CN107527031A - A kind of indoor objects detection method based on SSD - Google Patents
A kind of indoor objects detection method based on SSD Download PDFInfo
- Publication number
- CN107527031A CN107527031A CN201710724937.7A CN201710724937A CN107527031A CN 107527031 A CN107527031 A CN 107527031A CN 201710724937 A CN201710724937 A CN 201710724937A CN 107527031 A CN107527031 A CN 107527031A
- Authority
- CN
- China
- Prior art keywords
- module
- convolution
- dimension
- input
- branch road
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/35—Categorising the entire scene, e.g. birthday party or wedding scene
- G06V20/36—Indoor scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
-
- 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
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Image Analysis (AREA)
Abstract
A kind of indoor objects detection method based on SSD of the disclosure of the invention, belongs to field of target recognition, is a kind of new application method of the convolutional neural networks for indoor objects detection.What the present invention solved is indoor objects detection, but currently without the database for meeting this requirement of experiment marked, a database related to indoor objects is constructed for this, all images of the database are all with the openr angular samples in the visual field, meet the normal viewing angle of intelligent robot, there is otherness again in background, illumination and picture size etc..Artificial mark has been carried out to indoor frequent goal refrigerator, TV, bed, dining table, chair, sofa, tea table, closestool, Wash Units, bathing pool, cup etc.；Network and detector are extracted using the image training characteristics of acquisition, finally target to be identified is detected using the feature extraction network and detector that train.
Description
Technical field
Field of target recognition of the present invention, it is a kind of new application side of the convolutional neural networks for indoor objects detection
Method.
Background technology
Indoor objects identification is the process for a kind of target being separated and being identified from other targets under environment indoors,
The application field such as design, the retrieval of interior design plan of research and development, household monitoring system for household service robot has
Important meaning.But traditional technique in measuring accuracy and speed can not all reach requirement, more to increase so there is an urgent need for one kind
Imitate accurate detection method.
In computer vision field, there are all multi-methods to can be used for indoor objects detection, traditional object detection method base
In features such as SIFT, HOG, using sorting techniques such as SVM, Adaboot.But the feature of these hand-designeds is often to image
Sign it is limited in one's ability, cause target identification precision and positioning accuracy often relatively low, it is difficult to meet application request.And
, there is good improvement in a Ge Xin branch of the deep learning as machine learning in real-time and accuracy.In recent years, depth
Study is developed rapidly, and good performance is all presented in fields such as automatic speech recognition, natural language processing and image procossings.
In image processing field, deep learning is widely used to image classification, target detection and semantic segmentation.Convolutional neural networks are
The core of deep learning, the superperformance that convolutional neural networks obtain for deep learning are performed meritorious deeds never to be obliterated.In object detection field,
The models such as Faster R-CNN, YOLO and SSD have good Detection results, and convolutional neural networks play not wherein
Alternative effect.At the same time, Recognition with Recurrent Neural Network also has good development in recent years.Cyclic convolution network is for the time
The processing of sequence has its unique advantage compared to convolutional neural networks.And the model such as LSTM has greatly improved cyclic convolution
The gradient disappearance problem of network.
For the detection of indoor objects, there are some relatively good schemes, such as：Target inspection based on Faster R-CNN
Survey, RPN (Region Proposal Network) is formed using shared convolution net, Suggestion box is directly predicted with RPN, and
RPN overwhelming majority prediction is completed in GPU, convolution net and Fast R-CNN partial sharings, therefore greatly improves target inspection
The speed of survey, but can not still meet to require in real time., can be with based on YOLO (You Only Look Once) target detection
The position of target is directly returned out, uses preset frame mechanism due to no, YOLO accuracy of detection is not very high.In addition, also
There is the detection method based on SSD (Single Shot Multibox Detector), this method is also direct regressive object
Position and classification, accuracy of detection are improved with respect to YOLO, but still the mission requirements not reached.
Neutral net has millions of parameters, it is easy to the problem of over-fitting occurs.In order to solve asking for over-fitting
Topic, it is general using first pre-training goes out a model on large-scale dataset, then again with specific small-scale data set to this
Model is finely adjusted.Simultaneously the node of some hidden layers can be allowed not worked in training by dropout method, to reach
Prevent the purpose of over-fitting.
The content of the invention
The present invention proposes a kind of improved indoor objects detection based on SSD, for family's monitoring, service robot
Application scenarios.On the database of foundation, the experiment enriched, indoor common furniture is detected.
What the present invention solved is indoor objects detection, but currently without the data for meeting this requirement of experiment marked
Storehouse, a database related to indoor objects is constructed for this, all images of the database are all openr with the visual field
Angular samples, meet the normal viewing angle of intelligent robot, there is difference again in background, illumination and picture size etc.
Property.Indoor frequent goal refrigerator, TV, bed, dining table, chair, sofa, tea table, closestool, Wash Units, bathing pool, cup etc. are carried out
Artificial mark；Network and detector are extracted using the image training characteristics of acquisition, finally using the feature extraction net trained
Network and detector detect to target to be identified.Thus technical solution of the present invention is a kind of indoor objects detection based on SSD
Method, this method include:
Step 1:Obtain indoor target image to be detected；
Step 2：Feature extraction network is established, the global characteristics of network extraction target image are extracted using this feature；
Step 3：The global characteristics that step 2 is obtained input SSD detectors, testing result corresponding to acquisition；
It is characterized in that the feature extraction network of the step 2 includes：Three input modules, the first to the tenth a roll of product module
Block, the first to the 5th pond module, two contextual information extraction modules, a normalization module；Three described input moulds
Block is respectively an image to be detected and first and second zone bit information input module, and described image to be detected is as the first convolution
The input of module；First convolution module, the first pond module, the second convolution module, the second pond module, the 3rd convolution module,
3rd pond module, Volume Four volume module, the 4th pond module, the 5th convolution module, the 5th pond module, the 6th convolution mould
Block, the 7th convolution module, the 8th convolution module, the 9th convolution module, the tenth convolution module, the 11st convolution module level successively
Connection；Extra, the output of Volume Four volume module will also be input to normalizing together with the output of the first zone bit information input module
Change module, the output for then normalizing module is input to the first contextual information extraction module；Extra, the 7th convolution module
Output will also be input to the second contextual information extraction module together with the output of the second zone bit information input module；Finally will
The output of first and second contextual information extraction module, the 8th to the 11st convolution module is as the global characteristics extracted.
Further, described contextual information extraction module includes two convolution modules, a cascade module, a horizontal stroke
To feature extraction branch road and a longitudinal feature extraction branch road；The output of first convolution module is separately input to transverse features and carried
Branch road and longitudinal feature extraction branch road are taken, the output of transverse features the extraction branch road and longitudinal feature extraction branch road inputs jointly
To cascade module, the then output of cascade module is input to second convolution module；Wherein cascade module is a cascading layers.Should
Scheme extracts contextual information from horizontal and vertical two directions, is effectively extracted the Global Information of image.
Further, two convolution modules in the contextual information extraction module are all that convolution kernel size is 1*1, step
A length of 1, it is filled with 0 convolutional layer.The program can effectively extract characteristics of image, and help to prevent network from dissipating.
Further, described transverse features branch road and longitudinal feature branch road be all by a dimension conversion module, one
Dimension merging module, a cyclic convolution module, a dimension recovery module cascade composition successively；The dimension transformation module
Act as each dimensional information of input picture is exchanged into position, the number of dimensions for acting as reducing image of dimension merging module
The image dimension number of degrees after reduction are reverted to its original number of dimensions and order by amount, the acting as of the dimension recovery module；Follow
Ring convolution module is LSTM layers (shot and long term memory network), and its direction handled is controlled by the dimension conversion module of corresponding branch road.
The program takes full advantage of disposal ability of the cyclic convolution network to time series, and flexible Application to the feature of single image with carrying
Take, and LSTM has good inhibition to gradient disperse.
Further, the dimension of the input picture is n*c*w*h, and wherein n represents image number, and c represents port number, w
Represent that picture is wide, h represents that image is high；Dimension order is changed into h*n*w*c by the dimension conversion module in the transverse features branch road,
Dimension order is changed into w*n*h*c by the dimension conversion module in longitudinal feature branch road；The dimension of two branch roads merges
Module is all that second and third dimension are merged into a dimension.Program main function is in control cyclic convolution module
LSTM extracts the direction of characteristics of image, and transverse features extraction branch road is to follow image as a time series input per a line
Ring convolution module, longitudinal feature extraction branch road are that image is inputted into cyclic convolution module as a time series per a line.
Further, it by two convolution kernel sizes is 3*3 that the first and second convolution modules, which are all, step-length 1, is filled with 1
Convolutional layer composition；It by three convolution kernel sizes is 3*3 that 3rd to the 5th convolution module, which is, step-length 1, is filled with 1 convolution
Layer composition；It by a convolution kernel size is 3*3 that 6th convolution module, which is, step-length 1, is filled with 6 convolutional layer composition；Volume seven
It by a convolution kernel size is 1*1 that volume module, which is, step-length 1, is filled with 1 convolutional layer composition；8th and the 9th convolution module
It by a convolution kernel size is 1*1 to be, step-length 1, is filled with 0 convolutional layer and a convolution kernel size is 3*3, step-length 2,
It is filled with 1 convolutional layer composition；It by a convolution kernel size is 1*1 that 8th and the 9th convolution module, which is, step-length 1, filling
Convolutional layer and a convolution kernel size for 0 are 3*3, step-length 1, are filled with 0 convolutional layer composition.The program is fully extracted
The feature of image, is easy to the detector below effectively to detect target.
The present invention is improved on the basis of the method for existing detection to extraction characteristic, flexible Application circulation volume
Contextual information is added in characteristic spectrum by product network.In existing method, accuracy and speed is considered, select SSD as inspection
Survey scheme.SSD considers multiple features spectrum, selects the module of suitable characteristic spectrum addition, in the case where ensureing speed, will examine
Survey precision to improve, the actual demand preferably met.
Brief description of the drawings
Fig. 1 is the overall network structure of the present invention；
Fig. 2 is the contextual information extraction module of the present invention；
Fig. 3 is partial test result figure.
Embodiment
Groundwork of the present invention is divided into two parts of training and test, and all working is divided into six steps：
Step 1, structure database：For it is to be studied the problem of, build house data storehouse.Image selects from indoor design website
Take, visual angle is wider, and indoor frequent goal relativeness more can significantly embody.After some inappropriate images are excluded,
6000 multiple images are collected altogether, wherein 2/3rds are used as training sample, 1/3rd are used as test sample.Because sample is less,
To avoid over-fitting, having carried out random stripping and slicing to image in training, mirror image etc. operates, to increase sample size.
Step 2, the detection to sample object：Image all in database is manually marked, by the mesh in image
Its ground truth are marked out, that is, mark out position and the classification of target, classification is demarcated as 0,1,2....9, totally ten
Classification.
Step 3, improve SSD models：For room objects, have between thing and thing, thing and environment and closely contact, they
Be according to the mankind share custom place, so for room objects one of feature, improve SSD models, by global information
It is added among the detection to specific objective.Using feature extraction network extraction characteristics of image, using LSTM to time series
Disposal ability, model is built, the characteristic spectrum of image information is split in rows and columns respectively, be then input to cyclic convolution network and work as
In handled, then carry out that processing is in parallel and convolution operation to information, the characteristic spectrum of row processing and column processing be processed into one
Individual characteristic spectrum, it is then input in SSD detection network and is detected.Because high-level characteristic global information is lost seriously, and too
The characteristic amount of low layer is excessive, and lengthy and jumbled information is a lot, so the present invention adds contextual information extraction module the 4th
In the characteristic spectrum of the 7th two convolution module output.
Step 4, pre-training model：Because network model is larger, parameter is more, and sample is less, to prevent over-fitting, first will
In ImageNet, this larger database is trained model, obtains pre-training model.
Step 5, the improved model of training：On the basis of step 4, continue to carry out model using the database of oneself
Training.Data are pre-processed before training, expanding data sample.Training obtains final network model.
Step 6, test model：To training pattern carried out respectively on PASCAL VOC207 and the database of oneself
Test, obtains position and the classification of detection image.The precision that the test result present invention is detected compared to SSD for indoor objects has
Significantly improve, and remain to complete to detect with fast speed, the use demand that meets well.
6 steps more than, are improved the indoor objects detection method based on SSD, more fully make use of
The contextual information of image, target detection capabilities of the networking for indoor scene are effectively increased, and ensure that the detection of network
Speed.Partial test result is as shown in Figure 3.
Claims (6)
1. a kind of indoor objects detection method based on SSD, this method include:
Step 1:Obtain indoor target image to be detected；
Step 2：Feature extraction network is established, the global characteristics of network extraction target image are extracted using this feature；
Step 3：The global characteristics that step 2 is obtained input SSD detectors, testing result corresponding to acquisition；
It is characterized in that the feature extraction network of the step 2 includes：Three input modules, the first to the 11st convolution module,
First to the 5th pond module, two contextual information extraction modules, a normalization module；Three described input modules point
Not Wei an image to be detected and first and second zone bit information input module, described image to be detected is as the first convolution module
Input；First convolution module, the first pond module, the second convolution module, the second pond module, the 3rd convolution module, the 3rd
Pond module, Volume Four volume module, the 4th pond module, the 5th convolution module, the 5th pond module, the 6th convolution module,
Seven convolution modules, the 8th convolution module, the 9th convolution module, the tenth convolution module, the 11st convolution module cascade successively；Additionally
, the output of Volume Four volume module will also be input to normalization module together with the output of the first zone bit information input module,
Then the output for normalizing module is input to the first contextual information extraction module；Extra, the output of the 7th convolution module is also
The second contextual information extraction module is input to together with the output of the second zone bit information input module；Finally by first,
The output of two contextual information extraction modules, the 8th to the 11st convolution module is as the global characteristics extracted.
A kind of 2. indoor objects detection method based on SSD as claimed in claim 1, it is characterised in that the contextual information
Extraction module includes two convolution modules, a cascade module, a transverse features extraction branch road and a longitudinal feature extraction
Branch road；The output of first convolution module is separately input to transverse features extraction branch road and longitudinal feature extraction branch road, the horizontal stroke
Output to feature extraction branch road and longitudinal feature extraction branch road is input to cascade module jointly, and then the output of cascade module is defeated
Enter to second convolution module；Wherein cascade module is a cascading layers.
A kind of 3. indoor objects detection method based on SSD as claimed in claim 1 or 2, it is characterised in that first He
It by two convolution kernel sizes is 3*3 that second convolution module, which is all, step-length 1, is filled with 1 convolutional layer composition；3rd to the 5th
It by three convolution kernel sizes is 3*3 that convolution module, which is, step-length 1, is filled with 1 convolutional layer composition；6th convolution module be by
One convolution kernel size is 3*3, step-length 1, is filled with 6 convolutional layer composition；7th convolution module is big by a convolution kernel
Small is 1*1, step-length 1, is filled with 1 convolutional layer composition；It by a convolution kernel size is 1* that 8th and the 9th convolution module, which is,
1, step-length 1, is filled with 0 convolutional layer and a convolution kernel size is 3*3, step-length 2, is filled with 1 convolutional layer composition；
It by a convolution kernel size is 1*1 that 8th and the 9th convolution module, which is, step-length 1, is filled with 0 convolutional layer and a convolution kernel
Size is 3*3, step-length 1, is filled with 0 convolutional layer composition.
A kind of 4. indoor objects detection method based on SSD as claimed in claim 2, it is characterised in that the contextual information
Two convolution modules in extraction module are all that convolution kernel size is 1*1, step-length 1, are filled with 0 convolutional layer.
A kind of 5. indoor objects detection method based on SSD as claimed in claim 2, it is characterised in that described transverse features
Branch road and longitudinal feature branch road are all by a dimension conversion module, a dimension merging module, cyclic convolution module, one
Individual dimension recovery module cascades composition successively；The dimension transformation module is act as each dimensional information friendship of input picture
Change place, the number of dimensions for acting as reducing image of dimension merging module, the dimension recovery module act as reducing
The image dimension number of degrees afterwards revert to its original number of dimensions and order；Cyclic convolution module is LSTM layers, its handle direction by
The dimension conversion module control of corresponding branch road.
6. a kind of indoor objects detection method based on SSD as claimed in claim 5, it is characterised in that the input picture
Dimension is n*c*w*h, and wherein n represents image number, and c represents port number, and w represents that picture is wide, and h represents that image is high；The transverse direction
Dimension order is changed into h*n*w*c by dimension conversion module in feature branch road, the dimension conversion mould in longitudinal feature branch road
Dimension order is changed into w*n*h*c by block；The dimension merging module of two branch roads is all to be merged into second and third dimension
In one dimension.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710724937.7A CN107527031B (en) | 2017-08-22 | 2017-08-22 | SSD-based indoor target detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710724937.7A CN107527031B (en) | 2017-08-22 | 2017-08-22 | SSD-based indoor target detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107527031A true CN107527031A (en) | 2017-12-29 |
| CN107527031B CN107527031B (en) | 2021-02-05 |
Family
ID=60681878
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710724937.7A Active CN107527031B (en) | 2017-08-22 | 2017-08-22 | SSD-based indoor target detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107527031B (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108256567A (en) * | 2018-01-12 | 2018-07-06 | 环球大数据科技有限公司 | A kind of target identification method and system based on deep learning |
| CN108537117A (en) * | 2018-03-06 | 2018-09-14 | 哈尔滨思派科技有限公司 | A kind of occupant detection method and system based on deep learning |
| CN108710897A (en) * | 2018-04-24 | 2018-10-26 | 江苏科海智能系统有限公司 | A kind of online general target detecting system in distal end based on SSD-T |
| CN109241872A (en) * | 2018-08-20 | 2019-01-18 | 电子科技大学 | Image, semantic fast partition method based on multistage network |
| CN109284760A (en) * | 2018-08-10 | 2019-01-29 | 杭州群核信息技术有限公司 | A kind of furniture detection method and device based on depth convolutional neural networks |
| CN109308458A (en) * | 2018-08-31 | 2019-02-05 | 电子科技大学 | A method of small target deteection precision is promoted based on characteristic spectrum change of scale |
| CN109472315A (en) * | 2018-11-15 | 2019-03-15 | 江苏木盟智能科技有限公司 | A kind of object detection method and system separating convolution based on depth |
| CN109741318A (en) * | 2018-12-30 | 2019-05-10 | 北京工业大学 | The real-time detection method of single phase multiple dimensioned specific objective based on effective receptive field |
| CN109840502A (en) * | 2019-01-31 | 2019-06-04 | 深兰科技（上海）有限公司 | A kind of method and apparatus carrying out target detection based on SSD model |
| CN110222593A (en) * | 2019-05-18 | 2019-09-10 | 四川弘和通讯有限公司 | A kind of vehicle real-time detection method based on small-scale neural network |
| CN111460919A (en) * | 2020-03-13 | 2020-07-28 | 华南理工大学 | Monocular vision road target detection and distance estimation method based on improved YO L Ov3 |
| CN112446372A (en) * | 2020-12-08 | 2021-03-05 | 电子科技大学 | Text detection method based on channel grouping attention mechanism |
| TWI738009B (en) * | 2019-06-20 | 2021-09-01 | 和碩聯合科技股份有限公司 | Object detection system and object detection method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103198489A (en) * | 2013-04-27 | 2013-07-10 | 哈尔滨工业大学 | Automatic detection method of salient object based on salience density and edge response |
| CN106845549A (en) * | 2017-01-22 | 2017-06-13 | 珠海习悦信息技术有限公司 | A kind of method and device of the scene based on multi-task learning and target identification |
| US10216949B1 (en) * | 2013-09-20 | 2019-02-26 | Amazon Technologies, Inc. | Dynamic quorum membership changes |
-
2017
- 2017-08-22 CN CN201710724937.7A patent/CN107527031B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103198489A (en) * | 2013-04-27 | 2013-07-10 | 哈尔滨工业大学 | Automatic detection method of salient object based on salience density and edge response |
| US10216949B1 (en) * | 2013-09-20 | 2019-02-26 | Amazon Technologies, Inc. | Dynamic quorum membership changes |
| CN106845549A (en) * | 2017-01-22 | 2017-06-13 | 珠海习悦信息技术有限公司 | A kind of method and device of the scene based on multi-task learning and target identification |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108256567B (en) * | 2018-01-12 | 2020-08-25 | 环球大数据科技有限公司 | Target identification method and system based on deep learning |
| CN108256567A (en) * | 2018-01-12 | 2018-07-06 | 环球大数据科技有限公司 | A kind of target identification method and system based on deep learning |
| CN108537117A (en) * | 2018-03-06 | 2018-09-14 | 哈尔滨思派科技有限公司 | A kind of occupant detection method and system based on deep learning |
| CN108710897A (en) * | 2018-04-24 | 2018-10-26 | 江苏科海智能系统有限公司 | A kind of online general target detecting system in distal end based on SSD-T |
| CN109284760A (en) * | 2018-08-10 | 2019-01-29 | 杭州群核信息技术有限公司 | A kind of furniture detection method and device based on depth convolutional neural networks |
| CN109241872A (en) * | 2018-08-20 | 2019-01-18 | 电子科技大学 | Image, semantic fast partition method based on multistage network |
| CN109241872B (en) * | 2018-08-20 | 2022-03-18 | 电子科技大学 | Image semantic fast segmentation method based on multistage network |
| CN109308458A (en) * | 2018-08-31 | 2019-02-05 | 电子科技大学 | A method of small target deteection precision is promoted based on characteristic spectrum change of scale |
| CN109308458B (en) * | 2018-08-31 | 2022-03-15 | 电子科技大学 | Method for improving small target detection precision based on characteristic spectrum scale transformation |
| CN109472315B (en) * | 2018-11-15 | 2021-09-24 | 江苏木盟智能科技有限公司 | Target detection method and system based on depth separable convolution |
| CN109472315A (en) * | 2018-11-15 | 2019-03-15 | 江苏木盟智能科技有限公司 | A kind of object detection method and system separating convolution based on depth |
| CN109741318A (en) * | 2018-12-30 | 2019-05-10 | 北京工业大学 | The real-time detection method of single phase multiple dimensioned specific objective based on effective receptive field |
| CN109741318B (en) * | 2018-12-30 | 2022-03-29 | 北京工业大学 | Real-time detection method of single-stage multi-scale specific target based on effective receptive field |
| CN109840502B (en) * | 2019-01-31 | 2021-06-15 | 深兰科技（上海）有限公司 | Method and device for target detection based on SSD model |
| CN109840502A (en) * | 2019-01-31 | 2019-06-04 | 深兰科技（上海）有限公司 | A kind of method and apparatus carrying out target detection based on SSD model |
| CN110222593A (en) * | 2019-05-18 | 2019-09-10 | 四川弘和通讯有限公司 | A kind of vehicle real-time detection method based on small-scale neural network |
| TWI738009B (en) * | 2019-06-20 | 2021-09-01 | 和碩聯合科技股份有限公司 | Object detection system and object detection method |
| CN111460919A (en) * | 2020-03-13 | 2020-07-28 | 华南理工大学 | Monocular vision road target detection and distance estimation method based on improved YO L Ov3 |
| CN111460919B (en) * | 2020-03-13 | 2023-05-26 | 华南理工大学 | Monocular vision road target detection and distance estimation method based on improved YOLOv3 |
| CN112446372A (en) * | 2020-12-08 | 2021-03-05 | 电子科技大学 | Text detection method based on channel grouping attention mechanism |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107527031B (en) | 2021-02-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107527031A (en) | A kind of indoor objects detection method based on SSD | |
| CN111126202B (en) | Optical remote sensing image target detection method based on void feature pyramid network | |
| Li et al. | Building-a-nets: Robust building extraction from high-resolution remote sensing images with adversarial networks | |
| CN107316058A (en) | Improve the method for target detection performance by improving target classification and positional accuracy | |
| CN107358257B (en) | Under a kind of big data scene can incremental learning image classification training method | |
| CN109410242A (en) | Method for tracking target, system, equipment and medium based on double-current convolutional neural networks | |
| CN103871100B (en) | Tree modelling method for reconstructing based on a cloud Yu data-driven | |
| CN109446970A (en) | A kind of Intelligent Mobile Robot road scene recognition methods based on deep learning | |
| CN105631413A (en) | Cross-scene pedestrian searching method based on depth learning | |
| CN107154043A (en) | A kind of Lung neoplasm false positive sample suppressing method based on 3DCNN | |
| Han et al. | A new method in wheel hub surface defect detection: Object detection algorithm based on deep learning | |
| CN105740773A (en) | Deep learning and multi-scale information based behavior identification method | |
| CN109614488A (en) | Distribution network live line work condition distinguishing method based on text classification and image recognition | |
| CN109410190B (en) | Tower pole reverse-breaking detection model training method based on high-resolution remote sensing satellite image | |
| CN105868706A (en) | Method for identifying 3D model based on sparse coding | |
| CN107958067A (en) | It is a kind of based on without mark Automatic Feature Extraction extensive electric business picture retrieval system | |
| CN112818777B (en) | Remote sensing image target detection method based on dense connection and feature enhancement | |
| CN107527065A (en) | A kind of flower variety identification model method for building up based on convolutional neural networks | |
| Ning et al. | Point-voxel and bird-eye-view representation aggregation network for single stage 3D object detection | |
| Ouyang et al. | Aerial target detection based on the improved YOLOv3 algorithm | |
| Li et al. | A hierarchical deformable deep neural network and an aerial image benchmark dataset for surface multiview stereo reconstruction | |
| Zhang et al. | Recognition and detection of wolfberry in the natural background based on improved YOLOv5 network | |
| CN107578445A (en) | Image discriminant region extracting method based on convolution characteristic spectrum | |
| Li et al. | Detecting Plant Leaves Based on Vision Transformer Enhanced YOLOv5 | |
| Li et al. | Ctmu-net: an improved u-net for semantic segmentation of remote-sensing images based on the combined attention mechanism |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |