CN107766890B - Improved method for discriminant graph block learning in fine-grained identification - Google Patents
Improved method for discriminant graph block learning in fine-grained identification Download PDFInfo
- Publication number
- CN107766890B CN107766890B CN201711040828.XA CN201711040828A CN107766890B CN 107766890 B CN107766890 B CN 107766890B CN 201711040828 A CN201711040828 A CN 201711040828A CN 107766890 B CN107766890 B CN 107766890B
- Authority
- CN
- China
- Prior art keywords
- image
- discriminant
- vector
- multiplied
- detector
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 31
- 239000013598 vector Substances 0.000 claims abstract description 49
- 230000004044 response Effects 0.000 claims abstract description 22
- 238000011176 pooling Methods 0.000 claims abstract description 9
- 238000013528 artificial neural network Methods 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 9
- 238000001514 detection method Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 3
- 238000009825 accumulation Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 238000010586 diagram Methods 0.000 claims description 2
- 238000013527 convolutional neural network Methods 0.000 description 9
- 238000002372 labelling Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241000282472 Canis lupus familiaris Species 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
An improved method for discriminant graph block learning in fine-grained recognition comprises the following steps: extracting image blocks with distinguishing properties in an original image, comprising the following steps: obtaining a feature map from the original image through a convolution pooling layer in a convolution neural network, and regarding a vector of each space fixed position in the feature map as a detector corresponding to a corresponding position image block in the original image; supposing that a detector with the highest response in a discriminant area in an original image is learned, carrying out convolution operation on the detector and a feature map to obtain a new response map; selecting the position with the maximum value in the new response image to obtain an image block with distinguishing property; learning features of tiles having discriminative properties and for classification, comprising: obtaining a local saliency map according to an image block with distinguishing properties; the local saliency map is encoded using a spatially weighted fisher vector. The method learns the discriminant characteristics more suitable for a fine-grained identification task, and reduces the interference of background information in a discriminant image block so as to improve the classification precision.
Description
Technical Field
The invention relates to discriminative graph block learning in fine-grained identification. In particular, to an improved method for discriminant tile learning in fine-grained identification by spatially weighting an image descriptor to obtain spatially weighted fisher vectors according to a response map.
Background
In recent years, fine-grained recognition attracts more and more attention in the field of target recognition, which is to recognize subclasses of a certain large class of targets, such as flowers, birds, dogs, automobiles and the like, which generally have the same structure, so that how to learn features with distinguishing properties in images becomes a main task of fine-grained recognition.
In past research, the field of fine-grained identification mainly includes two tasks: local localization and characterization. In addition to image category labels, fine-grained datasets usually provide additional labeling of target bounding boxes and local parts, and many previous tasks more or less rely on these additional labeling, but fine-grained classification usually requires expert-level knowledge, and ordinary people have difficulty in completing the task, which makes manual labeling expensive. In recent years, more research is focused on methods without any additional labeled information, the method related to the invention only needs image category labels, and does not need local labeling, and local features with discriminability are learned by a weak supervision method.
For the feature description of images, CNN features have made a breakthrough on many benchmarks. The traditional method is to encode local information and then fuse the local information into an integral feature representation, the CNN features are different from the CNN features, the global learning can be directly realized, a feature extractor does not need to be designed manually, and the current fine-grained identification method is based on the CNN and is used for learning the fine and unique features of the image through other algorithms.
Disclosure of Invention
The invention aims to solve the technical problem of providing an improved method for learning discriminant image blocks in fine-grained identification, which can accurately learn detailed features and abandon disordered background information in small image blocks, thereby improving classification precision and avoiding the need of global feature assistance.
The technical scheme adopted by the invention is as follows: an improved method for learning discriminant graph blocks in fine-grained recognition comprises the following steps:
1) extracting image blocks with distinguishing properties in an original image, comprising the following steps:
(1) obtaining a characteristic image of C multiplied by H multiplied by W size from an original image through a convolution pooling layer in a convolution neural network, wherein C is the number of pipelines, H is the height, W is the width, and a C multiplied by 1 vector of each space fixed position in the characteristic image is regarded as a detector corresponding to a corresponding position image block in the original image;
(2) supposing that a detector with the highest response in a discriminant area in an original image is learned, carrying out convolution operation on the detector with the size of C multiplied by 1 and a feature map with the size of C multiplied by H multiplied by W to obtain a new response map with the size of H multiplied by W;
(3) selecting the position with the maximum value in the new H multiplied by W response image to obtain a 1 multiplied by 1 image block with distinguishing property;
2) learning features of tiles having discriminative properties and for classification, comprising:
(1) obtaining a local saliency map according to an image block with distinguishing properties;
(2) the local saliency map is encoded using a spatially weighted fisher vector.
Step 1) step (2) the step of learning the detector with the highest response in a discriminant region in the original image is as follows:
(1) assuming that the number of discriminative pattern block detectors of each type of image is n, the images share M types, and the number of required detectors is nM;
(2) respectively performing convolution operation on nM Cx 1 x 1 detectors and Cx H x W feature maps to obtain new feature maps, and performing global maximum pooling on the new feature maps to obtain a nM-dimension feature vector;
(3) averaging each class of eigenvectors in the nM dimensional eigenvectors to obtain an M dimensional vector;
(4) introducing the M-dimensional vector into a Softmax loss function, training a Cx 1 x 1 detector by using a back propagation algorithm, and obtaining a detector with the highest response in a discriminant region in an original image after the training is finished;
step 2) the step (1) of obtaining the local saliency map is as follows: a local saliency map Q is computed from a saliency map S derived from the original image, as follows:
where p is the pixel of the discriminative tile, i is the detection site, and when the ith detection site contains pixel p, then Di(p) 1, otherwise Di(p) 0, s (p) the saliency map of the whole image, q (p) the local saliency map, Z a normalized constant such that maxq (p) 1.
Step 2) the step (2) comprises the following steps:
suppose vector I is (z)1,…,zN) Is a series of D-dimension characteristic vectors extracted from an image, and the Fisher vector of a picture I encodes phi (I) to (u)1,v1,…,uk,vk) Is the mean square error ukSum covariance vkAccumulation of (u)k，vkWritten as follows:
where j 1, D represents the vector dimension, and o (μ) is definedk,σk,πkK1, …, K) are parameters of a gaussian mixture model, qikIs each vector z of mode k in the hybrid modeliWherein i is 1, …, N.
For each vector ziIntroducing a spatial weighting term Q (p)i)，ujkAnd vjkThe weighted result is expressed as:
wherein, Q (p)i) Is a local saliency map, uijk、vijkRespectively, formula (2) (3), important features can be learned by introducing spatial weights.
The improved method for learning the discriminant image block in the fine-grained recognition combines the CNN characteristic and the Fisher vector to learn the discriminant characteristic more suitable for a fine-grained recognition task, and reduces the interference of background information in the discriminant image block so as to improve the classification precision. The method mainly aims at the improvement of the existing discriminant graphic block learning, and for the discriminant region detected by the learned graphic block detector, except the discriminant feature of the target object, redundant background information often exists, so that the method utilizes the local saliency map and the Fisher vector coding to improve the limit, and fully utilizes the characteristics of the discriminant region to be effectively used for the classification task. The advantages are mainly reflected in that:
1) the novelty is as follows: the most efficient and popular feature representation method at present is CNN, but the present invention combines fisher vector coding with CNN features for specific problems. Because the most important thing in the fine-grained identification problem is the learning of the discriminant features, and the backgrounds in the data sets are usually very similar, the method introduces the Fisher vector in a novel way, and can effectively reduce the interference of the background information in the discriminant image blocks.
2) Effectiveness: compared with the original method, the local salient map and Fisher vector coding-based method designed by the invention can effectively learn local features, the traditional CNN usually needs a rectangle with a fixed size as input, which contains invalid information of the background, and the method can effectively reduce the interference of background noise, so that the learned local features are more discriminative, and the classification precision is improved.
3) The practicability is as follows: the method is simple and feasible, is an end-to-end network, and can be effectively used for fine-grained identification.
Drawings
FIG. 1 is a flow chart of an improved method of discriminative tile learning in fine-grained recognition of the present invention.
Detailed Description
1. An improved method for learning discriminant graph blocks in fine-grained recognition is characterized by comprising the following steps:
1) extracting image blocks with distinguishing properties in an original image, comprising the following steps:
(1) obtaining a characteristic image of C multiplied by H multiplied by W size from an original image through a convolution pooling layer in a convolution neural network, wherein C is the number of pipelines, H is the height, W is the width, and a C multiplied by 1 vector of each space fixed position in the characteristic image is regarded as a detector corresponding to a corresponding position image block in the original image;
(2) supposing that a detector with the highest response in a discriminant area in an original image is learned, carrying out convolution operation on the detector with the size of C multiplied by 1 and a feature map with the size of C multiplied by H multiplied by W to obtain a new response map with the size of H multiplied by W;
(3) selecting the position with the maximum value in the new H multiplied by W response image to obtain a 1 multiplied by 1 image block with distinguishing property;
2) learning features of tiles having discriminative properties and for classification, comprising:
(1) obtaining a local saliency map according to an image block with distinguishing properties;
(2) encoding the local saliency map using spatially weighted fisher vectors;
2. the improved method for discriminative patch learning in fine granularity recognition as claimed in claim 1, wherein the step 1) step (2) of learning the detector with the highest response in a discriminative region in the original image comprises the steps of:
(1) assuming that the number of discriminative pattern block detectors of each type of image is n, the images share M types, and the number of required detectors is nM;
(2) respectively performing convolution operation on nM Cx 1 x 1 detectors and Cx H x W feature maps to obtain new feature maps, and performing global maximum pooling on the new feature maps to obtain a nM-dimension feature vector;
(3) averaging each class of eigenvectors in the nM dimensional eigenvectors to obtain an M dimensional vector;
(4) introducing the M-dimensional vector into a Softmax loss function, training a Cx 1 x 1 detector by using a back propagation algorithm, and obtaining a detector with the highest response in a discriminant region in an original image after the training is finished;
3. the improved method for discriminative graph block learning in fine grain identification as claimed in claim 1, wherein the step 2) the step (1) is to obtain the local saliency map as follows: a local saliency map Q is computed from a saliency map S derived from the original image, as follows:
where p is the pixel of the discriminative tile, i is the detection site, and when the ith detection site contains pixel p, then Di(p) 1, otherwise Di(p) 0, s (p) the saliency map of the whole image, q (p) the local saliency map, Z a normalized constant such that maxq (p) 1.
4. The improved method for discriminative tile learning in fine granularity recognition as claimed in claim 1, wherein the step 2) and the step (2) comprise:
suppose vector I is (z)1,…,zN) Is a series of D-dimension characteristic vectors extracted from an image, and the Fisher vector of a picture I encodes phi (I) to (u)1,v1,…,uk,vk) Is the mean square error ukSum covariance vkAccumulation of (u)k，vkWritten as follows:
where j 1, D represents the vector dimension, and o (μ) is definedk,σk,πkK1, …, K) are parameters of a gaussian mixture model, qikIs each vector z of mode k in the hybrid modeliWherein i is 1, …, N.
For each vector ziIntroducing a spatial weighting term Q (p)i)，ujkAnd vjkThe weighted result is expressed as:
wherein, Q (p)i) Is a local saliency map, uijk、vijkRespectively, formula (2) (3), important features can be learned by introducing spatial weights.
Specific examples are given below in connection with fig. 1:
FIG. 1 depicts a flow diagram of the architecture of the present invention. The structure of the invention mainly comprises three parts, as shown in figure 1. The method is based on a VGG-16 model, and the model has 16 layers. The implementation process is divided into two stages: a training phase and a testing phase.
In the training stage, the parameters of the detector are mainly learned, and the process is shown as (i) and (ii) in fig. 1.
(1) First, the input image passes through a pre-trained convolutional neural network VGG-16, conv4-3 outputs a feature map with the size of 512 × 28 × 28, and therefore, the size of each detector is 512 × 1 × 1. Setting the number of detectors of each class to 10, there are 2000 detectors for the CUB200-2011 dataset;
(2) convolving each detector with the obtained 512 × 28 × 28 feature map to obtain a response map with the size of 28 × 28;
(3) after the response map is subjected to global maximum pooling, a 2000-dimensional feature vector is obtained,
(4) the feature vectors of each class in the 2000-dimensional vectors are averaged to obtain a 200-dimensional vector, the vector after the average value pooling is transmitted into a Softmax loss function to be trained through a back propagation algorithm, and then a detector capable of extracting discriminant image blocks for each class can be obtained.
And (3) a testing stage, as shown in (r) and (c) of FIG. 1. For the trained detector, the steps (1) to (3) in the training stage are repeated, so that a response graph with the size of 1 × 1 can be obtained, and the part with the distinguishing property of each picture can be identified. And then calculating a local saliency map of the image, wherein the local saliency map is obtained by two parts of the local map and the image saliency map, and a discriminant image block extracted from the original image is multiplied by the global saliency map, so as to obtain the local saliency map by the formula (1). The local saliency map is used to indicate the possibility that the pixel belongs to the foreground, which can effectively reduce the interference of the background. Designing weights for the Fisher vectors of the images according to the local saliency maps to obtain space-weighted Fisher vectors, and learning important features in a fine-grained identification task by introducing the weights so as to finally realize classification of the fine-grained images.
Claims (1)
1. An improved method for learning discriminant graph blocks in fine-grained recognition is characterized by comprising the following steps:
1) extracting image blocks with distinguishing properties in an original image, comprising the following steps:
(1) obtaining a characteristic diagram of C multiplied by H multiplied by W size from an original image through a convolution pooling layer in a convolution neural network, wherein C is the number of pipelines, H is the height, and W is the width;
(2) supposing that a detector with the highest response in a discriminant area in an original image is learned, carrying out convolution operation on the detector with the size of C multiplied by 1 and a feature map with the size of C multiplied by H multiplied by W to obtain a new response map with the size of H multiplied by W; the step of learning the detector with the highest response in a discriminant region in the original image is as follows:
(2.1) assuming that the number of discriminant pattern block detectors of each type of image is n, the images share M types, and the number of required detectors is nM;
(2.2) performing convolution operation on nM Cx 1 x 1 detectors and Cx H x W feature maps to obtain new feature maps, and performing global maximum pooling on the new feature maps to obtain an nM-dimension feature vector;
(2.3) averaging each type of feature vector in the feature vectors of nM dimension to obtain a vector of M dimension;
(2.4) transmitting the M-dimensional vector into a Softmax loss function, training a Cx 1 x 1 detector by using a back propagation algorithm, and obtaining a detector with the highest response in a discriminant area in an original image after the training is finished;
(3) selecting the position with the maximum value in the new H multiplied by W response image to obtain a 1 multiplied by 1 image block with distinguishing property;
2) learning features of tiles having discriminative properties and for classification, comprising:
(1) obtaining a local saliency map according to an image block with distinguishing properties; the local saliency map is obtained by: a local saliency map Q is computed from a saliency map S derived from the original image, as follows:
where p is the pixel of the discriminant block, m represents the mth detection position, m is 1, …, Α, and when the mth detection position contains pixel p, then D ism(p) 1, otherwise Dm(p) 0, s (p) is the saliency map of the whole image, q (p) is the local saliency map, Z is a normalized constant such that maxq (p) 1;
(2) encoding the local saliency map using spatially weighted fisher vectors; the method comprises the following steps:
suppose vector I is (z)1,…,zN) Is a series of D-dimension characteristic vectors extracted from an image, and the Fisher vector of a picture I encodes phi (I) to (u)1,v1,…,uk,vk) Is the mean square error ukSum covariance vkAccumulation of (u)k，vkWritten as follows:
where j 1, D represents the vector dimension, and o (μ) is definedk,σk,πkK1, …, K) are parameters of a gaussian mixture model, qikIs each vector z of mode k in the hybrid modeliWherein i ═ 1, …, N;
for each vector ziIntroducing a spatial weighting term Q (p)i)，ujkAnd vjkThe weighted result is expressed as:
wherein, Q (p)i) Is a local saliency map, uijk、vijkRespectively, formula (2) (3), important features can be learned by introducing spatial weights.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711040828.XA CN107766890B (en) | 2017-10-31 | 2017-10-31 | Improved method for discriminant graph block learning in fine-grained identification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711040828.XA CN107766890B (en) | 2017-10-31 | 2017-10-31 | Improved method for discriminant graph block learning in fine-grained identification |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107766890A CN107766890A (en) | 2018-03-06 |
| CN107766890B true CN107766890B (en) | 2021-09-14 |
Family
ID=61271840
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711040828.XA Expired - Fee Related CN107766890B (en) | 2017-10-31 | 2017-10-31 | Improved method for discriminant graph block learning in fine-grained identification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107766890B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR102102161B1 (en) * | 2018-05-18 | 2020-04-20 | 오드컨셉 주식회사 | Method, apparatus and computer program for extracting representative feature of object in image |
| CN110019915B (en) * | 2018-07-25 | 2022-04-12 | 北京京东尚科信息技术有限公司 | Method and device for detecting picture and computer readable storage medium |
| CN109409384A (en) * | 2018-09-30 | 2019-03-01 | 内蒙古科技大学 | Image-recognizing method, device, medium and equipment based on fine granularity image |
| CN109815973A (en) * | 2018-12-07 | 2019-05-28 | 天津大学 | A kind of deep learning method suitable for the identification of fish fine granularity |
| CN109948628B (en) * | 2019-03-15 | 2023-01-03 | 中山大学 | Target detection method based on discriminant region mining |
| CN110197202A (en) * | 2019-04-30 | 2019-09-03 | 杰创智能科技股份有限公司 | A kind of local feature fine granularity algorithm of target detection |
| CN110309858B (en) * | 2019-06-05 | 2022-07-01 | 大连理工大学 | Fine-grained image classification method based on discriminant learning |
| CN110363233B (en) * | 2019-06-28 | 2021-05-28 | 西安交通大学 | Fine-grained image recognition method and system of convolutional neural network based on block detector and feature fusion |
| CN110796183A (en) * | 2019-10-17 | 2020-02-14 | 大连理工大学 | Weak supervision fine-grained image classification algorithm based on relevance-guided discriminant learning |
| CN111062438B (en) * | 2019-12-17 | 2023-06-16 | 大连理工大学 | Image propagation weak supervision fine granularity image classification algorithm based on correlation learning |
| CN112927221B (en) * | 2020-12-09 | 2022-03-29 | 广州市玄武无线科技股份有限公司 | Image fine-grained feature-based reproduction detection method and system |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104573744A (en) * | 2015-01-19 | 2015-04-29 | 上海交通大学 | Fine granularity classification recognition method and object part location and feature extraction method thereof |
| WO2016168235A1 (en) * | 2015-04-17 | 2016-10-20 | Nec Laboratories America, Inc. | Fine-grained image classification by exploring bipartite-graph labels |
| CN106778705A (en) * | 2017-02-04 | 2017-05-31 | 中国科学院自动化研究所 | A kind of pedestrian's individuality dividing method and device |
| CN106778807A (en) * | 2016-11-22 | 2017-05-31 | 天津大学 | The fine granularity image classification method of dictionary pair is relied on based on public dictionary pair and class |
| CN106778804A (en) * | 2016-11-18 | 2017-05-31 | 天津大学 | The zero sample image sorting technique based on category attribute transfer learning |
| CN106951872A (en) * | 2017-03-24 | 2017-07-14 | 江苏大学 | A kind of recognition methods again of the pedestrian based on unsupervised depth model and hierarchy attributes |
-
2017
- 2017-10-31 CN CN201711040828.XA patent/CN107766890B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104573744A (en) * | 2015-01-19 | 2015-04-29 | 上海交通大学 | Fine granularity classification recognition method and object part location and feature extraction method thereof |
| WO2016168235A1 (en) * | 2015-04-17 | 2016-10-20 | Nec Laboratories America, Inc. | Fine-grained image classification by exploring bipartite-graph labels |
| CN106778804A (en) * | 2016-11-18 | 2017-05-31 | 天津大学 | The zero sample image sorting technique based on category attribute transfer learning |
| CN106778807A (en) * | 2016-11-22 | 2017-05-31 | 天津大学 | The fine granularity image classification method of dictionary pair is relied on based on public dictionary pair and class |
| CN106778705A (en) * | 2017-02-04 | 2017-05-31 | 中国科学院自动化研究所 | A kind of pedestrian's individuality dividing method and device |
| CN106951872A (en) * | 2017-03-24 | 2017-07-14 | 江苏大学 | A kind of recognition methods again of the pedestrian based on unsupervised depth model and hierarchy attributes |
Non-Patent Citations (3)
| Title |
|---|
| Fully Convolutional Attention Localization Networks: Efficient Attention Localization for Fine-Grained Recognition;Xiao Liu 等;《arXiv》;20160404;第1-10页 * |
| Picking Deep Filter Responses for Fine-grained Image Recognition;Xiaopeng Zhang 等;《2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)》;20161212;论文第1-5节 * |
| 自上而下注意图分割的细粒度图像分类;冯语姗 等;《中国图象图形学报》;20160930;第1147-1154页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107766890A (en) | 2018-03-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107766890B (en) | Improved method for discriminant graph block learning in fine-grained identification | |
| CN111583263B (en) | Point cloud segmentation method based on joint dynamic graph convolution | |
| CN103605972B (en) | Non-restricted environment face verification method based on block depth neural network | |
| CN104820718B (en) | Image classification and search method based on geographic location feature Yu overall Vision feature | |
| CN110717411A (en) | Pedestrian re-identification method based on deep layer feature fusion | |
| JP7439153B2 (en) | Lifted semantic graph embedding for omnidirectional location recognition | |
| CN105825502B (en) | A kind of Weakly supervised method for analyzing image of the dictionary study based on conspicuousness guidance | |
| CN108062574A (en) | A kind of Weakly supervised object detection method based on particular category space constraint | |
| CN105184298A (en) | Image classification method through fast and locality-constrained low-rank coding process | |
| CN101551809A (en) | Search method of SAR images classified based on Gauss hybrid model | |
| CN109145832A (en) | Polarimetric SAR image semisupervised classification method based on DSFNN Yu non local decision | |
| CN103927511A (en) | Image identification method based on difference feature description | |
| US20210319242A1 (en) | Dense and Discriminative Neural Network Architectures for Improved Object Detection and Instance Segmentation | |
| CN114329031B (en) | Fine-granularity bird image retrieval method based on graph neural network and deep hash | |
| CN111860297A (en) | SLAM loop detection method applied to indoor fixed space | |
| CN109492610B (en) | Pedestrian re-identification method and device and readable storage medium | |
| CN115690152A (en) | Target tracking method based on attention mechanism | |
| CN106874862A (en) | People counting method based on submodule technology and semi-supervised learning | |
| CN112232297A (en) | Remote sensing image scene classification method based on depth joint convolution activation | |
| CN114973305B (en) | Accurate human body analysis method for crowded people | |
| CN115565080A (en) | Multi-feature fusion SAR image sea ice classification method based on self-attention mechanism | |
| CN117011701A (en) | Remote sensing image feature extraction method for hierarchical feature autonomous learning | |
| Zhu et al. | Adaptive alignment network for person re-identification | |
| CN116935196A (en) | Novel brain inspired target detection method for blocked target in complex environment | |
| Schenkel et al. | Domain adaptation for semantic segmentation using convolutional neural networks |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210914 |