CN106157319B - The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks - Google Patents
The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks Download PDFInfo
- Publication number
- CN106157319B CN106157319B CN201610604732.0A CN201610604732A CN106157319B CN 106157319 B CN106157319 B CN 106157319B CN 201610604732 A CN201610604732 A CN 201610604732A CN 106157319 B CN106157319 B CN 106157319B
- Authority
- CN
- China
- Prior art keywords
- pixel
- conspicuousness
- level
- cnn
- region
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- 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]
-
- 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
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (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)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the conspicuousness detection methods that a kind of region based on convolutional neural networks and Pixel-level are merged, the object of this method research is still image, wherein the content of image can be arbitrary, the target of research is the target for attracting human eye attention to be found out from image, and different significance value is assigned for it.The present invention mainly proposes a kind of adaptive Area growth, and devises two CNN network structures, is respectively used to the prediction of Pixel-level conspicuousness and conspicuousness fusion.The two CNN network models are used for the training of network model, and final output and input picture Saliency maps of the same size using the legitimate reading of image as supervisory signals using image as input.The estimation of region class conspicuousness and the prediction of Pixel-level conspicuousness can be effectively performed in the present invention, obtain two Saliency maps, finally merged to obtain final Saliency maps by two Saliency maps and original image using the CNN for carrying out conspicuousness fusion.
Description
Technical field
The present invention relates to a kind of image processing methods based on deep learning, and in particular to one kind being based on convolutional neural networks
Region and Pixel-level fusion conspicuousness detection method.
Background technology
With the development and rise of deep learning, the conspicuousness detection technique based on deep learning also continues to develop.Significantly
Property detection can be divided into bottom-up data-driven model and top-down task-driven model two major classes.It is bottom-up aobvious
The detection of work property refers to finding out attractive target in figure, which can be any classification for given arbitrary piece image
Things.And top-down conspicuousness detection method usually finds out the target of given classification from given picture, and assign not
Same significance value.Currently, to the most study of bottom-up conspicuousness detection method.
Existing bottom-up conspicuousness detection method can be divided into two classes, the respectively method based on hand-designed feature
With the method based on convolutional neural networks.The surface information in image is usually utilized due to the method based on hand-designed feature
（Such as color, texture etc.）These manual features to carry out feature extraction, and extract cannot capture the deep layer of conspicuousness target
Characteristic and multi-scale information, to which good performance cannot be obtained.Recently with the rise of deep learning, there is part researcher to open
Begin to carry out conspicuousness target detection using convolutional neural networks.It is existing at present to be mostly based on convolutional neural networks（CNN）
Conspicuousness object detection method divide the image into multiple regions first, then to the trained CNN models in each region
To predict its significance value.But these methods cannot obtain the conspicuousness prediction result of accurate Pixel-level.
Invention content
In order to overcome the above problem, the present invention to propose a kind of new conspicuousness detection method based on CNN, that is, be based on volume
The conspicuousness detection method in the region and Pixel-level fusion of product neural network.This method includes three phases, respectively region class
Conspicuousness estimation, the prediction of Pixel-level conspicuousness and conspicuousness fusion, each stage are directed to a CNN model.This method can
The Saliency maps of accurate Pixel-level are obtained, so as to more effectively push the development of the related application based on conspicuousness.
The purpose of the present invention is what is be achieved through the following technical solutions：
The present invention provides the conspicuousness detection methods that a kind of region based on convolutional neural networks and Pixel-level are merged, should
The object of technique study is still image, and wherein the content of image can be arbitrary, and the target of research is found out from image
Attract the target of human eye attention, and assigns different significance value for it.The present invention mainly proposes a kind of adaptive area
Domain generation technique, and two CNN network structures are devised, it is respectively used to the prediction of Pixel-level conspicuousness and conspicuousness fusion.This two
A CNN network models are used for the training of network model using the legitimate reading of image as supervisory signals using image as input, and
Final output and input picture Saliency maps of the same size.The estimation of region class conspicuousness and pixel can be effectively performed in the present invention
Grade conspicuousness prediction, obtains two Saliency maps, finally using carrying out the CNN of conspicuousness fusion by two Saliency maps and original
Image is merged to obtain final Saliency maps, and whole system block diagram is as shown in Figure 1.
The specific implementation step of the present invention is as follows：
One, region class conspicuousness is estimated
The first step is split input picture I using adaptive Area growth
（1）Super-pixel segmentation is carried out to input picture I using SLIC algorithms, obtains n super-pixel；
（2）A simple feature vector is extracted from each super-pixel, for characterizing the characteristic of the super-pixel；
（3）Super-pixel is clustered to obtain different regions using an agglomerative clustering algorithm based on figure；
Second step carries out region significance estimation using Clarifai network models
（1）M super-pixel is randomly selected around the center line in each region；
（2）The m window of center construction of center as window of m super-pixel is set, and these windows include entire figure
Picture；
（3）M video in window of construction is passed through into CNN models, obtains m significance value；
（4）Calculate the mean value of m significance value and the significance value as the region；
Two, Pixel-level conspicuousness is predicted
（1）Using model VGGNet as pre-training model, by the last one module removal in VGGNet, to the 4th and
The output of five modules carries out operation of deconvoluting, and splices them in feature channel direction, learns for Analysis On Multi-scale Features；
Then it uses size to carry out convolution to spliced characteristic pattern for the convolution kernel of 1* 1 and obtains a probability graph；
（2）In Pixel-level CNN model training stages, calculated between probability graph and legitimate reading figure using fork entropy loss function
Error, and error is returned to update Pixel-level CNN model parameters；
（3）After Pixel-level CNN model trainings are complete, input picture I is directly inputted in Pixel-level CNN models and predicts it
Corresponding Pixel-level Saliency maps；
Three, conspicuousness merges
（1）Structure fusion CNN network structures：CNN network structures include a splicing layer, three convolutional layers and a loss
Layer；
（2）By input picture I and Step 1: two two Saliency maps are spliced into the image in 5 channels, then by it
It is sent into three convolutional layers；
（3）In the fusion CNN network training stages, the last one convolutional layer is calculated using the fork entropy loss function in loss layer
Output and true Saliency maps between error, and error is returned to update fusion CNN model parameters；
（4）When test, input picture I is directly inputted in trained fusion CNN models, the last one volume of the model
The Saliency maps that the output of lamination is as finally predicted.
The invention has the advantages that：
1, the present invention proposes a kind of new conspicuousness detection method based on CNN, and this method has given full play to region class
Conspicuousness estimates and the advantage of Pixel-level conspicuousness prediction, and achieves good conspicuousness detection performance.
2, the present invention proposes a kind of adaptive Area growth, which can be that different images generates different numbers
The region of amount, and the edge of object can be kept well.
3, the present invention devises a CNN network structure, which can effectively excavate the multiple dimensioned letter in image
Breath can also carry out and the relevant task of pixel classifications, such as image segmentation other than it can be used for the prediction of Pixel-level conspicuousness.
4, the present invention proposes a kind of new conspicuousness convergence strategy based on CNN, not only takes full advantage of Saliency maps
Between complementary information, also use information abundant in original image, to largely improve conspicuousness detection
Performance.
Description of the drawings
Fig. 1 is whole system block diagram of the present invention；
Fig. 2 is that adaptive region generates result example, and a- original images, b- legitimate readings, c- super-pixel segmentations are as a result, d-
Area generation result；
Fig. 3 is region class conspicuousness estimated result example, a- original images, b- legitimate readings, c- region class results；
Fig. 4 is Pixel-level CNN network structures；
Fig. 5 is Pixel-level conspicuousness prediction result example, a- original images, b- legitimate readings, c- Pixel-level results；
Fig. 6 is fusion CNN network structures；
Fig. 7 is present invention progress conspicuousness detection as a result, a- original images, b- legitimate readings, c- fusion results, d- pictures
Plain grade is as a result, e- region class results.
Specific implementation mode
Technical scheme of the present invention is further described below in conjunction with the accompanying drawings, however, it is not limited to this, every to this
Inventive technique scheme is modified or replaced equivalently, and without departing from the spirit of the technical scheme of the invention and range, should all be covered
In protection scope of the present invention.
The present invention provides the conspicuousness detection method that a kind of region based on convolutional neural networks and Pixel-level are merged, tools
Body implementation steps are as follows：
One, region class conspicuousness is estimated
In region class conspicuousness estimation procedure, wherein the first step is exactly that a large amount of region is generated from input picture.Most
Simple method is to carry out conspicuousness estimation using super-pixel as region, so that how to determine the super-pixel of segmentation
Number becomes highly difficult.If super-pixel number is very little, may be owed so that belonging to same conspicuousness mesh target area
Segmentation.If super-pixel number is too many, so that the region for belonging to conspicuousness target or background may be by over-segmentation.No matter
It is less divided or over-segmentation, all the significance value of conspicuousness target or background may be made inconsistent.Therefore, for different
Image, due to their different characteristics, it should be divided into the super-pixel of different number.To solve the above-mentioned problems, of the invention
A kind of adaptive Area growth is proposed to carry out image segmentation.An input picture I is given, which generates
The process of technology is as follows：
（1）Super-pixel segmentation is carried out to I using SLIC algorithms and obtains n super-pixel.Consider method effect and
Efficiency, in the present invention, n=300.
（2）A simple feature vector is extracted from each super-pixel（Contain the average color on Lab color spaces
With average spatial position coordinate）, for characterizing the characteristic of the super-pixel.
（3）Super-pixel is clustered to obtain different regions using an agglomerative clustering algorithm based on figure.
After the above process, the similar and adjacent super-pixel of color is usually clustered the same area in image I
In domain.For different images, the areal obtained after final cluster is also different, and will be far smaller than super-pixel
Number n.Fig. 2 provides the result example of adaptive region generation technique three obtained generation.
It is region significance estimation in next step after the region generated.The present invention uses Clarifai network models
（It is the CNN models that image classification task obtains top performance in ImageNet2013）To carry out region significance estimation.Tool
For body, m super-pixel is randomly selected around the center line in each region first, then the center of this m super-pixel is set
M window is built as the center of window, and these windows contain whole image.Super picture around selection region center line
Element be in order to（1）So that boundary of the center of constructed window as far away from region,（2）Window from different zones
In content it is as different as possible.In the present invention, when the number for the super-pixel for including in region is more than 5, m=5 are set, otherwise,
The value of m is set to the number of super-pixel.From the foregoing, it will be observed that for each region, m video in window will be constructed, by CNN models
Afterwards, m significance value will be obtained, their mean value and the significance value as the region are calculated, so that the region
Significance value it is more robust to noise.Fig. 3 gives the result example of three region class conspicuousnesses estimation.
Two, Pixel-level conspicuousness is predicted
Although the estimation of region class conspicuousness can obtain consistent and edge and keep good Saliency maps, picture cannot be obtained
The Saliency maps of plain class precision.In this regard, the present invention proposes a kind of CNN network structures（It is denoted as Pixel-level CNN）For carrying out pixel
Grade conspicuousness prediction.Pixel-level CNN is input with original image, to be output with the Saliency maps of the sizes such as original image.
Accurate conspicuousness prediction, the CNN structures should be deep layer and possess different strides in order to obtain（strides）It is multiple dimensioned
Stage, so that for image pixel study to the strong Analysis On Multi-scale Features of discriminating power.When training sample scale is smaller, accent
Start effectively to train such a network structure to be a very difficult task.Such issues that in order to overcome, one very well
Way be exactly use some trained models on large-scale dataset（For example those are extremely successful on ImageNet
Network model VGGNet and GoogleNet）As pre-training model, then to model in the small data set in the task of needs
It is finely adjusted, a strong model of learning ability is obtained so as to training.
The present invention modifies to it on the basis of VGGNet models, to build Pixel-level CNN models.VGGNet
By six modules（block）Composition, first five module is by convolutional layer（It is denoted as conv）With pond layer（It is denoted as pooling）Composition,
As shown in Figure 4.The last one module is made of a pond layer and two full articulamentums.The present invention is by last in VGGNet
A module removal.In order to which using the multi-scale information of image, the output that the present invention merges the 4th and the 5th module is more to realize
Scale feature learns.Since the output size of most latter two module is different and is much smaller than the size of original image, in order to make
Pixel-level CNN models can learn Analysis On Multi-scale Features and predict that the present invention first will be to last two for Pixel-level conspicuousness automatically
The output of a module carries out operation of deconvoluting（It is denoted as deconv）So that their size is consistent with original image, and
Feature channel direction splices them（It is denoted as concat）.Then it is the convolution kernel of 1* 1 to spliced spy to use size
Sign figure carries out convolution and obtains a probability graph, and in the probability graph, value means more greatly more notable.In test, the probability graph
The actually Saliency maps of input picture.In training, fork entropy loss function is used（It is denoted as loss）To calculate the probability
Error between figure and legitimate reading figure, and error is returned to update model parameter.This is arrived, entire Pixel-level CNN
Network structure all complete by structure, as shown in Figure 4.In model training stage, the stochastic gradient descent algorithm of standard by with
To minimize loss function.After model training is complete, image is directly inputted in model to predict that its corresponding Pixel-level is shown
Work property figure.Fig. 5 gives the example of three Pixel-level conspicuousness prediction results.
Three, conspicuousness merges
For a given image, two Saliency maps can effectively be obtained by the above process, respectively：Region class is aobvious
Work property figure and Pixel-level Saliency maps.Since the CNN models that they are the different information being utilized in image are calculated,
They have complementarity.If can effectively be merged to it, the performance of conspicuousness detection will be further increased.
The present invention designs a simple CNN network structure（It is denoted as fusion CNN）And learn a kind of nonlinear transformation to fill
Divide the complementary information between excavation regions grade Saliency maps and Pixel-level Saliency maps, high performance purpose is put forward to reach.It should
CNN network structures contain a splicing layer（concat）, three convolutional layers（conv）With a loss layer（loss）, such as Fig. 6
It is shown.Original image and its two Saliency maps are spliced into the image in 5 channels first, are then sent to subsequent
Three convolutional layers（Concrete configuration is shown in Fig. 6）.In test, conspicuousness that the output of the last one convolutional layer is as finally predicted
Figure.In training, using the fork entropy loss function in loss layer come the output for calculating the last one convolutional layer and true conspicuousness
Error between figure.From the foregoing, it will be observed that in conspicuousness fusion method proposed by the present invention, in addition to using two to obtain Saliency maps
Outside, original image is also used.This is because introduce the abundant information of original image can correct it is certain can not be by using only notable
The mistake that property figure is corrected when being merged.
Fusion CNN can be trained individually, also can carry out combination learning with the CNN networks of front to obtain lastness
The promotion of energy.In the estimation of region class conspicuousness, most start to need to generate multiple regions from input picture, then using area
Grade CNN carries out significance value estimation to each region.And Pixel-level CNN and fusion CNN then directly using image as input, and directly
It connects output and obtains Saliency maps, therefore be a process end to end.Therefore, it is difficult to which above three CNN network structures are incorporated
In the network unified to one, and carry out combination learning end to end.In order to simplify this process, finally, three CNN are advanced
Row individually training, then, Pixel-level CNN and fusion CNN further carry out combination learning on the basis of pre-training, such as Fig. 1 institutes
Show.When test, image is inputted into the frame in Fig. 1, the output for merging CNN is the Saliency maps finally predicted.Fig. 7 gives
The testing result example of four conspicuousness detection methods proposed by the present invention, as shown in Figure 7, result that the present invention detects with it is true
As a result very close to thus illustrating effectiveness of the invention.
Claims (2)
1. a kind of conspicuousness detection method of region and Pixel-level fusion based on convolutional neural networks, it is characterised in that the side
Steps are as follows for method：
Step 1: region class conspicuousness is estimated
The first step is split input picture I using adaptive Area growth
（1）Super-pixel segmentation is carried out to input picture I using SLIC algorithms, obtains n super-pixel；
（2）A simple feature vector is extracted from each super-pixel, for characterizing the characteristic of the super-pixel；
（3）Super-pixel is clustered to obtain different regions using an agglomerative clustering algorithm based on figure；
Second step carries out region significance estimation using Clarifai network models
（1）M super-pixel is randomly selected around the center line in each region；
（2）The m window of center construction of center as window of m super-pixel is set, and these windows include whole image；
（3）M video in window of construction is passed through into CNN models, obtains m significance value；
（4）Calculate the mean value of m significance value and the significance value as the region；
Step 2: Pixel-level conspicuousness is predicted
（1）Using model VGGNet as pre-training model, by the last one module removal in VGGNet, to the 4th and the 5th mould
The output of block carries out operation of deconvoluting, and splices them in feature channel direction, learns for Analysis On Multi-scale Features；Then
It uses size to carry out convolution to spliced characteristic pattern for the convolution kernel of 1* 1 and obtains a probability graph；
（2）In Pixel-level CNN model training stages, the mistake between probability graph and legitimate reading figure is calculated using fork entropy loss function
Difference, and error is returned to update Pixel-level CNN model parameters；
（3）After Pixel-level CNN model trainings are complete, input picture I is directly inputted in Pixel-level CNN models and predicts its correspondence
Pixel-level Saliency maps；
Step 3: conspicuousness merges
（1）Structure fusion CNN network structures：CNN network structures include a splicing layer, three convolutional layers and a loss layer；
（2）By input picture I and Step 1: two two Saliency maps are spliced into the image in 5 channels, then it is sent to
Three convolutional layers；
（3）In the fusion CNN network training stages, the defeated of the last one convolutional layer is calculated using the fork entropy loss function in loss layer
Go out the error between true Saliency maps, and error is returned to update fusion CNN model parameters；
（4）When test, input picture I is directly inputted in trained fusion CNN models, the last one convolutional layer of the model
The Saliency maps as finally predicted of output.
2. the conspicuousness detection method of the region and Pixel-level fusion according to claim 1 based on convolutional neural networks,
It is characterized in that n=300.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610604732.0A CN106157319B (en) | 2016-07-28 | 2016-07-28 | The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610604732.0A CN106157319B (en) | 2016-07-28 | 2016-07-28 | The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106157319A CN106157319A (en) | 2016-11-23 |
| CN106157319B true CN106157319B (en) | 2018-11-02 |
Family
ID=58060262
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610604732.0A Active CN106157319B (en) | 2016-07-28 | 2016-07-28 | The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106157319B (en) |
Families Citing this family (50)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106709532B (en) * | 2017-01-25 | 2020-03-10 | 京东方科技集团股份有限公司 | Image processing method and device |
| CN106911930A (en) * | 2017-03-03 | 2017-06-30 | 深圳市唯特视科技有限公司 | It is a kind of that the method for perceiving video reconstruction is compressed based on recursive convolution neutral net |
| CN106934397B (en) * | 2017-03-13 | 2020-09-01 | 北京市商汤科技开发有限公司 | Image processing method and device and electronic equipment |
| CN107016409A (en) * | 2017-03-20 | 2017-08-04 | 华中科技大学 | A kind of image classification method and system based on salient region of image |
| CN107169954B (en) * | 2017-04-18 | 2020-06-19 | 华南理工大学 | Image significance detection method based on parallel convolutional neural network |
| CN107194933A (en) * | 2017-04-24 | 2017-09-22 | 天津大学 | With reference to convolutional neural networks and the brain tumor dividing method and device of fuzzy reasoning |
| CN109118459B (en) * | 2017-06-23 | 2022-07-19 | 南开大学 | Image salient object detection method and device |
| CN107369160B (en) * | 2017-06-28 | 2020-04-03 | 苏州比格威医疗科技有限公司 | Choroid neogenesis blood vessel segmentation algorithm in OCT image |
| CN107437246B (en) * | 2017-07-05 | 2020-08-18 | 浙江大学 | Common significance detection method based on end-to-end full-convolution neural network |
| JP6907774B2 (en) * | 2017-07-14 | 2021-07-21 | オムロン株式会社 | Object detectors, object detection methods, and programs |
| CN107368831B (en) * | 2017-07-19 | 2019-08-02 | 中国人民解放军国防科学技术大学 | English words and digit recognition method in a kind of natural scene image |
| CN107506792B (en) * | 2017-08-16 | 2020-09-29 | 广西荷福智能科技有限公司 | Semi-supervised salient object detection method |
| US11222255B2 (en) * | 2017-08-17 | 2022-01-11 | Samsung Electronics Co., Ltd. | Neuromorphic processing apparatus |
| CN107730546B (en) * | 2017-08-25 | 2020-11-03 | 华北电力大学（保定） | Image depth feature determination method and system |
| CN107784308B (en) * | 2017-10-09 | 2020-04-03 | 哈尔滨工业大学 | Saliency target detection method based on chain type multi-scale full-convolution network |
| CN107766810B (en) * | 2017-10-10 | 2021-05-14 | 湖南省测绘科技研究所 | Cloud and shadow detection method |
| CN107886533B (en) * | 2017-10-26 | 2021-05-04 | 深圳大学 | Method, device and equipment for detecting visual saliency of three-dimensional image and storage medium |
| CN107945204B (en) * | 2017-10-27 | 2021-06-25 | 西安电子科技大学 | Pixel-level image matting method based on generation countermeasure network |
| CN107767383B (en) * | 2017-11-01 | 2021-05-11 | 太原理工大学 | Road image segmentation method based on superpixels |
| CN107945109B (en) * | 2017-11-06 | 2020-07-28 | 清华大学 | Image splicing method and device based on convolutional network |
| CN107967474A (en) * | 2017-11-24 | 2018-04-27 | 上海海事大学 | A kind of sea-surface target conspicuousness detection method based on convolutional neural networks |
| CN108345850B (en) * | 2018-01-23 | 2021-06-01 | 哈尔滨工业大学 | Scene text detection method based on region classification of stroke feature transformation and deep learning of superpixel |
| CN108389182B (en) * | 2018-01-24 | 2020-07-17 | 北京卓视智通科技有限责任公司 | Image quality detection method and device based on deep neural network |
| CN110111295B (en) * | 2018-02-01 | 2021-06-11 | 北京中科奥森数据科技有限公司 | Image collaborative saliency detection method and device |
| CN108961220B (en) * | 2018-06-14 | 2022-07-12 | 上海大学 | Image collaborative saliency detection method based on multilayer convolution feature fusion |
| CN109086777B (en) * | 2018-07-09 | 2021-09-28 | 南京师范大学 | Saliency map refining method based on global pixel characteristics |
| CN108960261B (en) * | 2018-07-25 | 2021-09-24 | 扬州万方电子技术有限责任公司 | Salient object detection method based on attention mechanism |
| CN109409222B (en) * | 2018-09-20 | 2020-10-30 | 中国地质大学（武汉） | Multi-view facial expression recognition method based on mobile terminal |
| CN109389056B (en) * | 2018-09-21 | 2020-05-26 | 北京航空航天大学 | Space-based multi-view-angle collaborative track surrounding environment detection method |
| RU2703270C1 (en) | 2018-10-31 | 2019-10-16 | Общество с ограниченной ответственностью "Аби Продакшн" | Optical character recognition using specialized confidence functions, implemented on the basis of neural networks |
| CN109409435B (en) * | 2018-11-01 | 2022-07-15 | 上海大学 | Depth perception significance detection method based on convolutional neural network |
| CN109934241B (en) * | 2019-03-28 | 2022-12-09 | 南开大学 | Image multi-scale information extraction method capable of being integrated into neural network architecture |
| CN110084221B (en) * | 2019-05-08 | 2023-02-03 | 南京云智控产业技术研究院有限公司 | Serialized human face key point detection method with relay supervision based on deep learning |
| CN110166850B (en) * | 2019-05-30 | 2020-11-06 | 上海交通大学 | Method and system for predicting panoramic video watching position by multiple CNN networks |
| CN110222704B (en) * | 2019-06-12 | 2022-04-01 | 北京邮电大学 | Weak supervision target detection method and device |
| CN110390327B (en) * | 2019-06-25 | 2022-06-28 | 北京百度网讯科技有限公司 | Foreground extraction method and device, computer equipment and storage medium |
| CN112149459B (en) * | 2019-06-27 | 2023-07-25 | 哈尔滨工业大学（深圳） | Video saliency object detection model and system based on cross attention mechanism |
| CN110472639B (en) * | 2019-08-05 | 2023-04-18 | 山东工商学院 | Target extraction method based on significance prior information |
| CN111583173B (en) * | 2020-03-20 | 2023-12-01 | 北京交通大学 | RGB-D image saliency target detection method |
| CN111311532B (en) * | 2020-03-26 | 2022-11-11 | 深圳市商汤科技有限公司 | Image processing method and device, electronic device and storage medium |
| CN111598841B (en) * | 2020-04-23 | 2022-04-15 | 南开大学 | Example significance detection method based on regularized dense connection feature pyramid |
| CN111260653B (en) * | 2020-04-27 | 2020-08-25 | 腾讯科技（深圳）有限公司 | Image segmentation method and device, storage medium and electronic equipment |
| CN111696021B (en) * | 2020-06-10 | 2023-03-28 | 中国人民武装警察部队工程大学 | Image self-adaptive steganalysis system and method based on significance detection |
| CN111915613B (en) * | 2020-08-11 | 2023-06-13 | 华侨大学 | Image instance segmentation method, device, equipment and storage medium |
| CN112465700B (en) * | 2020-11-26 | 2022-04-26 | 北京航空航天大学 | Image splicing positioning device and method based on depth clustering |
| CN112541912B (en) * | 2020-12-23 | 2024-03-12 | 中国矿业大学 | Rapid detection method and device for salient targets in mine sudden disaster scene |
| CN113034365B (en) * | 2021-03-19 | 2023-09-22 | 西安电子科技大学 | Multi-picture splicing method and system based on super pixels |
| CN114004775B (en) * | 2021-11-30 | 2023-07-04 | 四川大学 | Infrared and visible light image fusion method combining potential low-rank representation and convolutional neural network |
| CN114255351B (en) * | 2022-02-28 | 2022-05-27 | 魔门塔(苏州)科技有限公司 | Image processing method, device, medium, equipment and driving system |
| CN115965844B (en) * | 2023-01-04 | 2023-08-18 | 哈尔滨工业大学 | Multi-focus image fusion method based on visual saliency priori knowledge |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102222231A (en) * | 2011-05-26 | 2011-10-19 | 厦门大学 | Visual attention computational model based on guidance of dorsal pathway and processing method thereof |
| CN102567731A (en) * | 2011-12-06 | 2012-07-11 | 北京航空航天大学 | Extraction method for region of interest |
-
2016
- 2016-07-28 CN CN201610604732.0A patent/CN106157319B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102222231A (en) * | 2011-05-26 | 2011-10-19 | 厦门大学 | Visual attention computational model based on guidance of dorsal pathway and processing method thereof |
| CN102567731A (en) * | 2011-12-06 | 2012-07-11 | 北京航空航天大学 | Extraction method for region of interest |
Non-Patent Citations (4)
| Title |
|---|
| Offline Signature Verification based on ASIFT;Youbao Tang et.al;《International Confer-ence on Biometrics (ICB)》;20131231(第1期);第1-6页 * |
| Offline Text-Independent Writer Identification Based on Scale Invariant Feature Transform;Xiangqian Wu et.al;《IEEE Transactions on Information Forensics and Security》;20140331;第9卷(第3期);第526-536页 * |
| Saliency Detection based on Graph-Structural Agglomerative Clustering;Youbao Tang et.al;《ACM International Conference on Mul-timedia (ACMMM)》;20151231(第1期);第1083-1086页 * |
| 深度卷积神经网络的显著性检测;李岳云 等;《中国图像图形学报》;20160131;第21卷(第1期);第53-59页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106157319A (en) | 2016-11-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106157319B (en) | The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks | |
| CN110111335B (en) | Urban traffic scene semantic segmentation method and system for adaptive countermeasure learning | |
| CN107203781B (en) | End-to-end weak supervision target detection method based on significance guidance | |
| CN110852316B (en) | Image tampering detection and positioning method adopting convolution network with dense structure | |
| CN103984959B (en) | A kind of image classification method based on data and task-driven | |
| CN107480726A (en) | A kind of Scene Semantics dividing method based on full convolution and shot and long term mnemon | |
| CN109711413A (en) | Image, semantic dividing method based on deep learning | |
| CN105139395B (en) | SAR image segmentation method based on small echo pond convolutional neural networks | |
| CN109902806A (en) | Method is determined based on the noise image object boundary frame of convolutional neural networks | |
| CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
| CN107403430A (en) | A kind of RGBD image, semantics dividing method | |
| CN107680106A (en) | A kind of conspicuousness object detection method based on Faster R CNN | |
| CN109002755B (en) | Age estimation model construction method and estimation method based on face image | |
| CN109583340A (en) | A kind of video object detection method based on deep learning | |
| CN107066916B (en) | Scene semantic segmentation method based on deconvolution neural network | |
| CN106981080A (en) | Night unmanned vehicle scene depth method of estimation based on infrared image and radar data | |
| CN105912999A (en) | Human behavior identification method based on depth information | |
| CN104966286A (en) | 3D video saliency detection method | |
| CN107247952B (en) | Deep supervision-based visual saliency detection method for cyclic convolution neural network | |
| CN110334589A (en) | A kind of action identification method of the high timing 3D neural network based on empty convolution | |
| CN105184772A (en) | Adaptive color image segmentation method based on super pixels | |
| CN109858487A (en) | Weakly supervised semantic segmentation method based on watershed algorithm and image category label | |
| CN110472634A (en) | Change detecting method based on multiple dimensioned depth characteristic difference converged network | |
| CN112488229A (en) | Domain self-adaptive unsupervised target detection method based on feature separation and alignment | |
| CN111882620A (en) | Road drivable area segmentation method based on multi-scale information |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |