CN107818302A - Non-rigid multiple dimensioned object detecting method based on convolutional neural networks - Google Patents
Non-rigid multiple dimensioned object detecting method based on convolutional neural networks Download PDFInfo
- Publication number
- CN107818302A CN107818302A CN201710989778.3A CN201710989778A CN107818302A CN 107818302 A CN107818302 A CN 107818302A CN 201710989778 A CN201710989778 A CN 201710989778A CN 107818302 A CN107818302 A CN 107818302A
- Authority
- CN
- China
- Prior art keywords
- target
- detection
- network
- rigid
- multiple dimensioned
- 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
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of non-rigid multiple dimensioned object detecting method based on convolutional neural networks, projected depth network first, and with certain representational infrared pedestrian's data set cvc 09, cvc 14 (has the Small object of about 20x30 pixels, also there is the big target pedestrian of closer distance), and the Small object pedestrian data of oneself mark carry out the depth network learning and training of target detection, are then based on the different detection model of Hardware platform design, and assessed.Designed network, uses 6 candidate frames with 7x7 mesh generations and for each grid, completes target area nomination, realizes that pedestrian target detects by Classification and Identification and position regression analysis.Neutral net is used in the design, adds the ability in feature extraction of detection model.In view of the thought of mesh generation, realizes the real-time of detection, data set and multiple dimensioned training with different target yardstick, increases candidate frame quantity, the final detectability for strengthening Small object.
Description
Technical field
The present invention relates to the technical field that object detects in real time, and in particular to a kind of based on the non-rigid of convolutional neural networks
The detection method of multiple dimensioned object.
Background technology
As shown in figure 3, neuron is the mathematical modeling of mimic biology nerve construction, different inputs (such as different pixel
Point) summed by weighting (convolution algorithm in such as convolutional neural networks), it is same at node to cross nonlinear function (activation primitive)
Conversion obtain the output of the neurode.As shown in figure 4, numerous neurodes can form nervous layer, numerous nervous layers can be taken
Build neutral net.Neutral net is shown to training because input is by multilayered nonlinear conversion (nesting of nonlinear function)
The powerful capability of fitting of data (learning ability), so as to show good predictive ability.
For neural network model, there are training and two stages of test.Training is to train picture by inputting, and what is obtained is defeated
Go out and contrasted with standard output (mark), its error is embodied by loss function, reduces loss function by Optimal Parameters.Survey
Examination is then the assessment of the model performance to training, and the generalization ability of model can be evaluated by test result.
In Detection task, network model and three main plates:Region nomination (is chosen target Probability Area, often led to
Cross sliding window and selective search), but sliding window is a kind of method of exhaustion in fact in conventional method, and there are many redundancies in process, selectivity
Search then needs to take considerable time.Mesh generation, Mei Gequ are taken in detection end to end (positioning, identification integration)
Domain each determines whether target, and then synthesis draws end product.
The content of the invention
The invention solves technical problem to be:In the target of mobile device and small device in real time detection, due to not
Changeable with non-rigid multiscale target form, angle in scene, traditional detection method universality and robustness be not high.Using god
In scheme through network and grid division range searching, solve the problems, such as universality and robustness to a certain extent, while again
It ensure that the real-time of detection.
The present invention solve the technical scheme that uses of above-mentioned technical problem for:It is a kind of non-rigid more based on convolutional neural networks
Nanoscale object (such as infrared pedestrian image) detection method, construct and detected with the neutral net of powerful feature extraction and learning ability
System so that the abstract characteristics in data focusing study to non-rigid object different scale;It is again that program portable is flat to exploitation
Platform, by hardware device input picture and carry out the real-time detection of target.
Further, there is suitable and accurate calibration data set (supervised learning), due to neural metwork training data set
The middle training dataset lacked for smaller infrared target, and at this stage there is bottle in neutral net in terms of detection is compared with Small object
Neck, by the way that the data set of non-standard configuration is found, screened on network, write script and data set mark is converted to the available number of network
According on this basis, by shooting and marking, having expanded data, formd trainable data set.
Further, there is powerful ability in feature extraction and the model of learning ability, due to non-rigid targets (such as:People)
Diversity ratio is larger in different angle posture, and target has different scale, and conventional method performance is restricted under Nonuniform field scape,
The process employs the disaggregated model with powerful ability in feature extraction --- depth network, the model by different Internets it
Between intensive connection so that each layer characteristic use is abundant.
Further, real-time detection can be realized on a hardware platform.By extracting the feature of non-rigid targets, then by
Recurrence layer in neutral net and classification layer carry out target identification and location estimation to candidate frame, and process is simultaneously end to end for this
The identification and positioning of target are completed, is divided by area grid and aforesaid operations is implemented to each region, simplified and searched possibility
The process of target, target and non-targeted judgement are disposably completed, so as to ensure that the real-time of detection, last is non-very big
Value suppresses the detection output of place to go redundancy.
The present invention compared with prior art the advantages of be:
(1) feature of the present invention is the real-time that network structure ensure that detection end to end.Compared with universal model, use
More average pondization reduction operational data, with the thought of dense connection, is carried in the case where not increasing network depth
The performance of high network of network, while the framework of the neutral net is succinct, has high portability so that the network energy trained
Directly apply in imaging system target detection, realize the landing of detection technique, realize small device have it is intelligent
Detection.
(2) method that the present invention employs augmentation on the training data, while (target is full of the positive sample in database
Whole pictures) regard the training data of target detection, further directed to property the feature of learning objective in itself.
Brief description of the drawings
Fig. 1 is the flow chart that the inventive method network training process is realized;
Fig. 2 is the flow chart that the inventive method network detection process is realized;
Fig. 3 is the mathematical modeling of neuron node in neutral net;
The basic model of Fig. 4 neutral nets;
Fig. 5 is dense link model;
Fig. 6 is mesh generation input picture schematic diagram;
Fig. 7 signals network class after feature extraction identifies the net region where target;
Fig. 8 represents to return the region where target with candidate frame;
Fig. 9 is the final output after non-maxima suppression redundancy detection frame;
Figure 10 is that loss function changes chart in training process.
Embodiment
Below in conjunction with the accompanying drawings and embodiment further illustrates the present invention.
As shown in figure 1, a kind of non-rigid multiple dimensioned object detecting method based on convolutional neural networks of the present invention, this method
Comprise the following steps that:
Step 1, the image with mark;
Fig. 1 represents whole training process.Training data (image to be marked) with mark is inputted into neutral net, led to
Cross propagated forward and obtain output result.By output result and mark be compared to obtain error (weighed with loss function, training
Target is to reduce loss function), by the error gradient backpropagation undated parameter of neutral net, so circulation, make loss letter
Number reduces and convergence.
During training, translate, in addition to multiple dimensioned, adjustment picture contrast method except with traditional, also use
The mode of other data augmentation carrys out the increased view data for being used to train, such as:
(1) target in the data set of mark is cut out to come, turns into the fresh target that callout box is view picture figure size.
(2) next target image random combine will be cut out to be placed on new background image, turns into new training sample.
Step 2, with 7 × 7 mesh generations into different zones;
Mesh generation except carrying out 7 × 7 to input picture, division size can be selected according to specific tasks.Such as
For small compared with Small object in infrared pedestrian detection, using the above method., can be to nerve if detecting the fish of underwater large volume
In network characteristic layer rearward carry out mesh generation (because in a network, what convolutional layer above extracted be more edge and
Contour feature, below be then relatively abstract semantic feature), the information of characteristic layer is imported into the test section of network end-point
Point.
Divide yardstick and characteristic layer selection can the motor-driven selection according to specific tasks, it is few and as far as possible not with characteristic parameter
Premised on the amount of calculation for increasing network.
Step 3, feature extraction is carried out to entire image;
Feature extraction is completed by the convolutional layer in neutral net.The convolution nuclear parameter of neutral net convolutional layer is existing initial
Change, during training, by constantly regulate parameter, reduce loss function, finally obtain the network model trained (now
Parameter by training and change, enable to convolution operation to extract characteristics of image, then propagated forward, detects target).Number
Effect according to augmentation is exactly to increase training sample so that network can learn more abstract, more commonly to target in the training process
Feature.
Step 4, each region is detected with candidate frame;
The effect of candidate frame is that have individual priori to judge detection target.In specific Detection task, using the several of selection
The frame of fixed dimension positions to each regional area, then merges in subsequent steps and removes redundancy, to obtain most
Whole testing result.K mean cluster is then to draw the size of candidate frame according to markup information in training data.Candidate frame
Number can as requested depending on, quantity is more to cause speed to decline, precision improvement.K-means can be utilized to choose candidate frame
Size.
The classification and recurrence of step 5, candidate frame;
Classification and recurrence are by softmax (classification layer) and bbox regressor (frames conventional in neural network model
Return layer) complete.
Step 6, non-maxima suppression;
Non-maxima suppression removes it according to given threshold, the frame of selected threshold above confidence level and the classification of probability
His frame and classification, obtain positional information and classification information to the end.
Target is oriented in step 7, identification, and detection error is obtained according to label;
Detection process just with the neural network model for having trained parameter, is obtained and exported to the end, as shown in Figure 2.
In training process, then by exporting and marking contrast, error successively backpropagation, with gradient descent method undated parameter so that
The output of model approaches mark (reducing loss function value), completes the process of training study, as shown in Figure 1.
The present invention is that the method based on convolutional neural networks feature extraction carries out the non-rigid object of Small Target (such as row
People) detection, i.e., by building convolutional neural networks detection model, import image data base and be trained Optimized model parameter, reach
To good Detection results.Projected depth network first, and with certain representational infrared pedestrian's data set cvc-09,
Cvc-14 (has the Small object of about 20x30 pixels, the big target pedestrian for also having closer distance), and the Small object row of oneself mark
Personal data carries out the depth network learning and training of target detection, is then based on the different detection model of Hardware platform design, and add
To assess.Designed network, multiple candidate frames are used with 7x7 mesh generations and for each grid, target area is completed and carries
Name, realize that pedestrian target detects by Classification and Identification and position regression analysis.Neutral net is used in the design, adds detection
The ability in feature extraction of model.In view of the thought of mesh generation, the real-time of detection is realized, with the number of different target yardstick
According to collection and multiple dimensioned training, increase candidate frame quantity, the final detectability for strengthening Small object.
By the cascade of neutral net different layers characteristic pattern, strengthen the learning ability of network, while use multitiered network
Model approach, ensure have the more network number of plies under have less parameter, facilitate training, simplify calculating, beneficial to
Used under small volume, computational not high platform.
Network dense connection between layers and the combination of end-to-end technology, while have chosen with the method for mean cluster
Multiple (6 candidate frames in such as pedestrian detection) candidate frames, are connected to the network end by detector or transition zone go out, according to making
With one or more detector.Dense connection is as shown in Figure 5.
In Fig. 5, every layer using front layer output as input, for there is a legacy network of L layers, one shared L connects, for
Multitiered network, then there is L (L+1)/2.In upper figure, H4 layers can directly use original input information X0, while before also having used
Layer can so maximize the flowing of information to the information after X0 processing.In back-propagation process, X0 gradient information contains
Loss function directly to X0 derivative, is advantageous to gradient propagation, because dense connection and weights are shared so that in non-rigid object
In (such as pedestrian) detection, there is a network deep enough (more than 300 layers), and parameter (storage size only 60MB or so) few enough.
The nomination technology of mesh generation, data set is in combination, then good effect can be obtained above in detection, and
And it is different according to specific task, can on the different transition zone of multitiered network connecting detection device, it is comprehensive under multiple dimensioned basis
Close testing result.Can be also used in it is single go out connecting detection device (infrared pedestrian detection is such as connected at end) can then obtain compared with
Good result, flow carries out mesh generation as Figure 6-9, to image, after by extracting feature, to the image in each region
Classify with candidate frame and position, finally merge and remove redundancy, obtain detecting position and the classification information of target.
For darknet, hardware development platform employs GTX1080 and accelerates network instruction the neural network framework that the present invention uses
Practice, hardware is Nvidia jetson TX1 using platform.
Installation configuration cuda and cudnn first accelerates storehouse and darknet.On this basis, script is write using python
The labeled data of CVCX series is changed, is marked with lablImg and expands training sample.
By changing web results, network candidate frame is assessed, is drawn in network structure configuration using Route layers
Detector, using mode augmentation training datas such as translation, rotations, suitable batch data amount and data is selected to import number to control
Data are handled up in training process processed, select suitable learning rate and weights attenuation rate etc. to ensure the convergence of optimization process, and
Prevent from converging to locally optimal solution, the upper repetitive exercises of GPU reduce loss function, and partial routine is as shown in Figure 10.
Finally, assess and select most suitable training weights, be transplanted to using platform.
Claims (4)
1. a kind of non-rigid multiple dimensioned object detecting method based on convolutional neural networks, it is characterized in that:Construction has powerful spy
Sign extraction and the neutral net detecting system of learning ability so that in data focusing study to non-rigid object different scale
Abstract characteristics;Again by program portable to development platform, by hardware device input picture and the real-time detection of target is carried out.
2. a kind of non-rigid multiple dimensioned object detecting method based on convolutional neural networks according to claim 1, it is special
Sign is:With suitable and accurate calibration data set, smaller infrared target is directed to due to lacking in neural metwork training data set
Training dataset, and at this stage neutral net detection compared with Small object in terms of bottleneck be present, pass through on network find, screen
The data set of non-standard configuration, write script data set mark be converted to the available data of network, on this basis, by shooting and
Mark, has expanded data, has formd trainable data set.
3. a kind of non-rigid multiple dimensioned object detecting method based on convolutional neural networks according to claim 1, it is special
Sign is:Model with powerful ability in feature extraction and learning ability, due to non-rigid targets in different angle posture difference
It is bigger, and different scale be present in target, conventional method performance is restricted under Nonuniform field scape, and the process employs with strong
Disaggregated model --- the depth network of big ability in feature extraction, the model is by intensive connection between different Internets so that each
Layer characteristic use is abundant.
4. a kind of non-rigid multiple dimensioned object detecting method based on convolutional neural networks according to claim 1, it is special
Sign is:Real-time detection can be realized on a hardware platform;By extracting the feature of non-rigid targets, then by neutral net
Return layer and classification layer carries out target identification and location estimation to candidate frame, process completes the knowledge of target simultaneously end to end for this
Not and positioning, by area grid divide and to each region implement aforesaid operations, simplify searched may target process, one
Secondary property completes target and non-targeted judgement, so as to ensure that the real-time of detection, last non-maxima suppression place to go
The detection output of redundancy.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710989778.3A CN107818302A (en) | 2017-10-20 | 2017-10-20 | Non-rigid multiple dimensioned object detecting method based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710989778.3A CN107818302A (en) | 2017-10-20 | 2017-10-20 | Non-rigid multiple dimensioned object detecting method based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107818302A true CN107818302A (en) | 2018-03-20 |
Family
ID=61608634
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710989778.3A Pending CN107818302A (en) | 2017-10-20 | 2017-10-20 | Non-rigid multiple dimensioned object detecting method based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107818302A (en) |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108520286A (en) * | 2018-04-24 | 2018-09-11 | 青岛科技大学 | Infrared dark dim light small target deteection system based on convolutional Neural and candidate region |
CN108537329A (en) * | 2018-04-18 | 2018-09-14 | 中国科学院计算技术研究所 | A kind of method and apparatus carrying out operation using Volume R-CNN neural networks |
CN108549866A (en) * | 2018-04-12 | 2018-09-18 | 上海海事大学 | Remote sensing aeroplane recognition methods based on intensive convolutional neural networks |
CN108614996A (en) * | 2018-04-04 | 2018-10-02 | 杭州电子科技大学 | A kind of military ships based on deep learning, civilian boat automatic identifying method |
CN108647573A (en) * | 2018-04-04 | 2018-10-12 | 杭州电子科技大学 | A kind of military target recognition methods based on deep learning |
CN108764203A (en) * | 2018-06-06 | 2018-11-06 | 四川大学 | A kind of pedestrian's quantitative analysis and display systems towards urban planning |
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 |
CN109145713A (en) * | 2018-07-02 | 2019-01-04 | 南京师范大学 | A kind of Small object semantic segmentation method of combining target detection |
CN109272050A (en) * | 2018-09-30 | 2019-01-25 | 北京字节跳动网络技术有限公司 | Image processing method and device |
CN109376693A (en) * | 2018-11-22 | 2019-02-22 | 四川长虹电器股份有限公司 | Method for detecting human face and system |
CN109377508A (en) * | 2018-09-26 | 2019-02-22 | 北京字节跳动网络技术有限公司 | Image processing method and device |
CN109493320A (en) * | 2018-10-11 | 2019-03-19 | 苏州中科天启遥感科技有限公司 | Method for extracting remote sensing image road and system, storage medium, electronic equipment based on deep learning |
CN109543683A (en) * | 2018-11-30 | 2019-03-29 | 百度在线网络技术(北京)有限公司 | Image labeling modification method, device, equipment and medium |
CN109583456A (en) * | 2018-11-20 | 2019-04-05 | 西安电子科技大学 | Infrared surface object detection method based on Fusion Features and dense connection |
CN109598197A (en) * | 2018-10-31 | 2019-04-09 | 大连大学 | The design method of hourglass model based on intensive link block |
CN109614957A (en) * | 2019-01-04 | 2019-04-12 | 王卫星 | A kind of railway shunting signals security early warning system |
CN109919934A (en) * | 2019-03-11 | 2019-06-21 | 重庆邮电大学 | A kind of liquid crystal display panel defect inspection method based on the study of multi-source domain depth migration |
CN109948605A (en) * | 2019-02-15 | 2019-06-28 | 北京中电普华信息技术有限公司 | A kind of picture Enhancement Method and device for Small object |
CN110633594A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Target detection method and device |
CN111310831A (en) * | 2020-02-19 | 2020-06-19 | 中国科学院光电技术研究所 | Small target detection network layer-by-layer training method based on deep learning |
CN112153483A (en) * | 2019-06-28 | 2020-12-29 | 腾讯科技(深圳)有限公司 | Information implantation area detection method and device and electronic equipment |
CN112686152A (en) * | 2020-12-30 | 2021-04-20 | 广西慧云信息技术有限公司 | Crop pest and disease identification method with multi-size input and multi-size targets |
CN113033504A (en) * | 2021-05-19 | 2021-06-25 | 广东众聚人工智能科技有限公司 | Multi-scale video anomaly detection method |
CN113076966A (en) * | 2020-01-06 | 2021-07-06 | 字节跳动有限公司 | Image processing method and device, neural network training method and storage medium |
WO2021244079A1 (en) * | 2020-06-02 | 2021-12-09 | 苏州科技大学 | Method for detecting image target in smart home environment |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105844238A (en) * | 2016-03-23 | 2016-08-10 | 乐视云计算有限公司 | Method and system for discriminating videos |
CN106815596A (en) * | 2016-12-08 | 2017-06-09 | 中国银联股份有限公司 | A kind of Image Classifier method for building up and device |
CN106991646A (en) * | 2017-03-28 | 2017-07-28 | 福建帝视信息科技有限公司 | A kind of image super-resolution method based on intensive connection network |
CN107038448A (en) * | 2017-03-01 | 2017-08-11 | 中国科学院自动化研究所 | Target detection model building method |
CN107145845A (en) * | 2017-04-26 | 2017-09-08 | 中山大学 | The pedestrian detection method merged based on deep learning and multi-characteristic points |
-
2017
- 2017-10-20 CN CN201710989778.3A patent/CN107818302A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105844238A (en) * | 2016-03-23 | 2016-08-10 | 乐视云计算有限公司 | Method and system for discriminating videos |
CN106815596A (en) * | 2016-12-08 | 2017-06-09 | 中国银联股份有限公司 | A kind of Image Classifier method for building up and device |
CN107038448A (en) * | 2017-03-01 | 2017-08-11 | 中国科学院自动化研究所 | Target detection model building method |
CN106991646A (en) * | 2017-03-28 | 2017-07-28 | 福建帝视信息科技有限公司 | A kind of image super-resolution method based on intensive connection network |
CN107145845A (en) * | 2017-04-26 | 2017-09-08 | 中山大学 | The pedestrian detection method merged based on deep learning and multi-characteristic points |
Non-Patent Citations (2)
Title |
---|
JOSEPH REDMON等: ""YOLO9000: Better, Faster, Stronger"", 《ARXIV》 * |
ZHIQIANG SHEN等: ""DSOD: Learning Deeply Supervised Object Detectors from Scratch"", 《ARXIV》 * |
Cited By (32)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108614996A (en) * | 2018-04-04 | 2018-10-02 | 杭州电子科技大学 | A kind of military ships based on deep learning, civilian boat automatic identifying method |
CN108647573A (en) * | 2018-04-04 | 2018-10-12 | 杭州电子科技大学 | A kind of military target recognition methods based on deep learning |
CN108549866A (en) * | 2018-04-12 | 2018-09-18 | 上海海事大学 | Remote sensing aeroplane recognition methods based on intensive convolutional neural networks |
CN108537329A (en) * | 2018-04-18 | 2018-09-14 | 中国科学院计算技术研究所 | A kind of method and apparatus carrying out operation using Volume R-CNN neural networks |
CN108520286A (en) * | 2018-04-24 | 2018-09-11 | 青岛科技大学 | Infrared dark dim light small target deteection system based on convolutional Neural and candidate region |
CN108764203A (en) * | 2018-06-06 | 2018-11-06 | 四川大学 | A kind of pedestrian's quantitative analysis and display systems towards urban planning |
CN110633594A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Target detection method and device |
CN109145713A (en) * | 2018-07-02 | 2019-01-04 | 南京师范大学 | A kind of Small object semantic segmentation method of combining target detection |
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 |
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 |
CN109377508A (en) * | 2018-09-26 | 2019-02-22 | 北京字节跳动网络技术有限公司 | Image processing method and device |
CN109272050B (en) * | 2018-09-30 | 2019-11-22 | 北京字节跳动网络技术有限公司 | Image processing method and device |
CN109272050A (en) * | 2018-09-30 | 2019-01-25 | 北京字节跳动网络技术有限公司 | Image processing method and device |
CN109493320A (en) * | 2018-10-11 | 2019-03-19 | 苏州中科天启遥感科技有限公司 | Method for extracting remote sensing image road and system, storage medium, electronic equipment based on deep learning |
CN109493320B (en) * | 2018-10-11 | 2022-06-17 | 苏州中科天启遥感科技有限公司 | Remote sensing image road extraction method and system based on deep learning, storage medium and electronic equipment |
CN109598197A (en) * | 2018-10-31 | 2019-04-09 | 大连大学 | The design method of hourglass model based on intensive link block |
CN109583456A (en) * | 2018-11-20 | 2019-04-05 | 西安电子科技大学 | Infrared surface object detection method based on Fusion Features and dense connection |
CN109583456B (en) * | 2018-11-20 | 2023-04-28 | 西安电子科技大学 | Infrared surface target detection method based on feature fusion and dense connection |
CN109376693A (en) * | 2018-11-22 | 2019-02-22 | 四川长虹电器股份有限公司 | Method for detecting human face and system |
CN109543683A (en) * | 2018-11-30 | 2019-03-29 | 百度在线网络技术(北京)有限公司 | Image labeling modification method, device, equipment and medium |
CN109614957A (en) * | 2019-01-04 | 2019-04-12 | 王卫星 | A kind of railway shunting signals security early warning system |
CN109948605B (en) * | 2019-02-15 | 2021-05-28 | 北京中电普华信息技术有限公司 | Picture enhancement method and device for small target |
CN109948605A (en) * | 2019-02-15 | 2019-06-28 | 北京中电普华信息技术有限公司 | A kind of picture Enhancement Method and device for Small object |
CN109919934A (en) * | 2019-03-11 | 2019-06-21 | 重庆邮电大学 | A kind of liquid crystal display panel defect inspection method based on the study of multi-source domain depth migration |
CN112153483A (en) * | 2019-06-28 | 2020-12-29 | 腾讯科技(深圳)有限公司 | Information implantation area detection method and device and electronic equipment |
CN113076966A (en) * | 2020-01-06 | 2021-07-06 | 字节跳动有限公司 | Image processing method and device, neural network training method and storage medium |
CN113076966B (en) * | 2020-01-06 | 2023-06-13 | 字节跳动有限公司 | Image processing method and device, training method of neural network and storage medium |
CN111310831A (en) * | 2020-02-19 | 2020-06-19 | 中国科学院光电技术研究所 | Small target detection network layer-by-layer training method based on deep learning |
WO2021244079A1 (en) * | 2020-06-02 | 2021-12-09 | 苏州科技大学 | Method for detecting image target in smart home environment |
CN112686152A (en) * | 2020-12-30 | 2021-04-20 | 广西慧云信息技术有限公司 | Crop pest and disease identification method with multi-size input and multi-size targets |
CN112686152B (en) * | 2020-12-30 | 2023-06-09 | 广西慧云信息技术有限公司 | Crop pest identification method with multi-size input and multi-size targets |
CN113033504A (en) * | 2021-05-19 | 2021-06-25 | 广东众聚人工智能科技有限公司 | Multi-scale video anomaly detection method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107818302A (en) | Non-rigid multiple dimensioned object detecting method based on convolutional neural networks | |
WO2021139069A1 (en) | General target detection method for adaptive attention guidance mechanism | |
CN106970615B (en) | A kind of real-time online paths planning method of deeply study | |
CN111368896B (en) | Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network | |
CN105512640B (en) | A kind of people flow rate statistical method based on video sequence | |
CN107451602A (en) | A kind of fruits and vegetables detection method based on deep learning | |
CN107092870B (en) | A kind of high resolution image Semantic features extraction method | |
CN104217214B (en) | RGB D personage's Activity recognition methods based on configurable convolutional neural networks | |
CN106845499A (en) | A kind of image object detection method semantic based on natural language | |
CN109584248A (en) | Infrared surface object instance dividing method based on Fusion Features and dense connection network | |
CN109919108A (en) | Remote sensing images fast target detection method based on depth Hash auxiliary network | |
CN109977918A (en) | A kind of target detection and localization optimization method adapted to based on unsupervised domain | |
CN108805070A (en) | A kind of deep learning pedestrian detection method based on built-in terminal | |
CN108021947B (en) | A kind of layering extreme learning machine target identification method of view-based access control model | |
CN110516539A (en) | Remote sensing image building extracting method, system, storage medium and equipment based on confrontation network | |
CN108830188A (en) | Vehicle checking method based on deep learning | |
CN107609525A (en) | Remote Sensing Target detection method based on Pruning strategy structure convolutional neural networks | |
CN108875624A (en) | Method for detecting human face based on the multiple dimensioned dense Connection Neural Network of cascade | |
CN107392901A (en) | A kind of method for transmission line part intelligence automatic identification | |
CN107316058A (en) | Improve the method for target detection performance by improving target classification and positional accuracy | |
CN109241982A (en) | Object detection method based on depth layer convolutional neural networks | |
CN106980858A (en) | The language text detection of a kind of language text detection with alignment system and the application system and localization method | |
CN109508360A (en) | A kind of polynary flow data space-time autocorrelation analysis method of geography based on cellular automata | |
CN107529650A (en) | The structure and closed loop detection method of network model, related device and computer equipment | |
CN108921879A (en) | The motion target tracking method and system of CNN and Kalman filter based on regional choice |
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 |