)
th=log (Gh/Ph)
The position of candidate target can be finely adjusted by a convolutional layer, and the loss function being optimized is:
Wherein, Part I is prediction loss function, and Part II is regularization term, and λ is used for the shadow for controlling regularization term
Ring the factor, W*Can be obtained by gradient descent method.
As shown in figure 4, the part of detecting detected to lane line using the inventive method is included:
Step 1:Imageing sensor is set up, acquisition need to extract the coloured image of lane line.
Step 2:Given test image, the network that training is obtained perform a forward direction transmission, and each sliding window is through net
Network obtains low-dimensional vector (corresponding to the number of filter of last 1 convolutional layer), and this low-dimensional vector is imported into two convolution
Layer:Classification layer (cls) obtains the classification confidence of 2 class targets (lane line block/non-lane line block), and alignment layers (reg) are waited
Position and the dimensional information of frame are selected, the testing result of lane line discreet region is as shown in the little square frame in Fig. 5 (c).
Step 3:All of lane line block obtains lane line equation based on optimization method.Obtaining lane line discreet region
Afterwards, the central point of each discreet region is taken, and lane line equation is described with conic section:
Y=a0+a1x+a2x2
Wherein:(x, y) is the coordinate of the central point of lane line discreet region, can solve system of conics with optimization method
Number.
As shown in figure 5, detecting to lane line using the method for the present invention, adaboost methods are compared, significantly can be carried
The recall rate of high lane line;The depth convolutional neural networks method classified based on sliding window is compared, detection efficiency and inspection can be improved
Survey accuracy rate.
Above-described specific embodiment has been described in detail to technical scheme and beneficial effect, Ying Li
Solution is to the foregoing is only presently most preferred embodiment of the invention, is not limited to the present invention, all principle models in the present invention
Interior done any modification, supplement and equivalent etc. are enclosed, be should be included within the scope of the present invention.
Claims (4)
1. a kind of Small object method for quick based on depth convolutional neural networks, concretely comprises the following steps:
(1) gather training data and pretreatment is carried out to training data;
(2) network is trained using the training data of pretreatment, obtains Small object sorter model:
The pre-training stage:VGG16 networks are obtained using ImageNet database trainings;
The fine setting stage:Based on VGG16 networks, network training is carried out using the training data of pretreatment, obtain Small object point
Class device model;
(3) collecting test data pretreatment is carried out to test data, obtains the coloured image of Small object;
(4) test image is given, the network that training is obtained performs a forward direction transmission, and each sliding window obtains little through network
The position of the classification confidence and candidate frame of target or non-Small object and dimensional information, its classification confidence and fixed threshold are entered
Row compares, and obtains the small target deteection result of high-class confidence level;
(5) Small object equation is obtained based on optimization method using the small target deteection result of all high confidence levels;
The Small object of high-class confidence level herein is the Small object that classification confidence exceedes fixed threshold.
2. Small object method for quick according to claim 1 based on depth convolutional neural networks, it is characterised in that:Institute
The step of stating (1) concretely comprises the following steps:
(1-1) imageing sensor is set up, acquisition need to extract the coloured image of Small object;
(1-2) training data is cleaned, removes dimmer appearance and the invalid targets being blocked;
(1-3) the Small object delineator in coloured image is fixed using own calibration tool;
(1-4) training data is extended, obtains enough training datas.
3. Small object method for quick according to claim 1 based on depth convolutional neural networks, it is characterised in that:It is micro-
Tune stage concretely comprises the following steps:
(1) sliding window is selected on the convolution characteristic pattern of VGG16 networks last shared convolutional layers as candidate frame, and adopt
Half-pixel accuracy sliding window;
(2) on the basis of VGG16 network structures, the 5th pond layer is deleted, retains other convolutional layers and pond layer;Increase by one
Convolutional layer of the convolution kernel for 3x3;It is all of full articulamentum in 1x1 convolutional layer alternate networks with two convolution kernels, is improved
VGG16 network structures;
(3) network training is carried out using improved VGG16 network structures using the candidate frame selected, obtains Small object grader mould
Type.
4. Small object method for quick according to claim 1 based on depth convolutional neural networks, it is characterised in that:Step
Suddenly calculated using GPU in (2).
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611129536.9A CN106599827A (en) | 2016-12-09 | 2016-12-09 | Small target rapid detection method based on deep convolution neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611129536.9A CN106599827A (en) | 2016-12-09 | 2016-12-09 | Small target rapid detection method based on deep convolution neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106599827A true CN106599827A (en) | 2017-04-26 |
Family
ID=58597994
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611129536.9A Pending CN106599827A (en) | 2016-12-09 | 2016-12-09 | Small target rapid detection method based on deep convolution neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106599827A (en) |
Cited By (46)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107154021A (en) * | 2017-05-03 | 2017-09-12 | 西安电子科技大学 | Image super-resolution method based on deep layer thresholding 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 |
CN107341517A (en) * | 2017-07-07 | 2017-11-10 | 哈尔滨工业大学 | The multiple dimensioned wisp detection method of Fusion Features between a kind of level based on deep learning |
CN107491733A (en) * | 2017-07-19 | 2017-12-19 | 南京农业大学 | A kind of chrysanthemum recognition methods based on deep neural network |
CN107563446A (en) * | 2017-09-05 | 2018-01-09 | 华中科技大学 | A kind of micro OS object detection method |
CN107844790A (en) * | 2017-11-15 | 2018-03-27 | 上海捷售智能科技有限公司 | A kind of vegetable identification and POS and method based on image recognition |
CN107871162A (en) * | 2017-11-16 | 2018-04-03 | 厦门美图之家科技有限公司 | A kind of image processing method and mobile terminal based on convolutional neural networks |
CN107944860A (en) * | 2017-11-15 | 2018-04-20 | 上海捷售智能科技有限公司 | A kind of bakery identification and cash register system and method based on neutral net |
CN108009528A (en) * | 2017-12-26 | 2018-05-08 | 广州广电运通金融电子股份有限公司 | Face authentication method, device, computer equipment and storage medium based on Triplet Loss |
CN108122003A (en) * | 2017-12-19 | 2018-06-05 | 西北工业大学 | A kind of Weak target recognition methods based on deep neural network |
CN108171212A (en) * | 2018-01-19 | 2018-06-15 | 百度在线网络技术(北京)有限公司 | For detecting the method and apparatus of target |
CN108230354A (en) * | 2017-05-18 | 2018-06-29 | 深圳市商汤科技有限公司 | Target following, network training method, device, electronic equipment and storage medium |
CN108229524A (en) * | 2017-05-25 | 2018-06-29 | 北京航空航天大学 | A kind of chimney and condensing tower detection method based on remote sensing images |
CN108416378A (en) * | 2018-02-28 | 2018-08-17 | 电子科技大学 | A kind of large scene SAR target identification methods based on deep neural network |
CN108520273A (en) * | 2018-03-26 | 2018-09-11 | 天津大学 | A kind of quick detection recognition method of dense small item based on target detection |
CN108564601A (en) * | 2018-05-07 | 2018-09-21 | 北京禾泽方圆智能科技有限公司 | A kind of fruit recognition and tracking method and system based on deep learning algorithm |
CN108573238A (en) * | 2018-04-23 | 2018-09-25 | 济南浪潮高新科技投资发展有限公司 | A kind of vehicle checking method based on dual network structure |
CN108805198A (en) * | 2018-06-08 | 2018-11-13 | Oppo广东移动通信有限公司 | Image processing method, device, computer readable storage medium and electronic equipment |
CN108985222A (en) * | 2018-07-12 | 2018-12-11 | 天津艾思科尔科技有限公司 | A kind of deep learning network model and system for making and receiving calls identification |
CN109002745A (en) * | 2017-06-06 | 2018-12-14 | 武汉小狮科技有限公司 | A kind of lane line real-time detection method based on deep learning and tracking technique |
CN109059954A (en) * | 2018-06-29 | 2018-12-21 | 广东星舆科技有限公司 | The method and system for supporting high-precision map lane line real time fusion to update |
CN109389046A (en) * | 2018-09-11 | 2019-02-26 | 昆山星际舟智能科技有限公司 | Round-the-clock object identification and method for detecting lane lines for automatic Pilot |
CN109446919A (en) * | 2018-09-30 | 2019-03-08 | 贵州大学 | A kind of vision lane keeping method based on end-to-end study |
CN109558902A (en) * | 2018-11-20 | 2019-04-02 | 成都通甲优博科技有限责任公司 | A kind of fast target detection method |
CN109583322A (en) * | 2018-11-09 | 2019-04-05 | 长沙小钴科技有限公司 | A kind of recognition of face depth network training method and system |
CN109598339A (en) * | 2018-12-07 | 2019-04-09 | 电子科技大学 | A kind of vehicle attitude detection method based on grid convolutional network |
CN109598268A (en) * | 2018-11-23 | 2019-04-09 | 安徽大学 | A kind of RGB-D well-marked target detection method based on single flow depth degree network |
CN109624850A (en) * | 2017-10-05 | 2019-04-16 | 斯特拉德视觉公司 | Monitor the method for the blind spot of vehicle and the blind spot monitoring device using this method |
CN109697449A (en) * | 2017-10-20 | 2019-04-30 | 杭州海康威视数字技术股份有限公司 | A kind of object detection method, device and electronic equipment |
CN109829390A (en) * | 2019-01-09 | 2019-05-31 | 浙江新再灵科技股份有限公司 | A kind of elevator intelligent scheduling system and method based on deep learning |
CN109919135A (en) * | 2019-03-27 | 2019-06-21 | 华瑞新智科技(北京)有限公司 | Behavioral value method, apparatus based on deep learning |
CN110210603A (en) * | 2019-06-10 | 2019-09-06 | 长沙理工大学 | Counter model construction method, method of counting and the device of crowd |
CN110399868A (en) * | 2018-04-19 | 2019-11-01 | 北京大学深圳研究生院 | A kind of seashore wetland birds detection method |
CN110796144A (en) * | 2019-11-08 | 2020-02-14 | 北京字节跳动网络技术有限公司 | License plate detection method, device, equipment and storage medium |
CN110796048A (en) * | 2019-10-18 | 2020-02-14 | 武汉大学 | Ship target real-time detection method based on deep neural network |
CN111142091A (en) * | 2020-01-10 | 2020-05-12 | 清华大学 | Automatic driving system laser radar online calibration method fusing vehicle-mounted information |
CN111209829A (en) * | 2019-12-31 | 2020-05-29 | 浙江大学 | Vision-based mobile visual body static medium and small scale target identification method |
CN111439259A (en) * | 2020-03-23 | 2020-07-24 | 成都睿芯行科技有限公司 | Agricultural garden scene lane deviation early warning control method and system based on end-to-end convolutional neural network |
CN111507226A (en) * | 2020-04-10 | 2020-08-07 | 北京觉非科技有限公司 | Road image recognition model modeling method, image recognition method and electronic equipment |
CN112132145A (en) * | 2020-08-03 | 2020-12-25 | 深圳大学 | Image classification method and system based on model extended convolutional neural network |
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 |
CN112906626A (en) * | 2021-03-12 | 2021-06-04 | 李辉 | Fault identification method based on artificial intelligence |
CN112949731A (en) * | 2021-03-11 | 2021-06-11 | 江苏禹空间科技有限公司 | Target detection method, device, storage medium and equipment based on multi-expert model |
CN113272829A (en) * | 2019-01-14 | 2021-08-17 | 西门子股份公司 | Hardware accelerator extension for migratory learning-extend/complete training to the edge |
CN113313733A (en) * | 2021-05-19 | 2021-08-27 | 西华大学 | Hierarchical unmanned aerial vehicle target tracking method based on shared convolution |
CN111094952B (en) * | 2017-09-06 | 2021-11-16 | 科磊股份有限公司 | Method and system for defect detection and classification using a unified neural network |
-
2016
- 2016-12-09 CN CN201611129536.9A patent/CN106599827A/en active Pending
Non-Patent Citations (3)
Title |
---|
刘博: "《基于深度学习的目标检测技术的研究》", 《北京邮电大学硕博论文》 * |
汤鹏杰 等: "《并行交叉的深度卷积神经网络模型》", 《中国图象图形学报》 * |
芮挺 等: "《基于深度卷积神经网络的行人检测》", 《计算机工程与应用》 * |
Cited By (62)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107154021B (en) * | 2017-05-03 | 2019-08-13 | 西安电子科技大学 | Image super-resolution method based on deep layer thresholding convolutional neural networks |
CN107154021A (en) * | 2017-05-03 | 2017-09-12 | 西安电子科技大学 | Image super-resolution method based on deep layer thresholding convolutional neural networks |
CN108230354B (en) * | 2017-05-18 | 2022-05-10 | 深圳市商汤科技有限公司 | Target tracking method, network training method, device, electronic equipment and storage medium |
CN108230354A (en) * | 2017-05-18 | 2018-06-29 | 深圳市商汤科技有限公司 | Target following, network training method, device, electronic equipment and storage medium |
CN108229524A (en) * | 2017-05-25 | 2018-06-29 | 北京航空航天大学 | A kind of chimney and condensing tower detection method based on remote sensing images |
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 |
CN109002745A (en) * | 2017-06-06 | 2018-12-14 | 武汉小狮科技有限公司 | A kind of lane line real-time detection method based on deep learning and tracking technique |
CN107341517A (en) * | 2017-07-07 | 2017-11-10 | 哈尔滨工业大学 | The multiple dimensioned wisp detection method of Fusion Features between a kind of level based on deep learning |
CN107341517B (en) * | 2017-07-07 | 2020-08-11 | 哈尔滨工业大学 | Multi-scale small object detection method based on deep learning inter-level feature fusion |
CN107491733A (en) * | 2017-07-19 | 2017-12-19 | 南京农业大学 | A kind of chrysanthemum recognition methods based on deep neural network |
CN107563446B (en) * | 2017-09-05 | 2020-08-18 | 华中科技大学 | Target detection method for micro-operation system |
CN107563446A (en) * | 2017-09-05 | 2018-01-09 | 华中科技大学 | A kind of micro OS object detection method |
CN111094952B (en) * | 2017-09-06 | 2021-11-16 | 科磊股份有限公司 | Method and system for defect detection and classification using a unified neural network |
CN109624850A (en) * | 2017-10-05 | 2019-04-16 | 斯特拉德视觉公司 | Monitor the method for the blind spot of vehicle and the blind spot monitoring device using this method |
CN109697449A (en) * | 2017-10-20 | 2019-04-30 | 杭州海康威视数字技术股份有限公司 | A kind of object detection method, device and electronic equipment |
CN107944860A (en) * | 2017-11-15 | 2018-04-20 | 上海捷售智能科技有限公司 | A kind of bakery identification and cash register system and method based on neutral net |
CN107844790A (en) * | 2017-11-15 | 2018-03-27 | 上海捷售智能科技有限公司 | A kind of vegetable identification and POS and method based on image recognition |
CN107871162A (en) * | 2017-11-16 | 2018-04-03 | 厦门美图之家科技有限公司 | A kind of image processing method and mobile terminal based on convolutional neural networks |
CN107871162B (en) * | 2017-11-16 | 2020-06-05 | 厦门美图之家科技有限公司 | Image processing method based on convolutional neural network and mobile terminal |
CN108122003A (en) * | 2017-12-19 | 2018-06-05 | 西北工业大学 | A kind of Weak target recognition methods based on deep neural network |
CN108009528B (en) * | 2017-12-26 | 2020-04-07 | 广州广电运通金融电子股份有限公司 | Triple Loss-based face authentication method and device, computer equipment and storage medium |
WO2019128367A1 (en) * | 2017-12-26 | 2019-07-04 | 广州广电运通金融电子股份有限公司 | Face verification method and apparatus based on triplet loss, and computer device and storage medium |
CN108009528A (en) * | 2017-12-26 | 2018-05-08 | 广州广电运通金融电子股份有限公司 | Face authentication method, device, computer equipment and storage medium based on Triplet Loss |
CN108171212A (en) * | 2018-01-19 | 2018-06-15 | 百度在线网络技术(北京)有限公司 | For detecting the method and apparatus of target |
CN108416378B (en) * | 2018-02-28 | 2020-04-14 | 电子科技大学 | Large-scene SAR target recognition method based on deep neural network |
CN108416378A (en) * | 2018-02-28 | 2018-08-17 | 电子科技大学 | A kind of large scene SAR target identification methods based on deep neural network |
CN108520273A (en) * | 2018-03-26 | 2018-09-11 | 天津大学 | A kind of quick detection recognition method of dense small item based on target detection |
CN110399868A (en) * | 2018-04-19 | 2019-11-01 | 北京大学深圳研究生院 | A kind of seashore wetland birds detection method |
CN110399868B (en) * | 2018-04-19 | 2022-09-09 | 北京大学深圳研究生院 | Coastal wetland bird detection method |
CN108573238A (en) * | 2018-04-23 | 2018-09-25 | 济南浪潮高新科技投资发展有限公司 | A kind of vehicle checking method based on dual network structure |
CN108564601A (en) * | 2018-05-07 | 2018-09-21 | 北京禾泽方圆智能科技有限公司 | A kind of fruit recognition and tracking method and system based on deep learning algorithm |
CN108564601B (en) * | 2018-05-07 | 2021-03-02 | 北京禾泽方圆智能科技有限公司 | Fruit identification tracking method and system based on deep learning algorithm |
CN108805198A (en) * | 2018-06-08 | 2018-11-13 | Oppo广东移动通信有限公司 | Image processing method, device, computer readable storage medium and electronic equipment |
CN109059954A (en) * | 2018-06-29 | 2018-12-21 | 广东星舆科技有限公司 | The method and system for supporting high-precision map lane line real time fusion to update |
CN108985222B (en) * | 2018-07-12 | 2024-02-20 | 天津艾思科尔科技有限公司 | Deep learning network model and system for recognition of incoming calls |
CN108985222A (en) * | 2018-07-12 | 2018-12-11 | 天津艾思科尔科技有限公司 | A kind of deep learning network model and system for making and receiving calls identification |
CN109389046A (en) * | 2018-09-11 | 2019-02-26 | 昆山星际舟智能科技有限公司 | Round-the-clock object identification and method for detecting lane lines for automatic Pilot |
CN109446919A (en) * | 2018-09-30 | 2019-03-08 | 贵州大学 | A kind of vision lane keeping method based on end-to-end study |
CN109583322A (en) * | 2018-11-09 | 2019-04-05 | 长沙小钴科技有限公司 | A kind of recognition of face depth network training method and system |
CN109583322B (en) * | 2018-11-09 | 2020-07-17 | 长沙小钴科技有限公司 | Face recognition deep network training method and system |
CN109558902A (en) * | 2018-11-20 | 2019-04-02 | 成都通甲优博科技有限责任公司 | A kind of fast target detection method |
CN109598268A (en) * | 2018-11-23 | 2019-04-09 | 安徽大学 | A kind of RGB-D well-marked target detection method based on single flow depth degree network |
CN109598339A (en) * | 2018-12-07 | 2019-04-09 | 电子科技大学 | A kind of vehicle attitude detection method based on grid convolutional network |
CN109829390A (en) * | 2019-01-09 | 2019-05-31 | 浙江新再灵科技股份有限公司 | A kind of elevator intelligent scheduling system and method based on deep learning |
CN113272829A (en) * | 2019-01-14 | 2021-08-17 | 西门子股份公司 | Hardware accelerator extension for migratory learning-extend/complete training to the edge |
CN109919135A (en) * | 2019-03-27 | 2019-06-21 | 华瑞新智科技(北京)有限公司 | Behavioral value method, apparatus based on deep learning |
CN110210603A (en) * | 2019-06-10 | 2019-09-06 | 长沙理工大学 | Counter model construction method, method of counting and the device of crowd |
CN110796048A (en) * | 2019-10-18 | 2020-02-14 | 武汉大学 | Ship target real-time detection method based on deep neural network |
CN110796048B (en) * | 2019-10-18 | 2022-11-04 | 武汉大学 | Ship target real-time detection method based on deep neural network |
CN110796144A (en) * | 2019-11-08 | 2020-02-14 | 北京字节跳动网络技术有限公司 | License plate detection method, device, equipment and storage medium |
CN111209829A (en) * | 2019-12-31 | 2020-05-29 | 浙江大学 | Vision-based mobile visual body static medium and small scale target identification method |
CN111142091A (en) * | 2020-01-10 | 2020-05-12 | 清华大学 | Automatic driving system laser radar online calibration method fusing vehicle-mounted information |
CN111439259A (en) * | 2020-03-23 | 2020-07-24 | 成都睿芯行科技有限公司 | Agricultural garden scene lane deviation early warning control method and system based on end-to-end convolutional neural network |
CN111507226B (en) * | 2020-04-10 | 2023-08-11 | 北京觉非科技有限公司 | Road image recognition model modeling method, image recognition method and electronic equipment |
CN111507226A (en) * | 2020-04-10 | 2020-08-07 | 北京觉非科技有限公司 | Road image recognition model modeling method, image recognition method and electronic equipment |
CN112132145A (en) * | 2020-08-03 | 2020-12-25 | 深圳大学 | Image classification method and system based on model extended convolutional neural network |
CN112132145B (en) * | 2020-08-03 | 2023-08-01 | 深圳大学 | Image classification method and system based on model extended convolutional neural network |
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 |
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 |
CN112949731A (en) * | 2021-03-11 | 2021-06-11 | 江苏禹空间科技有限公司 | Target detection method, device, storage medium and equipment based on multi-expert model |
CN112906626A (en) * | 2021-03-12 | 2021-06-04 | 李辉 | Fault identification method based on artificial intelligence |
CN113313733A (en) * | 2021-05-19 | 2021-08-27 | 西华大学 | Hierarchical unmanned aerial vehicle target tracking method based on shared convolution |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106599827A (en) | Small target rapid detection method based on deep convolution neural network | |
CN108985194B (en) | Intelligent vehicle travelable area identification method based on image semantic segmentation | |
EP3614308B1 (en) | Joint deep learning for land cover and land use classification | |
CN105957342B (en) | Track grade road plotting method and system based on crowdsourcing space-time big data | |
CN109932730B (en) | Laser radar target detection method based on multi-scale monopole three-dimensional detection network | |
CN106446914A (en) | Road detection based on superpixels and convolution neural network | |
CN107368787A (en) | A kind of Traffic Sign Recognition algorithm that application is driven towards depth intelligence | |
CN109447033A (en) | Vehicle front obstacle detection method based on YOLO | |
CN108830188A (en) | Vehicle checking method based on deep learning | |
CN109902806A (en) | Method is determined based on the noise image object boundary frame of convolutional neural networks | |
CN108171112A (en) | Vehicle identification and tracking based on convolutional neural networks | |
CN110197215A (en) | A kind of ground perception point cloud semantic segmentation method of autonomous driving | |
CN107818302A (en) | Non-rigid multiple dimensioned object detecting method based on convolutional neural networks | |
CN106250812A (en) | A kind of model recognizing method based on quick R CNN deep neural network | |
Smolyakov et al. | Self-driving car steering angle prediction based on deep neural network an example of CarND udacity simulator | |
Tümen et al. | Intersections and crosswalk detection using deep learning and image processing techniques | |
CN106169188A (en) | A kind of method for tracing object based on the search of Monte Carlo tree | |
CN106873585A (en) | One kind navigation method for searching, robot and system | |
CN108537102A (en) | High Resolution SAR image classification method based on sparse features and condition random field | |
CN104866869A (en) | Time sequence SAR (Synthetic Aperture Radar) image classification method on the basis of distribution difference and incremental learning | |
CN110135502A (en) | A kind of image fine granularity recognition methods based on intensified learning strategy | |
CN104331716A (en) | SVM active learning classification algorithm for large-scale training data | |
CN110009648A (en) | Trackside image Method of Vehicle Segmentation based on depth Fusion Features convolutional neural networks | |
Xing et al. | Traffic sign recognition using guided image filtering | |
CN105654516A (en) | Method for detecting small moving object on ground on basis of satellite image with target significance |
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 | ||
TA01 | Transfer of patent application right | ||
TA01 | Transfer of patent application right |
Effective date of registration: 20190227 Address after: 310018 18 learning street, Xiasha Higher Education Park, Hangzhou, Zhejiang Applicant after: Zhejiang Gongshang University Applicant after: Zhejiang Uniview Technology Co., Ltd. Address before: 310018 18 learning street, Xiasha Higher Education Park, Hangzhou, Zhejiang Applicant before: Zhejiang Gongshang University |
|
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170426 |