CN108921850A - A kind of extracting method of the image local feature based on image Segmentation Technology - Google Patents
A kind of extracting method of the image local feature based on image Segmentation Technology Download PDFInfo
- Publication number
- CN108921850A CN108921850A CN201810336591.8A CN201810336591A CN108921850A CN 108921850 A CN108921850 A CN 108921850A CN 201810336591 A CN201810336591 A CN 201810336591A CN 108921850 A CN108921850 A CN 108921850A
- Authority
- CN
- China
- Prior art keywords
- image
- network
- feature
- characteristic pattern
- training
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of extracting methods of image local feature based on image Segmentation Technology, include the following steps：1, Image Segmentation Model is constructed；2, image inputs CNN network, obtains the characteristic pattern of multilayer, after down-sampled by multiple convolutional layers and pondization, has obtained the information of script pixel position when characteristic pattern and down-sampled corresponding every characteristic pattern；3, pixel is assigned to the position of script by up-sampling module to up-sampling by characteristic pattern again；4, softmax loss is calculated to location of pixels each on newly generated characteristic pattern；5, the continuous iteration above process completes the process of the building and training of Image Segmentation Model until the penalty values of passback are small to receiving in range；6, the extraction that network completes image local feature is extracted by training characteristics.Different degrees of accurate retrieval, precise positioning target and target site may be implemented in the present invention, so that extracting key position feature carries out careful aspect ratio pair.
Description
Technical field
The present invention relates to technical field of image processing more particularly to a kind of image local features based on image Segmentation Technology
Extracting method.
Background technique
Image retrieval is the process using picture search similar image or same target image.Precisely retrieval requires to search for
To image be all same target image, rather than similar purpose.The process of image retrieval includes the extraction of characteristics of image, figure
As the storage of feature, the comparison of feature is ranked up output result by similarity.There are many method of image retrieval on the market,
It is wherein main to innovate the extracting mode for all focusing on characteristics of image, the row of the way of contrast of feature and last output result
Sequence method.Key of the invention is using image Segmentation Technology, and depth model goes to divide image or target progress pixel scale
It cuts, the position of precise positioning target position (pedestrian, vehicle is in the position of image) or target part (trunk of people, four limbs), needle
Corresponding extraction feature is come to these positions, to the comparison of Partial Feature can be careful whether tell be same target, phase
Than in global characteristics, the detailed information that such local feature includes is more, is more conducive to the differentiation of target.
The feature extraction scheme of most of prior art is carried out based on global characteristics, more careful local feature
Few feature extractions for pixel scale are substantially and are gone using the method for image gradient or the image block selected using frame
Carry out the extraction of local feature.In the image retrieval task of similar image, such feature can satisfy wanting for task substantially
Ask, but it is careful to the retrieval tasks for retrieving same target when, such feature can not meet mission requirements well.
In addition, it is the image for only including target image that retrieval image, which generally requires input picture, when carrying out retrieval tasks,
Or input full figure (figure comprising various scenes and target) after, one-time detection is carried out before retrieving, could to target into
Row image retrieval.
Summary of the invention
In view of the above drawbacks of the prior art, technical problem to be solved by the invention is to provide one kind based on image point
The extracting method of the image local feature of technology is cut, so as to solve the deficiencies in the prior art.
To achieve the above object, the present invention provides a kind of extraction sides of image local feature based on image Segmentation Technology
Method includes the following steps：
Step 1, building Image Segmentation Model, parted pattern is made of the network layer of multiple and different functions, including CNN model
Structure, batchnorm layer structure, deconv layers of structure；
Step 2, image input CNN network, obtain the characteristic pattern of multilayer, by multiple convolutional layers and pondization it is down-sampled it
Afterwards, the information of script pixel position when characteristic pattern and down-sampled corresponding every characteristic pattern has been obtained；
Characteristic pattern is passed through up-sampling module to up-sampling, according to the position of pixel script on every layer of characteristic pattern before by step 3
Pixel, is assigned to the position of script by confidence breath again, is consistent the feature of the characteristic pattern generated and original input picture；
Step 4 calculates location of pixels each on newly generated characteristic pattern softmax loss, to each pixel
Position is classified, by comparing image tag and the network output marked by hand as a result, calculating for each pixel position
The loss set is lost by passback, the parameter in training pattern；
Step 5, the continuous iteration above process complete image segmentation mould until the penalty values of passback are small to receiving in range
The process of the building and training of type；
Step 6, the building for completing Image Segmentation Model extract network by training characteristics with after the process of training, according to
The output image of upper segmentation network, the feature of different parts is extracted to original image, completes the extraction of image local feature.
A kind of extracting method of above-mentioned image local feature based on image Segmentation Technology, the feature of the step 6 mention
Taking network includes model construction part and feature training part.
A kind of extracting method of above-mentioned image local feature based on image Segmentation Technology, the feature extraction network are special
It levies the component extracted to design using residual error, by dividing the image into the output of network model, input feature vector extracts network
Residual error network, it is common to supervise finally using triple loss function and softmax loss function by a series of residual error layer
The training of this feature extraction network.
A kind of extracting method of above-mentioned image local feature based on image Segmentation Technology, the feature extraction network are defeated
The image entered is the output according to segmentation network as a result, extracting the image of corresponding position from original image, takes one piece of rectangular area, square
The size in shape region is set according to actual retrieval demand, on location of pixels belonging to the classification that in rectangular area, needs
Original image corresponding position pixel is filled, remaining position all sets 0.
A kind of extracting method of above-mentioned image local feature based on image Segmentation Technology, the feature extraction network exist
When training, the input picture of network is triple image, includes the image of two same targets, a different target in triple
Image；Training terminates, and when users use, image is opened in the input one of feature extraction network.
A kind of extracting method of above-mentioned image local feature based on image Segmentation Technology, the feature extraction network are defeated
Entering image can be the target image being partitioned into from entire scene, be also possible to the topography of target image.
The beneficial effects of the invention are as follows：
The present invention utilizes image Segmentation Technology, and characteristics of image part is melted using image segmentation network and feature extraction network
Network method is closed, network is divided the image into and introduces in feature extraction network, the difference of grade is marked according to target, may be implemented
Different degrees of accurate retrieval, precise positioning target and target site positioning, so that it is careful to extract the progress of key position feature
Aspect ratio pair, significantly more efficient same clarification of objective and other feature differentiations can be come.
It is described further below with reference to technical effect of the attached drawing to design of the invention, specific structure and generation, with
It is fully understood from the purpose of the present invention, feature and effect.
Detailed description of the invention
Fig. 1 is block flow diagram of the invention.
Specific embodiment
As shown in Figure 1, a kind of extracting method of the image local feature based on image Segmentation Technology, includes the following steps：
Step 1, building Image Segmentation Model, parted pattern is made of the network layer of multiple and different functions, including CNN model
Structure, batchnorm layer structure, deconv layers of structure；
Step 2, image input CNN network, obtain the characteristic pattern of multilayer, by multiple convolutional layers and pondization it is down-sampled it
Afterwards, the information of script pixel position when characteristic pattern and down-sampled corresponding every characteristic pattern has been obtained；
Characteristic pattern is passed through up-sampling module to up-sampling, according to the position of pixel script on every layer of characteristic pattern before by step 3
Pixel, is assigned to the position of script by confidence breath again, is consistent the feature of the characteristic pattern generated and original input picture；
Step 4 calculates location of pixels each on newly generated characteristic pattern softmax loss, to each pixel
Position is classified, by comparing image tag and the network output marked by hand as a result, calculating for each pixel position
The loss set is lost by passback, the parameter in training pattern；
Step 5, the continuous iteration above process complete image segmentation mould until the penalty values of passback are small to receiving in range
The process of the building and training of type；
Step 6, the building for completing Image Segmentation Model extract network by training characteristics with after the process of training, according to
The output image of upper segmentation network, the feature of different parts is extracted to original image, completes the extraction of image local feature.
In the present embodiment, the feature extraction network of the step 6 includes model construction part and feature training part.
In the present embodiment, component that the feature extraction network characterization extracts is using residual error design, by by image
Divide the output of network model, input feature vector extracts the residual error network of network, by a series of residual error layer, finally uses ternary
Group loss function and softmax loss function, supervise the training of this feature extraction network jointly.
In the present embodiment, the image of the feature extraction network inputs is the output according to segmentation network as a result, from original image
The upper image for extracting corresponding position, takes one piece of rectangular area, and the size of rectangular area is set according to actual retrieval demand, for
In rectangular area, original image corresponding position pixel is filled on location of pixels belonging to the classification that needs, remaining position all sets 0.
In the present embodiment, for the feature extraction network in training, the input picture of network is triple image, triple
In include two same targets image, the image of a different target；Training terminates, when users use, feature extraction net
Image is opened in the input one of network.
In the present embodiment, the feature extraction network inputs image can be the target figure being partitioned into from entire scene
Picture is also possible to the topography of target image.
General image retrieval process is divided into image characteristics extraction, image storage, retrieval three processes of image feature comparison,
This method is that innovation is made mainly for image characteristics extraction.Characteristics of image part uses image segmentation network and feature extraction net
The converged network method of network divides the image into network and introduces in feature extraction network, and the difference of grade is marked according to target, can
To realize different degrees of accurate retrieval.When mark is just for target and background, the characteristics of image of extraction will be limited to target
Region generates the target signature that an image local is used for characteristic key at this time.When mark has been carried out more for target itself
Fine-grained mark (for example, he has marked out the number of people, people's trunk, four limbs and belongings of people etc. for people), this
When obtain the retrieval character being composed of target local feature.Retrieval performance in the case of two kinds will be better than individually making
The effect retrieved with single image.
Image object refers to splitting target from image background from pixel scale in image, according to point
Fine-grained difference is cut, the part of target can also be separated to reach the local shape factor of target from target
Purpose out.Image partition method used in the present invention is a kind of image segmentation side using deep neural network CNN
Method, segmentation network is by original input picture according to training on training set as a result, generating output one and input image size
Consistent image, the pixel of this image correspond to the classification information of each pixel in original image.When in training data
When classification only includes target and two classifications of background, the pixel for exporting image will be only comprising two kinds of pixel values, to distinguish prospect
With background.The feature of target part is taken out according to the contexts distinguished, is put into feature extraction branch by the present invention, generates
Feature be exactly the feature used for image retrieval.
Next the substantially process of lower image segmentation is described.Firstly, it is necessary to construct Image Segmentation Model, parted pattern is by more
The network layer of a different function forms, and other than common CNN model structure, is also added into batchnorm layers, deconv layers etc.
Different structure.Overall model frame is as shown in Figure 1.
Image first passes through a CNN network, obtains the characteristic pattern of multilayer, is passing through multiple convolutional layers and pond
After (pooling, down-sampled), characteristic pattern has been obtained, and script pixel position when corresponding every characteristic pattern is down-sampled
Information.
After this, these characteristic patterns are by up-sampling, according to the location information of pixel script on every layer of characteristic pattern before,
Pixel is assigned to the position of script again, makes the characteristic pattern generated, is consistent as far as possible with the feature of original input picture.
Softmax loss is calculated to location of pixels each on newly generated characteristic pattern, to the position of each pixel into
Row classification, by comparing image tag and the network output marked by hand as a result, calculating the damage for each location of pixels
It loses (loss), is lost by passback, have the function that parameter in training pattern.Make in the output image of model, each pixel
A corresponding classification, according to the difference of classification, so that it may extract the image of corresponding position for the position of the classification of needs.
The continuous iteration above process, until the penalty values of passback are small to receiving in range, this completes image segmentations
The process of the building and training of model.
But since the present invention be directed to image retrieval tasks, the convolutional layer of segmentation network or the feature of pond layer are directly extracted
It is not enough to reach the demand of retrieval tasks, so introducing a feature extraction network here.This network is as segmentation network
One branch completes and then individually trains this component in segmentation network training.This feature extraction network is mainly
The output image for dividing network according to upper one extracts the feature of different parts to original image.This feature extraction network
It is divided into model construction part and feature training part.Drag building part is introduced first：
The component of feature extraction is designed using residual error, passes through the characteristic pattern and image segmentation for generating above-mentioned network
The output of network, the residual error network that input feature vector extracts, by a series of residual error layer, finally using triple loss function with
Softmax loss function supervises the training of this feature extraction network jointly.The image that input feature vector extracts network is basis point
The output of network is cut as a result, extracting the image of corresponding position from original image, takes one piece of rectangular area, the size of rectangular area according to
Actual retrieval demand setting, for filling original image corresponding positions on location of pixels belonging to the classification that in rectangular area, needs
Pixel is set, remaining position all sets 0.The classification of triple loss and softmax feature one supervision feature, a supervision are different
The dispersion degree of target signature, it is therefore an objective to keep the feature resolution ability generated stronger, it is more sensitive to same clarification of objective.
Training part, due to needing to use triple loss, so the tissue in input picture needs to carry out some places
Reason, the image of 2 different images and another different target of same target is combined, and is inputted in network together.At this
In invention, input picture can be the target image being partitioned into from entire scene, be also possible to the topography of target image.
Remaining training process is similar with the segmentation training process of network.
In the output par, c of feature extraction network, in conjunction with image segmentation network before output as a result, can correspond to
Segmentation result extracts the feature of image section and target part, carries out the comparison of target local detail.
After characteristic pattern and segmentation result input feature vector are extracted network, it can set and mention according to actual items demand
The form of feature is taken, different local features can be obtained to target different parts according to segmentation result, can also integrally extracted
Target part feature.When extracting entire target signature, there are two or more targets in permission input picture.When selection mentions
When taking target part provincial characteristics, input picture only allows that there are a targets.The difference with comparison method is selected according to user,
The selection mode of this local feature is also more flexible.
The preferred embodiment of the present invention has been described in detail above.It should be appreciated that those skilled in the art without
It needs creative work according to the present invention can conceive and makes many modifications and variations.Therefore, all technologies in the art
Personnel are available by logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea
Technical solution, all should be within the scope of protection determined by the claims.
Claims (6)
1. a kind of extracting method of the image local feature based on image Segmentation Technology, which is characterized in that include the following steps：
Step 1, building Image Segmentation Model, parted pattern is made of the network layer of multiple and different functions, including CNN model knot
Structure, batchnorm layer structure, deconv layers of structure；
Step 2, image input CNN network, obtain the characteristic pattern of multilayer, after down-sampled by multiple convolutional layers and pondization,
The information of script pixel position when characteristic pattern and down-sampled corresponding every characteristic pattern is obtained；
Characteristic pattern is passed through up-sampling module to up-sampling by step 3, is believed according to the position of pixel script on every layer of characteristic pattern before
Pixel, is assigned to the position of script by breath again, is consistent the feature of the characteristic pattern generated and original input picture；
Step 4 calculates location of pixels each on newly generated characteristic pattern softmax loss, to the position of each pixel
Classify, by comparing image tag and the network output marked by hand as a result, calculating for each location of pixels
Loss is lost by passback, the parameter in training pattern；
Step 5, the continuous iteration above process complete Image Segmentation Model until the penalty values of passback are small to receiving in range
The process of building and training；
Step 6, the building for completing Image Segmentation Model extract network by training characteristics, according to upper one with after the process of training
The output image of a segmentation network, the feature of different parts is extracted to original image, completes the extraction of image local feature.
2. a kind of extracting method of the image local feature based on image Segmentation Technology as described in claim 1, feature exist
In the feature extraction network of the step 6 includes model construction part and feature training part.
3. a kind of extracting method of the image local feature based on image Segmentation Technology as claimed in claim 2, feature exist
In：The component that the feature extraction network characterization extracts is designed using residual error, by dividing the image into the defeated of network model
Out, input feature vector extracts the residual error network of network, by a series of residual error layer, finally using triple loss function with
Softmax loss function supervises the training of this feature extraction network jointly.
4. a kind of extracting method of the image local feature based on image Segmentation Technology as claimed in claim 2, feature exist
In：The image of the feature extraction network inputs is the output according to segmentation network as a result, extracting corresponding position from original image
Image takes one piece of rectangular area, and the size of rectangular area is set according to actual retrieval demand, for needing in rectangular area
Classification belonging to original image corresponding position pixel is filled on location of pixels, remaining position all sets 0.
5. a kind of extracting method of the image local feature based on image Segmentation Technology as claimed in claim 4, feature exist
In：For the feature extraction network in training, the input picture of network is triple image, includes two same mesh in triple
Target image, the image of a different target；Training terminates, and when users use, image is opened in the input one of feature extraction network.
6. a kind of extracting method of the image local feature based on image Segmentation Technology as claimed in claim 4, feature exist
In：The feature extraction network inputs image can be the target image being partitioned into from entire scene, be also possible to target figure
The topography of picture.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810336591.8A CN108921850B (en) | 2018-04-16 | 2018-04-16 | Image local feature extraction method based on image segmentation technology |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810336591.8A CN108921850B (en) | 2018-04-16 | 2018-04-16 | Image local feature extraction method based on image segmentation technology |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108921850A true CN108921850A (en) | 2018-11-30 |
| CN108921850B CN108921850B (en) | 2022-05-17 |
Family
ID=64402935
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810336591.8A Active CN108921850B (en) | 2018-04-16 | 2018-04-16 | Image local feature extraction method based on image segmentation technology |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921850B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109960742A (en) * | 2019-02-18 | 2019-07-02 | 苏州科达科技股份有限公司 | The searching method and device of local message |
| CN110458849A (en) * | 2019-07-26 | 2019-11-15 | 山东大学 | A kind of image partition method based on characteristic modification |
| WO2020118826A1 (en) * | 2018-12-14 | 2020-06-18 | 深圳先进技术研究院 | Left ventricle image segmenting method and apparatus, and device and storage medium |
| CN113255760A (en) * | 2021-05-20 | 2021-08-13 | 推想医疗科技股份有限公司 | Method for training image processing model, method and device for image processing |
| CN115661449A (en) * | 2022-09-22 | 2023-01-31 | 北京百度网讯科技有限公司 | Image segmentation and training method and device of image segmentation model |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN106384100A (en) * | 2016-09-28 | 2017-02-08 | 武汉大学 | Component-based fine vehicle model recognition method |
| CN106649487A (en) * | 2016-10-09 | 2017-05-10 | 苏州大学 | Image retrieval method based on interest target |
| CN106980641A (en) * | 2017-02-09 | 2017-07-25 | 上海交通大学 | The quick picture retrieval system of unsupervised Hash and method based on convolutional neural networks |
| CN107016681A (en) * | 2017-03-29 | 2017-08-04 | 浙江师范大学 | Brain MRI lesion segmentation approach based on full convolutional network |
| CN107203999A (en) * | 2017-04-28 | 2017-09-26 | 北京航空航天大学 | A kind of skin lens image automatic division method based on full convolutional neural networks |
| CN107729818A (en) * | 2017-09-21 | 2018-02-23 | 北京航空航天大学 | A kind of multiple features fusion vehicle recognition methods again based on deep learning |
| CN107784282A (en) * | 2017-10-24 | 2018-03-09 | 北京旷视科技有限公司 | The recognition methods of object properties, apparatus and system |
-
2018
- 2018-04-16 CN CN201810336591.8A patent/CN108921850B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN106384100A (en) * | 2016-09-28 | 2017-02-08 | 武汉大学 | Component-based fine vehicle model recognition method |
| CN106649487A (en) * | 2016-10-09 | 2017-05-10 | 苏州大学 | Image retrieval method based on interest target |
| CN106980641A (en) * | 2017-02-09 | 2017-07-25 | 上海交通大学 | The quick picture retrieval system of unsupervised Hash and method based on convolutional neural networks |
| CN107016681A (en) * | 2017-03-29 | 2017-08-04 | 浙江师范大学 | Brain MRI lesion segmentation approach based on full convolutional network |
| CN107203999A (en) * | 2017-04-28 | 2017-09-26 | 北京航空航天大学 | A kind of skin lens image automatic division method based on full convolutional neural networks |
| CN107729818A (en) * | 2017-09-21 | 2018-02-23 | 北京航空航天大学 | A kind of multiple features fusion vehicle recognition methods again based on deep learning |
| CN107784282A (en) * | 2017-10-24 | 2018-03-09 | 北京旷视科技有限公司 | The recognition methods of object properties, apparatus and system |
Non-Patent Citations (3)
| Title |
|---|
| YUHENG SONG ET AL.: "Image Segmentation Techniques Overview", 《2017 ASIA MODELLING SYMPOSIUM (AMS)》 * |
| 张永良: "基于深度学习的小面积指纹匹配方法", 《计算机应用》 * |
| 陈鸿翔: "基于卷积神经网络的图像语义分割", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020118826A1 (en) * | 2018-12-14 | 2020-06-18 | 深圳先进技术研究院 | Left ventricle image segmenting method and apparatus, and device and storage medium |
| CN109960742A (en) * | 2019-02-18 | 2019-07-02 | 苏州科达科技股份有限公司 | The searching method and device of local message |
| CN109960742B (en) * | 2019-02-18 | 2021-11-05 | 苏州科达科技股份有限公司 | Local information searching method and device |
| CN110458849A (en) * | 2019-07-26 | 2019-11-15 | 山东大学 | A kind of image partition method based on characteristic modification |
| CN110458849B (en) * | 2019-07-26 | 2023-04-25 | 山东大学 | Image segmentation method based on feature correction |
| CN113255760A (en) * | 2021-05-20 | 2021-08-13 | 推想医疗科技股份有限公司 | Method for training image processing model, method and device for image processing |
| CN115661449A (en) * | 2022-09-22 | 2023-01-31 | 北京百度网讯科技有限公司 | Image segmentation and training method and device of image segmentation model |
| CN115661449B (en) * | 2022-09-22 | 2023-11-21 | 北京百度网讯科技有限公司 | Image segmentation and training method and device for image segmentation model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108921850B (en) | 2022-05-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921850A (en) | A kind of extracting method of the image local feature based on image Segmentation Technology | |
| Tao et al. | An object detection system based on YOLO in traffic scene | |
| Liang et al. | Proposal-free network for instance-level object segmentation | |
| Zhou et al. | Learning deep features for scene recognition using places database | |
| CN110335277A (en) | Image processing method, device, computer readable storage medium and computer equipment | |
| Wang et al. | Training deeper convolutional networks with deep supervision | |
| CN110188635A (en) | A kind of plant pest recognition methods based on attention mechanism and multi-level convolution feature | |
| CN104899590B (en) | A kind of unmanned plane sensation target follower method and system | |
| CN109886998A (en) | Multi-object tracking method, device, computer installation and computer storage medium | |
| CN108304873A (en) | Object detection method based on high-resolution optical satellite remote-sensing image and its system | |
| CN110853026B (en) | Remote sensing image change detection method integrating deep learning and region segmentation | |
| CN109145939A (en) | A kind of binary channels convolutional neural networks semantic segmentation method of Small object sensitivity | |
| CN109919981A (en) | A kind of multi-object tracking method of the multiple features fusion based on Kalman filtering auxiliary | |
| CN109376603A (en) | A kind of video frequency identifying method, device, computer equipment and storage medium | |
| CN107871124A (en) | A kind of Remote Sensing Target detection method based on deep neural network | |
| CN105608692B (en) | Polarization SAR image segmentation method based on deconvolution network and sparse classification | |
| CN109299688A (en) | Ship Detection based on deformable fast convolution neural network | |
| CN110084318A (en) | A kind of image-recognizing method of combination convolutional neural networks and gradient boosted tree | |
| CN106250874A (en) | A kind of dress ornament and the recognition methods of carry-on articles and device | |
| CN109801292A (en) | A kind of bituminous highway crack image partition method based on generation confrontation network | |
| CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
| CN110110755A (en) | Based on the pedestrian of PTGAN Regional disparity and multiple branches weight recognition detection algorithm and device | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| CN105303163B (en) | A kind of method and detection device of target detection | |
| CN109766873A (en) | pedestrian re-identification method based on hybrid deformable convolution |
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 |