CN109740686A - A kind of deep learning image multiple labeling classification method based on pool area and Fusion Features - Google Patents
A kind of deep learning image multiple labeling classification method based on pool area and Fusion Features Download PDFInfo
- Publication number
- CN109740686A CN109740686A CN201910019115.8A CN201910019115A CN109740686A CN 109740686 A CN109740686 A CN 109740686A CN 201910019115 A CN201910019115 A CN 201910019115A CN 109740686 A CN109740686 A CN 109740686A
- Authority
- CN
- China
- Prior art keywords
- feature
- interest
- fusion features
- pool area
- candidate 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.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The deep learning image multiple labeling classification method based on pool area and Fusion Features that the present invention provides a kind of the steps include: 1) to carry out candidate region extraction to image to be processed, generate the candidate region of different scale；2) feature of interest figure is extracted by the conventional part of the convolutional neural networks of pre-training in candidate region, generate multichannel feature of interest Fig. 3 of different scale) the feature of interest figure of different scale is subjected to pool area, so that various sizes of feature of interest figure becomes identical size；4) the feature of interest figure of multiple identical sizes is merged, generates last feature vector；5) feature vector after Fusion Features is entered into classifier, predicts the probability in image containing object；The present invention is based on convolutional neural networks, and pool area and Fusion Features is added, provide not only the method that Fusion Features are carried out in convolutional neural networks, and the method for a high precision image multiple labeling is provided, facilitate researcher and engineering application personnel to improve image classification effect in practical applications.
Description
Technical field
The invention belongs to computer image processing technology fields, are related to a kind of depth based on pool area and Fusion Features
Degree study image multiple labeling classification method.
Background technique
The classification of image multiple labeling is the basic task of image procossing, it is therefore an objective to be identified in image comprising which class object
Body.Image classification and the classification of image multiple labeling are two tasks, and image classification is a kind of special circumstances of the more classification tasks of image,
A kind of object is contained only in image.In this year, due to the fast development of deep learning, image classification task also achieves considerable
Progress, the image classification based on convolutional neural networks are even more to have reached and the comparable level of the mankind.Then the multiple labeling of image point
For class due to the multiple objects in image to be identified, task is more complicated, and effect need to be improved.However in actual application
In, multiple labeling image classification has many application scenarios.There are below in research and application for image multiple labeling task at present
Problem:
On the one hand, image multiple labeling is classified the object due to identify multiple types, and scene is often relative complex,
It sometimes will overlap between object and object, increase the difficulty of identification.
On the other hand, multiple labeling classification requires have certain contiguity between label and label to a certain extent, uses
Deep learning model can capture being associated between label and label, and model can tend to convergence however actual scene in training
In, there may be lesser relevance between label and label.
There are also on the one hand, image classification task needs flag data to complete the training of model.In practical applications, relatively
For image classification, the every picture of multiple labeling task will mark multiple labels, increase label difficulty.
These problems will lead to user and be difficult with image multiple labeling, often convert multiple images classification for problem and appoint
Business weakens ease for use of the multiple labeling in practical field application to a certain extent.
Summary of the invention
Technical problem solved by the invention is in view of the deficiencies of the prior art, to provide a kind of based on pool area and spy
Levy fusion deep learning image multiple labeling classification method, provided for user it is clear in structure understandable, while still have high-precision figure
The method classified as multiple labeling.
Technical solution provided by the present invention are as follows:
A kind of deep learning image multiple labeling classification method based on pool area and Fusion Features comprising the steps of:
Step 1): candidate region extraction is carried out to image to be processed；It is extracted from picture using candidate region extraction algorithm
Several various sizes of candidate region P ∈ N × W outl×Hl(l=1,2....N), for subsequent step processing；Wherein N is candidate regions
The quantity in domain, WlAnd HlIt is the length and width of candidate region respectively, the scale of each candidate region is different；
Step 2): the candidate region P ∈ N × W obtained according to step 1)l×Hl(l=1,2....N), it is good into pre-training
Convolutional neural networks conventional part, generate multichannel feature of interest figure F ∈ N × C × w of different scalel×hl(l=1,
2....N)；Wherein N is feature of interest figure quantity, and C is feature of interest figure port number, wlAnd hlIt is single feature channel respectively
Length and width；Pre-training convolutional neural networks can be used, feature extraction is carried out to candidate region, generating has high abstraction
The feature of interest figure of feature；
Step 3): the feature of interest figure F ∈ N × C × w obtained according to step 2)l×hl(l=1,2....N) area is carried out
Domain pond generates the identical feature of interest figure of scale；wl×hlThe length and width for indicating each feature of interest figure, by it
The size of all feature of interest figures is normalized to w ' × h ', the feature of interest figure after normalization are as follows: and F ∈ N × C × w ' ×
h′；
Step 4): several normalized feature of interest figure F ∈ N × C × w ' obtained according to step 3) × h ' progress is special
Sign fusion generates the feature vector being abstracted with altitude feature, V ∈ T；Wherein T indicates the species number of object；
Step 5): the feature vector obtained according to step 4) classifies into classifier, predicts label vector p ∈ T；In advance
Each position p of direction finding amount piIndicate the probability containing object i；Given threshold θ, piIf being more than that threshold θ contains object i；
The candidate region extraction algorithm mentioned in the step 1), including but not limited to SelectiveSearch,
The candidate regions such as EdgeBoxes extraction algorithm.
The method for extracting feature of interest figure by convolutional neural networks mentioned in the step 2), including but not limited to
Other convolutional neural networks such as Alexnet, VGG, Resnet.
The step 3) the following steps are included:
Step 3.1): by each wl×hlInterested characteristic pattern be divided into a sub- characteristic pattern of w ' × h ', that is, use transverse and longitudinal
The method of cutting, the size of each subcharacter figure are as follows:
Step 3.2): each subcharacter figure is subjected to maximum pond, extracts maximum value characteristic point；Each subcharacter figure is most
Big characteristic point reassembles into new characteristic pattern, and the size of new characteristic pattern is w ' × h '；It is special that each feature of interest figure is subjected to son
Sign figure divides, maximum pond, and feature recombination, all feature of interest figures are normalized to w ' × h '；
The step 4) the following steps are included:
Step 4.1): several layers are entered according to normalization feature of interest figure F ∈ N × C × w ' × h ' that step 4) obtains
Full Connection Neural Network carries out feature extraction, generates feature of interest vector；First normalization feature of interest figure is carried out once
Convolution, convolution kernel size are w ' × h ', export as Fc ∈ N × C ', enter back into the full Connection Neural Network of several layers, export as Fs ∈
N × T, wherein N indicates feature of interest vector number, and T is the dimension of each vector；
Step 4.2): Fusion Features are carried out according to multiple feature vector Fs ∈ N × T that step 4.1) obtains, generate feature
Vector V ∈ T；Set Fs=(t1,t2,.....tN)t1∈ N × 1, V=(v1,v2,.....vN), then vi=max (t1)；
The prediction result vector mentioned in the step 5) is activated by feature vector and is obtained, p=δ (V), wherein pi=δ
(Vi), Activiation method includes but is not limited to sigmoid, the activation primitives such as tanh, softmax.
The utility model has the advantages that
The deep learning image multiple labeling classification method based on pool area and Fusion Features that the present invention provides a kind of, knot
Structure is clear, but maintains higher multiple labeling nicety of grading.This method has focused on a kind of framework, and model has certain
Washability, user can select appropriately sized model for oneself real data size.User only needs to be grasped basic
Deep learning and the knowledge of convolutional neural networks it will be appreciated that this method structure and self-actualization.The present invention has stronger
Expansion, user can formulate data label according to self-demand, complete training and be applied in actual production.
Detailed description of the invention
Fig. 1 is the method for the invention flow chart；
Fig. 2 is integrated stand composition of the present invention；
Fig. 3 is that schematic diagram is extracted in candidate region；
Fig. 4 is Fusion Features schematic diagram；
Fig. 5 is multiple labeling classification results schematic diagram；
Specific embodiment
To keep the purpose of the present invention, mentality of designing and advantage clearer, below in conjunction with specific example, and referring to attached drawing,
Invention is further described in detail.
The present invention provides a kind of deep learning image multiple labeling classification side based on pool area and Fusion Features
Method (title), as shown in Figure 1, including five key steps: picture candidate region is extracted, and different scale candidate region is generated；It waits
Favored area enters pre-training convolutional neural networks, generates feature of interest figure；Feature of interest graph region pond, generates identical ruler
Spend characteristic pattern；Multiple feature of interest figure Fusion Features generate feature vector；Feature vector is circulated into classifier, predicts figure
The kind of object contained in piece；
The committed step that method of the invention is related to is described in detail one by one below, it is shown that specific step is as follows:
Step 1 carries out candidate region extraction to image to be processed；As shown in Fig. 2, picture generally uses within 1000px
Resolution ratio, several various sizes of candidate region P ∈ N × W are extracted from picture using EdgeBoxes algorithml×Hl(l=
1,2....N) after, being extracted using EdgeBoxes algorithm, N is generally 200 or so；WlAnd HlAccording in image object it is big
Depending on small and shape.
Step 2, according to step 1) obtained candidate region P ∈ N × Wl×Hl(l=1,2....N), it is good into pre-training
VGG in, generate multichannel feature of interest figure F ∈ N × C × w of different scalel×hl(l=1,2....N)；Wherein N is sense
Interest characteristics figure quantity, generally 200 or so, C are that feature of interest figure port number is 256, wlAnd hlIt is that single feature is logical respectively
The length and width in road, depending on the size of candidate region；Convolution process can reduce the size of image while extracting feature；
Step 3, according to step 2) obtained feature of interest figure F ∈ N × C × wl×hl(l=1,2....N) area is carried out
Domain pond generates the identical feature of interest figure of scale；As shown in figure 4, wl×hlIndicate each feature of interest figure length and
The size of its all feature of interest figure is normalized to w ' × h ', value 40*40 by width.By each wl×hlSense it is emerging
The characteristic pattern of interest is divided into 40 × 40 sub- characteristic patterns, that is, the method for using transverse and longitudinal cutting, the size of each subcharacter figure are as follows:Each subcharacter figure is subjected to maximum pond, extracts maximum value characteristic point；The maximum feature of each subcharacter figure
Point reassembles into new characteristic pattern, and the size of new characteristic pattern is 40 × 40；Each feature of interest figure is carried out subcharacter figure to draw
Point, maximum pond, feature recombination, the feature of interest figure after normalization are as follows: F ∈ 200 × 256 × 40 × 40；
Step 4 enters several according to the normalization feature of interest figure F ∈ 200 × 256 × 40 × 40 that step 3 obtains
The full Connection Neural Network of layer carries out feature extraction, generates feature of interest vector；Normalization feature of interest figure is first carried out one
Secondary convolution, convolution kernel size are 40 × 40, export as Fc ∈ 200 × 256, enter back into the full Connection Neural Network of several layers, export
For Fs ∈ 256 × 20, object category is 20 classes, and the output dimension connected entirely is 20；The multiple feature vector Fs ∈ 256 that will be obtained
× 20 carry out Fusion Features, and maximum pond is carried out in 256 dimensions, generates feature vector V ∈ 20；
Step 5 is activated on each position, is predicted last according to the feature vector V ∈ 20 that step 4 generates
As a result p ∈ 20.Activiation method are as follows: p=sigmoid (V), given threshold 0.5, if predicted value is greater than 0.5 on each position,
Then contain the category；As a result as shown in Figure 5.
Claims (7)
1. a kind of deep learning image multiple labeling classification method based on pool area and Fusion Features, which is characterized in that include
Following steps:
Step 1): candidate region extraction is carried out to image to be processed；Several differences are extracted from picture using candidate frame algorithm
Candidate region P ∈ N × W of sizel×Hl(l=1,2....N), for subsequent step processing；Wherein N is the quantity of candidate region,
WlAnd HlIt is the length and width of candidate region respectively, the scale of each candidate region is different；
Step 2): the candidate region P ∈ N × W obtained according to step 1)l×Hl(l=1,2....N), the volume good into pre-training
The conventional part of product neural network, generates multichannel feature of interest figure F ∈ N × C × w of different scalel×hl(l=1,
2....N)；Wherein N is feature of interest figure quantity, and C is feature of interest figure port number, wlAnd hlIt is single feature channel respectively
Length and width；Pre-training convolutional neural networks can be used, feature extraction is carried out to candidate region, generating has high abstraction
The feature of interest figure of feature；
Step 3): the feature of interest figure F ∈ N × C × w obtained according to step 2)l×hl(l=1,2....N) pool area is carried out
Change, generates the identical feature of interest figure of scale；wl×hlThe length and width for indicating each feature of interest figure, is owned
The size of feature of interest figure be normalized to w ' × h ', the feature of interest figure after normalization are as follows: F ∈ N × C × w ' × h '；
Step 4): melted according to several normalized feature of interest figure F ∈ N × C × w ' × h ' carry out features that step 3) obtains
It closes, generates the feature vector being abstracted with altitude feature, V ∈ T；Wherein T indicates the species number of object；
Step 5): the feature vector obtained according to step 4) classifies into classifier, predicts label vector p ∈ T；Pre- direction finding
Measure each position p of piIndicate the probability containing object i；Given threshold θ, piIf being more than that threshold θ contains object i.
2. a kind of deep learning image multiple labeling classification side based on pool area and Fusion Features according to claim 1
Method, which is characterized in that the candidate region extraction algorithm that the step 1) is mentioned, including but not limited to SelectiveSearch,
The candidate regions such as EdgeBoxes extraction algorithm.
3. a kind of deep learning image multiple labeling classification side based on pool area and Fusion Features according to claim 1
Method, which is characterized in that the method for extracting feature of interest figure by convolutional neural networks that the step 2) is mentioned, including but not
It is limited to Alexnet, other convolutional neural networks such as VGG, Resnet.
4. a kind of deep learning image multiple labeling classification side based on pool area and Fusion Features according to claim 1
Method, which is characterized in that the step 3) the following steps are included:
Step 3.1): by each wl×hlInterested characteristic pattern be divided into a sub- characteristic pattern of w ' × h ', that is, use transverse and longitudinal cutting
Method, the size of each subcharacter figure are as follows:
Step 3.2): each subcharacter figure is subjected to maximum pond, extracts maximum value characteristic point；The maximum of each subcharacter figure is special
Sign point reassembles into new characteristic pattern, and the size of new characteristic pattern is w ' × h '；Each feature of interest figure is subjected to subcharacter figure
It divides, maximum pond, feature recombination, all feature of interest figures are normalized to w ' × h '.
5. a kind of deep learning image multiple labeling classification side based on pool area and Fusion Features according to claim 1
Method, which is characterized in that the step 4) the following steps are included:
Step 4.1): the normalization feature of interest figure F ∈ N × C × w ' × h ' obtained according to step 4) enters several layers and connects entirely
It connects neural network and carries out feature extraction, generate feature of interest vector；Normalization feature of interest figure is first subjected to a convolution,
Convolution kernel size is w ' × h ', exports as Fc ∈ N × C ', enters back into the full Connection Neural Network of several layers, is exported as Fs ∈ N × T,
Wherein N indicates feature of interest vector number, and T is the dimension of each vector；
Step 4.2): Fusion Features are carried out according to multiple feature vector Fs ∈ N × T that step 4.1) obtains, generate feature vector V
∈T；Set Fs=(t1,t2,.....tN)t1∈ N × 1, V=(v1,v2,.....vN), then vi=max (t1)。
6. a kind of deep learning image multiple labeling classification side based on pool area and Fusion Features according to claim 1
Method, which is characterized in that in the step 5), prediction result vector is activated by feature vector and obtained, p=δ (V), wherein pi=δ
(Vi), Activiation method includes but is not limited to sigmoid, the activation primitives such as tanh, softmax.
7. a kind of deep learning image multiple labeling classification side based on pool area and Fusion Features according to claim 1
Method, which is characterized in that the process comprising training pattern；Deep learning is divided into training process and prediction process；Trained error side
Journey are as follows:
Wherein l and p respectively indicate label and predicted vector.Error includes but is not limited to Euclidean distance.It is mentioned due to step 2)
Convolutional Neural neural network extracts the network for the pre-training that feature uses, therefore only training step 4) the full coupling part mentioned.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910019115.8A CN109740686A (en) | 2019-01-09 | 2019-01-09 | A kind of deep learning image multiple labeling classification method based on pool area and Fusion Features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910019115.8A CN109740686A (en) | 2019-01-09 | 2019-01-09 | A kind of deep learning image multiple labeling classification method based on pool area and Fusion Features |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109740686A true CN109740686A (en) | 2019-05-10 |
Family
ID=66364065
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910019115.8A Pending CN109740686A (en) | 2019-01-09 | 2019-01-09 | A kind of deep learning image multiple labeling classification method based on pool area and Fusion Features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740686A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110348537A (en) * | 2019-07-18 | 2019-10-18 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN110378297A (en) * | 2019-07-23 | 2019-10-25 | 河北师范大学 | A kind of Remote Sensing Target detection method based on deep learning |
| CN110427970A (en) * | 2019-07-05 | 2019-11-08 | 平安科技（深圳）有限公司 | Image classification method, device, computer equipment and storage medium |
| CN110825904A (en) * | 2019-10-24 | 2020-02-21 | 腾讯科技（深圳）有限公司 | Image matching method and device, electronic equipment and storage medium |
| CN111383429A (en) * | 2020-03-04 | 2020-07-07 | 西安咏圣达电子科技有限公司 | Method, system, device and storage medium for detecting dress of workers in construction site |
| CN112071421A (en) * | 2020-09-01 | 2020-12-11 | 深圳高性能医疗器械国家研究院有限公司 | Deep learning estimation method and application thereof |
| CN112101373A (en) * | 2019-06-18 | 2020-12-18 | 富士通株式会社 | Object detection method and device based on deep learning network and electronic equipment |
| CN112308115A (en) * | 2020-09-25 | 2021-02-02 | 安徽工业大学 | Multi-label image deep learning classification method and equipment |
| CN112418261A (en) * | 2020-09-17 | 2021-02-26 | 电子科技大学 | Human body image multi-attribute classification method based on prior prototype attention mechanism |
| CN112419292A (en) * | 2020-11-30 | 2021-02-26 | 深圳云天励飞技术股份有限公司 | Pathological image processing method and device, electronic equipment and storage medium |
| CN113609323A (en) * | 2021-07-20 | 2021-11-05 | 上海德衡数据科技有限公司 | Image dimension reduction method and system based on neural network |
| CN113705361A (en) * | 2021-08-03 | 2021-11-26 | 北京百度网讯科技有限公司 | Method and device for detecting model in living body and electronic equipment |
| CN114202805A (en) * | 2021-11-24 | 2022-03-18 | 北京百度网讯科技有限公司 | Living body detection method, living body detection device, electronic apparatus, and storage medium |
-
2019
- 2019-01-09 CN CN201910019115.8A patent/CN109740686A/en active Pending
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112101373A (en) * | 2019-06-18 | 2020-12-18 | 富士通株式会社 | Object detection method and device based on deep learning network and electronic equipment |
| JP7501125B2 (en) | 2019-06-18 | 2024-06-18 | 富士通株式会社 | Method, device and electronic device for object detection based on deep learning network |
| CN110427970A (en) * | 2019-07-05 | 2019-11-08 | 平安科技（深圳）有限公司 | Image classification method, device, computer equipment and storage medium |
| CN110427970B (en) * | 2019-07-05 | 2023-08-01 | 平安科技（深圳）有限公司 | Image classification method, apparatus, computer device and storage medium |
| CN110348537A (en) * | 2019-07-18 | 2019-10-18 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| US11481574B2 (en) | 2019-07-18 | 2022-10-25 | Beijing Sensetime Technology Development Co., Ltd. | Image processing method and device, and storage medium |
| CN110378297B (en) * | 2019-07-23 | 2022-02-11 | 河北师范大学 | Remote sensing image target detection method and device based on deep learning and storage medium |
| CN110378297A (en) * | 2019-07-23 | 2019-10-25 | 河北师范大学 | A kind of Remote Sensing Target detection method based on deep learning |
| CN110825904A (en) * | 2019-10-24 | 2020-02-21 | 腾讯科技（深圳）有限公司 | Image matching method and device, electronic equipment and storage medium |
| CN111383429A (en) * | 2020-03-04 | 2020-07-07 | 西安咏圣达电子科技有限公司 | Method, system, device and storage medium for detecting dress of workers in construction site |
| CN112071421A (en) * | 2020-09-01 | 2020-12-11 | 深圳高性能医疗器械国家研究院有限公司 | Deep learning estimation method and application thereof |
| CN112418261B (en) * | 2020-09-17 | 2022-05-03 | 电子科技大学 | Human body image multi-attribute classification method based on prior prototype attention mechanism |
| CN112418261A (en) * | 2020-09-17 | 2021-02-26 | 电子科技大学 | Human body image multi-attribute classification method based on prior prototype attention mechanism |
| CN112308115B (en) * | 2020-09-25 | 2023-05-26 | 安徽工业大学 | Multi-label image deep learning classification method and equipment |
| CN112308115A (en) * | 2020-09-25 | 2021-02-02 | 安徽工业大学 | Multi-label image deep learning classification method and equipment |
| CN112419292A (en) * | 2020-11-30 | 2021-02-26 | 深圳云天励飞技术股份有限公司 | Pathological image processing method and device, electronic equipment and storage medium |
| CN112419292B (en) * | 2020-11-30 | 2024-03-26 | 深圳云天励飞技术股份有限公司 | Pathological image processing method and device, electronic equipment and storage medium |
| CN113609323A (en) * | 2021-07-20 | 2021-11-05 | 上海德衡数据科技有限公司 | Image dimension reduction method and system based on neural network |
| CN113609323B (en) * | 2021-07-20 | 2024-04-23 | 上海德衡数据科技有限公司 | Image dimension reduction method and system based on neural network |
| CN113705361A (en) * | 2021-08-03 | 2021-11-26 | 北京百度网讯科技有限公司 | Method and device for detecting model in living body and electronic equipment |
| CN114202805A (en) * | 2021-11-24 | 2022-03-18 | 北京百度网讯科技有限公司 | Living body detection method, living body detection device, electronic apparatus, and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740686A (en) | A kind of deep learning image multiple labeling classification method based on pool area and Fusion Features | |
| El Achi et al. | Automated diagnosis of lymphoma with digital pathology images using deep learning | |
| Wang et al. | OAENet: Oriented attention ensemble for accurate facial expression recognition | |
| Krause et al. | The unreasonable effectiveness of noisy data for fine-grained recognition | |
| Pan et al. | Object detection based on saturation of visual perception | |
| Wu et al. | Research on image text recognition based on canny edge detection algorithm and k-means algorithm | |
| CN110533024B (en) | Double-quadratic pooling fine-grained image classification method based on multi-scale ROI (region of interest) features | |
| CN103020971A (en) | Method for automatically segmenting target objects from images | |
| CN111652273A (en) | Deep learning-based RGB-D image classification method | |
| CN111553351A (en) | Semantic segmentation based text detection method for arbitrary scene shape | |
| CN106203448A (en) | A kind of scene classification method based on Nonlinear Scale Space Theory | |
| CN109034213A (en) | Hyperspectral image classification method and system based on joint entropy principle | |
| Hu et al. | Research on the cascade vehicle detection method based on CNN | |
| Yeasmin et al. | Image classification for identifying social gathering types | |
| Weng et al. | Traffic scene perception based on joint object detection and semantic segmentation | |
| Wei et al. | Defective samples simulation through neural style transfer for automatic surface defect segment | |
| Torres | A framework for the unsupervised and semi-supervised analysis of visual frames | |
| CN105844299B (en) | A kind of image classification method based on bag of words | |
| Lu et al. | HEp-2 cell image classification method based on very deep convolutional networks with small datasets | |
| Zhou et al. | Semantic image segmentation using low-level features and contextual cues | |
| Chen et al. | ADOSMNet: a novel visual affordance detection network with object shape mask guided feature encoders | |
| CN114299342B (en) | Unknown mark classification method in multi-mark picture classification based on deep learning | |
| Wei et al. | Multiscale feature U-Net for remote sensing image segmentation | |
| Boroujerdi et al. | Deep interactive region segmentation and captioning | |
| Li et al. | Dermoscopy lesion classification based on GANs and a fuzzy rank-based ensemble of CNN models |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190510 |