CN110110780B - Image classification method based on antagonistic neural network and massive noise data - Google Patents
Image classification method based on antagonistic neural network and massive noise data Download PDFInfo
- Publication number
- CN110110780B CN110110780B CN201910358002.0A CN201910358002A CN110110780B CN 110110780 B CN110110780 B CN 110110780B CN 201910358002 A CN201910358002 A CN 201910358002A CN 110110780 B CN110110780 B CN 110110780B
- Authority
- CN
- China
- Prior art keywords
- data
- network
- neural network
- loss
- label
- 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
Images
Classifications
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
A picture classification method based on an anti-neural network and massive noise data solves the problem of insufficient data when a picture classification task is performed by using deep learning. The method comprises the steps of simultaneously using a network image and an image in a standard data set as training data to be sent into a convolutional neural network, and respectively representing the category and the data source on the image content by using a category label and a source label; and respectively calculating the label prediction loss and the source identification loss to obtain joint loss, and enabling the performances of the network data and the standard data to tend to be consistent by minimizing the joint loss. The method can be used for various fine classification tasks of computer vision, and a large amount of network data can be used as training data to effectively improve the classification performance of the convolutional neural network. In the method, a countermeasure strategy is adopted in the training process, so that the model is insensitive to the source of data, and network data and standard data can be well mixed together.
Description
Technical Field
The invention belongs to the technical field of computer vision, and relates to a method for training a convolutional neural network by using a network image, in particular to a picture classification method based on an antagonistic neural network and massive noise data.
Background
In recent years, deep learning has been greatly advanced in many computer vision tasks such as image classification, target detection, scene recognition, etc., and many image classification tasks based on convolutional neural networks have been achieved with remarkable results. However, training of the convolutional neural network requires a large number of image data sets with good labels, and manual labeling not only faces high cost but also requires a large amount of time, and large-scale data needs to be collected and labeled again for each new classification task.
As an alternative, using images in a network allows for a faster and easier collection of large amounts of image data. Although the network data inevitably carries some noise, a huge amount of network data can make up for this drawback. Document 1 shows that fine tuning a convolutional neural network using extensive network data is more efficient than using only a small standard dataset (i.e., a manually labeled dataset). Document 2 proposes a two-stage method of training a depth model using noise network data and neural network portability. Document 3 proposes a method for expanding a training set by collecting a large number of local regions from a large number of network images, which can generate a feature representation of a convolutional neural network with more discriminativity, and thereby improve classification accuracy. Document 4 processes noise data using a probabilistic framework to model the relationship between an image, a clean label, and a noise label in an end-to-end structure.
In general, the content of network images is typically more complex than standard data sets. For example, a network image contains a plurality of objects and includes only one target object therein, the target object is located at the edge of the image, or the size is small, which makes it difficult to distinguish the target objects. Therefore, the network image is noisy and has a gap from a well-labeled image, which may be caused by the factors of richer network image content, inconsistent position and scale of the object with the standard data set, noise label of the image, and the like. Currently, studies using network images for picture classification generally focus only on removing noisy data, and ignore the gap between network data and standard data sets.
Reference documents:
document 1: gong, d., and Wang, d.z. extracting visual knowledge from the web with multimodal learning. In IJCAI,2017.
Document 2: chen, X, and Gupta, A.Webly super search of connected networks. InICCV,2015.
Document 3: xu, z; huang, s.; zhang, y.; and Tao, d.assessment string supervision web data for fine-grained harvesting in CVPR,2015.
Document 4: xiao, t.; xia, t.; yang, y.; huang, c.; and Wang, x.learning from a mobile device equipped data for image classification. In ICCV,2015.
Disclosure of Invention
The invention provides a good solution for the problem of less data faced by the image classification task by training the convolutional neural network by utilizing the easily-obtained and massive network images. But the images in the network are noisy, on one hand because pictures in parts of the network may be completely irrelevant to the classification task, and on the other hand, due to the dimensions, positions and background clutter of objects in the images, gaps exist between the network images and the standard data set. Noise data prevalent in networks limits the improvement of classification results, so much existing work is done to investigate how to remove content-independent images for the first problem, while the second problem of gaps is not currently of interest to researchers.
The invention aims to solve the problem of difference between a network image and a standard data set, and utilizes network data with noise to classify pictures.
The technical scheme of the invention is as follows:
a picture classification method based on a countermeasure neural network and massive noise data comprises the following steps:
a. simultaneously taking a mass network image with noise and an image in a standard data set as training data to be sent into a convolutional neural network, and respectively representing the category of image content and the source of the image (network data or standard data) by using a category label and a source label for an image sample of each training data; the category label represents the category of the sub-classification task and is used for calculating the label prediction loss; the source label indicates whether the source of the image is a network or a standard data set and is used for calculating source identification loss;
b. according to the last layer of characteristics of the convolutional neural network in the step a, calculating label prediction loss for image categories by using a Softmax function, wherein the label prediction loss is smaller when the recognition accuracy of a classification task is higher;
in order to realize image classification and minimize label prediction loss to optimize a picture classification task, the higher the classification accuracy is, the smaller the label prediction loss is, and the lower the classification accuracy is, the larger the label prediction loss is;
c. b, calculating source identification loss by using a cross entropy loss function according to the last layer of characteristics of the convolutional neural network in the step a and based on the thought of the antagonistic neural network, wherein in order to learn from network data, the classification accuracy of data sources needs to be as low as possible, so that the value of the source identification loss needs to be increased;
in order to efficiently utilize network data with noise and maximize source identification loss to optimize a source classification task, the lower the classification accuracy, the greater the source identification loss, and the higher the classification accuracy, the smaller the source identification loss;
d. calculating to obtain a joint loss according to the label prediction loss in the step b and the source identification loss in the step c, taking a negative number from the value of the source identification loss and weighting and summing the negative number and the label prediction loss to obtain the joint loss so as to facilitate network optimization;
e. and (d) in the process of training the convolutional neural network, the joint loss in the step d is minimized, the network model tends to better classify the picture categories, and meanwhile, the performance of the network data and the performance of the standard data tend to be consistent, so that the influence of noise is weakened, the method is used for picture classification, and the accuracy of picture category label classification is improved.
For the convolutional neural network trained by using the network image with noise, the performance of the network data and the performance of the standard data gradually tend to be consistent by optimizing the joint loss, so that the number of training samples is increased, the influence of the noise is weakened, and the classification performance is improved.
The invention has the advantages and beneficial effects that: the method can effectively eliminate the difference between the network data set and the standard data set, can be applied to different data sets and network models, and has certain robustness. For the phenomenon of insufficient data when the convolutional neural network is used for carrying out the picture classification task, the method can rapidly utilize massive data in the network at low cost. In Food-101, dog-120 and Indor-67 data sets, compared with a model trained only by using standard data, the method provided by the invention has better classification accuracy. In summary, the invention starts with a new point of view to deal with the gap between the network data and the standard data, and gradually makes the network data and the standard data consistent in the process of training the convolutional neural network, so as to reduce the influence of noise data.
Drawings
Fig. 1 is a flow chart of a picture classification method based on a countering neural network and massive noise data.
Fig. 2 is a schematic diagram of a network structure of a picture classification method based on a countering neural network and massive noise data. In the figure, the left area is a network image used for training and an image in a standard data set, the upper area on the right side is a network structure for carrying out image classification by using a general convolutional neural network, and the lower area on the right side is a network structure designed by the image classification method based on the anti-neural network and massive noise data.
FIG. 3 is a diagram illustrating the influence of the value of the parameter λ in the present invention.
Detailed Description
The following describes embodiments of the present invention in further detail with reference to the accompanying drawings. The following examples are intended to illustrate the invention but are not intended to limit the scope of the invention.
As shown in fig. 1, a flow chart of a picture classification method based on a countering neural network and massive noise data includes the following specific steps:
step one, according to a classification task of a standard data set, massive data with noise are collected from a network to form a network data set with the same category as the standard data set, and the standard data set and the network data set are sent to a convolutional neural network together for network training. Specifically, the invention verifies the effectiveness of the method in three standard data sets of Food-101, dog-120 and Indor-67, wherein the three standard data sets respectively comprise 101000 images, 20580 images and 15620 images, and correspondingly, the network images of 240096 (Food), 52115 (Dog) and 76907 (Indoor scene) are respectively used as the network data sets. Each of which has two kinds of labels, namely a category label and a source label. The class label is an object class label in the image classification task and is used for calculating the label prediction loss; the source tag is used to mark whether the source of the image is a network data set or a standard data set for calculating a source identification loss.
And step two, in the convolutional neural network training process, extracting the characteristics of the last layer of full connection layer of the network model, and calculating the label prediction loss between the predicted label and the actual real label for the image category by using a Softmax function.
And thirdly, extracting the characteristics of the last full connection layer of the network model in the training process of the convolutional neural network, and calculating the source identification loss between the predicted source label and the actual source label by using a cross entropy loss function based on the thought of the antagonistic neural network.
Step four, according to a formula L = L c -λL d Loss of source identification L d Predicting loss L by adopting a weighted summation mode and a label after taking a negative value c Adding to obtain combined loss L, wherein the over-parameter lambda controls the specific gravity of the two losses, and the lambda takes different values between 0 and 0.6 respectively to carry out experiments, and determining the value according to the experimental result, as shown in FIG. 3. When the joint loss is smaller, the recognition accuracy of the fine classification task is higher, and the accuracy of the data source classification task is lower, namely the network data and the standard data tend to be consistent in performance.
And fifthly, in the process of training the convolutional neural network, optimizing the convolutional neural network with the minimum joint loss as a target until the expression of the network model and the value of the joint loss tend to be stable. The resulting network model learns knowledge from a large amount of noisy network data and is used for classification of images.
Test sets of three standard data sets of Food-101, dog-120 and Indor-67 are respectively sent into a trained network model for classification according to the steps, and classification accuracy rates of 89.35%, 87.07% and 84.59% are respectively obtained. While other prior art methods only achieve classification results on these three datasets up to 88.28% (Hassmanne jad, H.; matrella, G.; ciampolini, P.; deMunari, I.; mordonini, M.; and Cagnoni, S.2016.Food recognition using polypeptide discrete networks), 85.90% (Krause, J.; sapp, B.; hod, A.; zhou, H.; toshev, A.; duerig, T.; philibin, J.; fei-Fei, L.2016.The unreasenable expression of information data of finish-simulation. In.) and 83.75% (ECgung, S.2016. W., W.103.; W.S.sub., W.26; W.sub.sub.sub.26) of the classification results of the classification of the I., L.2016. Sub.sub.sub.26). In contrast, the present invention can effectively utilize massive noise data in the network for image classification.
Fig. 2 shows a schematic diagram of the method, in which the core problem of the algorithm at each stage, the training process, and the system input and output are visually described. Fig. 2 and fig. 1 have the same meaning, but different levels of abstraction, which mainly help to explain the parts in fig. 1.
FIG. 3 shows the classification accuracy when the parameter λ takes different values, and the method of the present invention was used to take different values between 0 and 0.6 for λ in three standard data sets of Food-101, dog-120 and Indor-67 and experiments were performed. The three polylines in the graph correspond to the experimental results on the Food-101, dog-120 and Indor-67 data sets from top to bottom. When the value of the parameter lambda is zero, the network is equivalent to a general unmodified convolutional neural network, and experiments show that the image classification effect of a standard data set by directly utilizing a network image with noise is not ideal. When the value of the parameter lambda is too large, the network is more inclined to mix the images from two different sources, so that the experimental result is not good. According to the experimental result, when the value of the parameter lambda is 0.1, the method obtains the highest classification accuracy in three different classification tasks, and three broken lines in the graph reach the peak value.
Claims (3)
1. A picture classification method based on an antagonistic neural network and massive noise data is used for training a convolutional neural network to enable network data to be consistent with standard data so as to improve the classification performance of a model, and is characterized by comprising the following steps:
a. simultaneously taking the massive network images with noise and the images in the standard data set as training data to be sent to a convolutional neural network, and respectively representing the category of the image content and the data source of the image by using a category label and a source label for an image sample of each training data;
b. b, according to the last layer of characteristics of the convolutional neural network in the step a, calculating label prediction loss for the image category by using a Softmax function, wherein the label prediction loss is smaller when the recognition accuracy of the classification task is higher;
c. according to the last layer of characteristics of the convolutional neural network in the step a, based on the thought of the antagonistic neural network, a cross entropy loss function is used for calculating source identification loss, and in order to learn from network data, the classification accuracy of data sources needs to be reduced, so that the value of the source identification loss needs to be increased;
d. calculating to obtain combined loss according to the label prediction loss in the step b and the source identification loss in the step c, and according to a formula L = L in order to facilitate network optimization c -λL d Loss of source identification L d After taking negative number, the value of (A) is weighted and summed, and the loss L is predicted by the label c Adding to obtain a combined loss L, wherein the ratio of the two losses is controlled by a super parameter lambda, and the value range is 0-0.6;
e. and (e) in the process of training the convolutional neural network, the joint loss in the step (d) is minimized, the network model tends to better classify the picture categories, and meanwhile, the performances of the network data and the standard data tend to be consistent, so that the influence of noise is weakened, and the accuracy of picture category label classification is improved.
2. The picture classification method based on the countermeasure neural network and the massive noise data as claimed in claim 1, wherein: the category label in the step a represents the category of the sub-classification task, and the source label represents whether the source of the image is a network or a standard data set.
3. The picture classification method based on the countermeasure neural network and the massive noise data as claimed in claim 1, wherein: the picture classification method based on the anti-neural network and the massive noise data aims at solving the noise problem caused by the difference between the network picture and the pictures in the standard data set due to the fact that the massive pictures from the network are noisy and the scale, the position and the background environment of the objects in the images, so that the difference is reduced, and the picture classification result is improved.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910358002.0A CN110110780B (en) | 2019-04-30 | 2019-04-30 | Image classification method based on antagonistic neural network and massive noise data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910358002.0A CN110110780B (en) | 2019-04-30 | 2019-04-30 | Image classification method based on antagonistic neural network and massive noise data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110110780A CN110110780A (en) | 2019-08-09 |
| CN110110780B true CN110110780B (en) | 2023-04-07 |
Family
ID=67487578
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910358002.0A Active CN110110780B (en) | 2019-04-30 | 2019-04-30 | Image classification method based on antagonistic neural network and massive noise data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110110780B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111242131B (en) * | 2020-01-06 | 2024-05-10 | 北京十六进制科技有限公司 | Method, storage medium and device for identifying images in intelligent paper reading |
| CN111299763B (en) * | 2020-02-28 | 2021-09-21 | 华南理工大学 | Anti-noise-interference laser visual welding seam automatic tracking method and system |
| CN111507419B (en) * | 2020-04-22 | 2022-09-30 | 腾讯科技（深圳）有限公司 | Training method and device of image classification model |
| CN111784595B (en) * | 2020-06-10 | 2023-08-29 | 北京科技大学 | Dynamic tag smooth weighting loss method and device based on historical record |
| CN113283578B (en) * | 2021-04-14 | 2024-07-23 | 南京大学 | Data denoising method based on marker risk control |
| CN115099115B (en) * | 2022-08-24 | 2022-12-30 | 国网江西省电力有限公司电力科学研究院 | Indoor substation ventilation noise reduction deep reinforcement learning design method |
| CN115618935B (en) * | 2022-12-21 | 2023-05-05 | 北京航空航天大学 | Robustness loss function searching method and system for classification task tag noise |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH0773329A (en) * | 1993-04-29 | 1995-03-17 | Canon Inc | Method and device for processing image |
| CN103984959A (en) * | 2014-05-26 | 2014-08-13 | 中国科学院自动化研究所 | Data-driven and task-driven image classification method |
| DE102015009981A1 (en) * | 2015-07-31 | 2017-02-02 | Eberhard Karls Universität Tübingen | Method and apparatus for image synthesis |
| CN106682694A (en) * | 2016-12-27 | 2017-05-17 | 复旦大学 | Sensitive image identification method based on depth learning |
| CN107220600A (en) * | 2017-05-17 | 2017-09-29 | 清华大学深圳研究生院 | A kind of Picture Generation Method and generation confrontation network based on deep learning |
| CN107463953A (en) * | 2017-07-21 | 2017-12-12 | 上海交通大学 | Image classification method and system based on quality insertion in the case of label is noisy |
| CN108257347A (en) * | 2018-01-10 | 2018-07-06 | 安徽大学 | A kind of flame image sequence sorting technique and device using convolutional neural networks |
| CN108416382A (en) * | 2018-03-01 | 2018-08-17 | 南开大学 | One kind is based on iteration sampling and a pair of of modified Web graph of multi-tag as training convolutional neural networks method |
| KR101913952B1 (en) * | 2017-07-11 | 2018-10-31 | 경북대학교 산학협력단 | Automatic Recognition Method of iPSC Colony through V-CNN Approach |
| CN108805188A (en) * | 2018-05-29 | 2018-11-13 | 徐州工程学院 | A kind of feature based recalibration generates the image classification method of confrontation network |
| CN108960127A (en) * | 2018-06-29 | 2018-12-07 | 厦门大学 | Pedestrian's recognition methods again is blocked based on the study of adaptive depth measure |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9704068B2 (en) * | 2012-06-22 | 2017-07-11 | Google Inc. | System and method for labelling aerial images |
| CN107437100A (en) * | 2017-08-08 | 2017-12-05 | 重庆邮电大学 | A kind of picture position Forecasting Methodology based on the association study of cross-module state |
| CN107729992B (en) * | 2017-10-27 | 2020-12-29 | 深圳市未来媒体技术研究院 | Deep learning method based on back propagation |
| CN108229404B (en) * | 2018-01-09 | 2022-03-08 | 东南大学 | Radar echo signal target identification method based on deep learning |
| CN109325527A (en) * | 2018-09-05 | 2019-02-12 | 哈尔滨工程大学 | A kind of sidescan-sonar image target automatic classification method based on transfer learning and deep learning |
| CN109614979B (en) * | 2018-10-11 | 2023-05-02 | 北京大学 | Data augmentation method and image classification method based on selection and generation |
-
2019
- 2019-04-30 CN CN201910358002.0A patent/CN110110780B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH0773329A (en) * | 1993-04-29 | 1995-03-17 | Canon Inc | Method and device for processing image |
| CN103984959A (en) * | 2014-05-26 | 2014-08-13 | 中国科学院自动化研究所 | Data-driven and task-driven image classification method |
| DE102015009981A1 (en) * | 2015-07-31 | 2017-02-02 | Eberhard Karls Universität Tübingen | Method and apparatus for image synthesis |
| CN106682694A (en) * | 2016-12-27 | 2017-05-17 | 复旦大学 | Sensitive image identification method based on depth learning |
| CN107220600A (en) * | 2017-05-17 | 2017-09-29 | 清华大学深圳研究生院 | A kind of Picture Generation Method and generation confrontation network based on deep learning |
| KR101913952B1 (en) * | 2017-07-11 | 2018-10-31 | 경북대학교 산학협력단 | Automatic Recognition Method of iPSC Colony through V-CNN Approach |
| CN107463953A (en) * | 2017-07-21 | 2017-12-12 | 上海交通大学 | Image classification method and system based on quality insertion in the case of label is noisy |
| CN108257347A (en) * | 2018-01-10 | 2018-07-06 | 安徽大学 | A kind of flame image sequence sorting technique and device using convolutional neural networks |
| CN108416382A (en) * | 2018-03-01 | 2018-08-17 | 南开大学 | One kind is based on iteration sampling and a pair of of modified Web graph of multi-tag as training convolutional neural networks method |
| CN108805188A (en) * | 2018-05-29 | 2018-11-13 | 徐州工程学院 | A kind of feature based recalibration generates the image classification method of confrontation network |
| CN108960127A (en) * | 2018-06-29 | 2018-12-07 | 厦门大学 | Pedestrian's recognition methods again is blocked based on the study of adaptive depth measure |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110110780A (en) | 2019-08-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110110780B (en) | Image classification method based on antagonistic neural network and massive noise data | |
| Chen et al. | Learning context flexible attention model for long-term visual place recognition | |
| CN109446970B (en) | Transformer substation inspection robot road scene recognition method based on deep learning | |
| CN101477633B (en) | Method for automatically estimating visual significance of image and video | |
| CN110009008B (en) | Method for automatically classifying immune fixed electrophoretogram based on extracted features | |
| CN110751195B (en) | Fine-grained image classification method based on improved YOLOv3 | |
| CN110705412A (en) | Video target detection method based on motion history image | |
| CN114596273B (en) | Intelligent detection method for multiple defects of ceramic substrate by using YOLOV4 network | |
| CN114155474A (en) | Damage identification technology based on video semantic segmentation algorithm | |
| CN115205727A (en) | Experiment intelligent scoring method and system based on unsupervised learning | |
| CN114612450B (en) | Image detection segmentation method and system based on data augmentation machine vision and electronic equipment | |
| CN115019039A (en) | Example segmentation method and system combining self-supervision and global information enhancement | |
| CN113673534B (en) | RGB-D image fruit detection method based on FASTER RCNN | |
| CN111832497B (en) | Text detection post-processing method based on geometric features | |
| CN116912184B (en) | Weak supervision depth restoration image tampering positioning method and system based on tampering area separation and area constraint loss | |
| CN111062388B (en) | Advertisement character recognition method, system, medium and equipment based on deep learning | |
| CN115359411B (en) | Substation environment understanding method based on improved deep Lab V < 3+ > network | |
| CN111369124A (en) | Image aesthetic prediction method based on self-generation global features and attention | |
| CN105373802A (en) | Scene image classification method based on interval Type-2 fuzzy support vector machine | |
| Wu et al. | Express parcel detection based on improved faster regions with CNN features | |
| Praganingrum et al. | Image Processing Applications in Construction Projects: Challenges and Opportunities | |
| Prabhakaran et al. | Comprehensive Analysis of Defect Detection Through Image Processing and Machine Learning for Photovoltaic Panels | |
| CN116188774B (en) | Hyperspectral image instance segmentation method and building instance segmentation method | |
| CN116468960B (en) | Video image analysis and retrieval method and system | |
| CN118429389B (en) | Target tracking method and system based on multiscale aggregation attention feature extraction network |
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 |