CN113780284A - Logo detection method based on target detection and metric learning - Google Patents
Logo detection method based on target detection and metric learning Download PDFInfo
- Publication number
- CN113780284A CN113780284A CN202111090918.6A CN202111090918A CN113780284A CN 113780284 A CN113780284 A CN 113780284A CN 202111090918 A CN202111090918 A CN 202111090918A CN 113780284 A CN113780284 A CN 113780284A
- Authority
- CN
- China
- Prior art keywords
- logo
- detection
- picture
- model
- candidate
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 97
- 238000000605 extraction Methods 0.000 claims abstract description 46
- 238000012549 training Methods 0.000 claims abstract description 34
- 238000000034 method Methods 0.000 claims abstract description 18
- 239000013598 vector Substances 0.000 claims description 39
- 238000005516 engineering process Methods 0.000 claims description 8
- 238000013145 classification model Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 4
- 230000009286 beneficial effect Effects 0.000 description 4
- 238000010276 construction Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/55—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06N3/08—Learning methods
Abstract
The invention discloses a logo detection method based on target detection and metric learning, which is characterized by comprising the following steps of: the method specifically comprises the following steps: step S1, constructing and training a logo detection model; step S2, constructing and training a logo feature extraction model; and step S3, detecting whether the candidate logo target exists in the picture to be detected, and determining whether the candidate logo target is a logo category in the logo retrieval picture library. The logo position in commodity poster data can be efficiently detected, the candidate area of the logo is determined, newly added brands can be identified, and redundant detection is achieved. The model does not need to be retrained, and the extraction accuracy of the brand logo features can be improved more effectively. The complexity of the system is greatly simplified, meanwhile, the recall rate is improved through logo detection, the accuracy rate is improved through logo feature extraction and retrieval, and compared with a single target detection method, the identification effect is better.
Description
Technical Field
The invention relates to the field of computer vision, in particular to a logo detection method based on target detection and metric learning.
Background
With the development of online shopping, the infringement problem of commodity pictures in web pages also becomes more serious. For a large number of commodity pictures, if manual review is performed, a large amount of manpower and material resources are consumed, so that automatic logo infringement detection becomes very important.
In the prior art, for a logo with infringement, when the number of categories is small, basic requirements can be met through target detection, but when the category of the infringement logo is continuously expanded, retraining the model every time becomes tedious, and time is wasted, so that the extensibility of the logo detection model becomes important.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide a logo detection method which is good in expansibility, efficient and accurate. Whether a logo exists in the picture to be detected is identified by using a detection technology, then logo picture characteristics are extracted through a characteristic extraction model based on metric learning, similar logo pictures are recalled by using a retrieval technology, and the category of the logo is judged through voting.
In order to solve the above technical problem, the present invention provides a logo detection method based on target detection and metric learning, which is characterized in that a candidate logo target in a picture is detected by using a target detection technology, the candidate logo target is subjected to feature extraction by using a metric learning technology, and finally the candidate logo is distinguished through logo retrieval to determine the type thereof, specifically comprising the following steps:
step S1, constructing and training a logo detection model, specifically comprising:
step S1-1, constructing a Logo detection data set, wherein commodity pictures in the Logo detection data set are taken from a Logo database, and the Logo database is a Logo Det-3K;
s1-2, constructing and training a logo detection model, wherein the logo detection model is a two-classification model developed based on yolov5 target detection algorithm and used for detecting whether a picture contains a logo or not;
step S2, constructing and training a logo feature extraction model, specifically comprising:
step S2-3, constructing a logo classification data set;
step S2-4, constructing and training a logo feature extraction model, wherein the logo feature extraction model is constructed based on metric learning, a logo classification data set is input for training, and after the logo feature extraction model is trained, the features of the last convolutional layer are taken out to serve as picture features;
step S2-5, constructing a logo retrieval picture library for candidate logo retrieval judgment;
step S2-6, extracting the feature vector of the logo picture in the logo retrieval picture library, inputting the logo picture into a logo feature extraction model, outputting and storing the corresponding feature vector for retrieval;
step S3, detecting whether the candidate logo target exists in the picture to be detected, if so, further extracting logo picture characteristics, comparing the logo picture characteristics with a logo retrieval picture library, and determining whether the logo is the logo category in the logo retrieval picture library;
step S3-7, detecting the picture to be detected by using the trained logo detection model to obtain the position of the candidate logo, and intercepting the candidate logo from the original picture;
step S3-8, scaling the picture resolution of the intercepted candidate logo to 256 × 256, inputting a logo feature extraction model to perform feature extraction to obtain feature vectors, calculating cosine similarity between the feature vectors of the candidate logo and the feature vectors stored in the logo search picture library in the step S2-6, and returning 10 nearest samples according to cosine distance, wherein the cosine distance is defined as follows:
a, B are feature vectors of vectors A and B, respectively, and the category of the logo picture is determined through sample voting.
Step S1-2 further includes that when the commodity pictures in the logo detection data set are sent to the logo detection model for training, the brand types are all changed into single types, the labels are set to be the logos, the specific brand of the logo is not distinguished at the moment, meanwhile, the type confidence coefficient parameter of the logo detection model is set to be 0.2, at the moment, a large number of redundant logo detection targets exist, the high recall rate of a logo detection part can be guaranteed, and the loss situation of the logo is reduced as much as possible.
And constructing a logo classification data set in the step S2-3, selecting 500 brands, intercepting 200 logo pictures of each brand, and resetting the size of each logo picture to 256 × 256.
In step S2-5, a logo feature extraction model is constructed based on an efficientnet algorithm, for the feature of the last convolutional layer, the feature is converted into a 1792 dimensional feature by avgpoling, the training loss is based on an ArcFace algorithm, and the mathematical expression thereof is as follows:
in step S3-8, the sample voting rule is: and the cosine distance returns the nearest 10 samples, if the same type belongs to the types, the number of the same type is not less than 7, and the maximum similarity value is greater than 0.6, the type of the logo to be detected is determined as the voting type, otherwise, the logo to be detected has no specific type and is directly discarded.
In step S2-5, the hyper-parameter of the ArcFace algorithm is configured as: the weight s is 30, the margin is 0.5, and the initial learning rate of model training is 1 e-4.
The invention achieves the following beneficial effects:
1. the method can efficiently detect the logo position in the commodity poster data, determine the candidate area of the logo, and realize identification of newly added brands through mixed training of multiple brands of logos without distinguishing specific brands, thereby effectively improving the model recall rate and realizing redundant detection.
2. The feature extraction model constructed based on metric learning is beneficial to extraction of the feature vectors of the newly added brand logo, retraining of the model is not needed, and the accuracy rate of extraction of the brand logo features can be improved more effectively based on the metric learning method.
3. Compared with the traditional method for directly detecting and identifying the logo based on target detection, the method can increase different brands of logos at will without retraining models, greatly simplifies the complexity of the system, improves the recall rate through logo detection, improves the accuracy rate through logo feature extraction and retrieval, and has better identification effect compared with the method for detecting a single target.
Drawings
FIG. 1 is a schematic flow diagram of a method of an exemplary embodiment of the present invention;
fig. 2 is a schematic diagram of a model structure of an exemplary embodiment of the present invention.
Detailed Description
The invention provides a logo detection method based on target detection and metric learning, which is characterized in that a target detection technology is utilized to position candidate logos in a commodity picture, and then a metric learning technology is utilized to retrieve a detection result to determine the category of the logo, and comprises the following steps:
according to the construction of Logo detection data, because the types and types of the logos are more, the workload of labeling the detection data is larger, and a large amount of manpower is consumed if the data is crawled from the head to label the Logo detection data, the Logo detection data can be constructed by using some public Logo data, such as Logo-3k, Logo-2k and the like.
Training of a logo detection model, wherein the logo detection model is mainly developed based on yolov5, and because a candidate region of a logo is located, the specific category of the logo can be disregarded. Based on the above reasons, when the Logo detection model is trained, only one model of two classifications is trained, and only the foreground and the background in the picture are considered. Therefore, the detection part only needs to ensure the recall height of the model, and all the candidate logo areas can be detected.
And performing model reasoning on the commodity poster picture by using the trained model, and positioning the candidate logo area.
And 2, constructing and training a logo feature extraction model.
And (3) constructing logo classification data, wherein the data can be constructed based on the detection data in the step 1, and because the logo area box is selected in the construction process of the detection data, the classification data set can be constructed by only taking the corresponding logo data from the marked document.
The logo feature extraction model is constructed and trained, feature extraction can be achieved in a mode of representation learning and measurement learning, due to the fact that differences among the logos are not large, details are emphasized, the model trained by the classification method is poor in logo picture extraction effect, and therefore measurement learning is adopted. The metric learning can fully excavate subtle differences among logo pictures, amplify the differences and realize the separation of different logo features on a vector space. The process of model training is the same as the normal process of classification model training, except that the classification loss function is replaced by the measurement loss function, and the cross entropy in the model is replaced by ArcFace. And after the training of the feature extraction model is finished, taking out the last feature layer of the model as the feature of the picture. The ArcFace over-parameter configuration is as follows: the weight s is 30, the margin is 0.5, and the initial learning rate of model training is 1 e-4.
The construction of logo picture library, this part is mainly used as the recalled picture library, and the construction of the picture library data requires higher quality. Because the features extracted from the high-quality picture can represent the central vector of the category, the retrieval result can be more accurate. The pictures are different according to categories, and at least 200 pictures are selected from each category.
And extracting logo characteristic vectors, namely inputting pictures in a logo picture library into a logo characteristic extraction model, outputting corresponding characteristic vectors and storing the characteristic vectors. The saved feature vectors are used for retrieval of subsequent models.
And 3, retrieving and recalling the logo picture.
And (3) firstly carrying out first-step target detection on a new poster picture, positioning a candidate region of the logo, then inputting the scratched candidate region into a logo feature extraction model for feature extraction, calculating cosine distances between the extracted features and the feature vectors stored in the step (2) to determine similarity, returning 10 nearest samples according to the distances, and finally determining the category through sample voting.
In step 2, the problem of class imbalance exists in the disclosed data set, and the training difficulty is increased if the feature extraction model is trained directly based on the data set. The model is used for extracting features, and the problem of long tail is not needed to be solved by spending time, so the class with less data amount is removed in the model training,
in the step 2, the model is mainly constructed based on an efficientnet model, the last convolution layer of the efficient model is converted into a 1792 dimensional feature through Avgpoling, other structures are kept unchanged, the training loss is based on ArcFace, and the mathematical expression is as follows:
and 3, extracting features of the candidate region positioned in the step 1 through the model in the step 2, and calculating cosine similarity with the logo feature vector in the step 2 to obtain the final logo category. Wherein the cosine distance is defined as follows:
where A, B are the feature vectors of vectors a and B, respectively.
The method comprises the steps of constructing and training a logo detection model based on YOLOv5 and used for detecting whether candidate logos exist in pictures, constructing a feature vector extraction model based on efficentnetb 4 and used for logo picture feature vector extraction, constructing a recall model based on picture similarity, obtaining optimal practice parameters by adjusting rules, recalling top10 similar logos each time, and finally determining logo classes according to the number of the same classes not less than 7 and the maximum similarity more than 0.6.
The invention will be further described with reference to the drawings and the exemplary embodiments:
as shown in fig. 1, a logo detection algorithm based on object detection and metric learning of this example mainly includes the following steps:
step S1: and (4) collecting and labeling logo detection data, wherein sufficient data are collected to train a target detection model.
Step S2: training of logo detection models, wherein the detection models are mainly trained based on yolov5, and the detection models are mainly used for detecting candidate areas in poster data.
Step S3: and (4) positioning the logo of the poster data by using a trained detection algorithm, and determining the position of the logo in the poster.
Step S4: and constructing logo classification data, wherein the part is mainly based on logo detection data, and deducting the detected labeling frame to obtain logo classification data.
Step S5: and (3) constructing and training a logo feature extraction model, wherein the logo feature extraction model is mainly constructed based on metric learning.
Step S6: and performing feature extraction on logo data in the galery library to generate a feature vector so as to facilitate subsequent logo retrieval.
Step S7: and searching the detected candidate region to determine the category of the last logo.
In step S1, a detection model may be trained based on public data sets of logo3k and logo2k, or data may be crawled to perform labeling to construct a detection data set.
In step S2, the inspection model is mainly trained based on yolov5, and may be based on other inspection algorithms such as yolov4 and SSD. The single-stage model is mainly used for realizing high detection speed and high efficiency.
In step S3, the trained detection model is used to perform detection and positioning on the test data to obtain a logo region candidate for the picture data.
In step S4, classification data for metric learning is constructed, and this classification data is constructed mainly based on the above-described detection data.
In step S5, the feature extraction model is mainly constructed by using a metric learning algorithm, and the model training is mainly performed by using the loss of arcfacce.
In step S6, feature extraction is performed on the data in the galery library to obtain a final feature vector, and this step is mainly to perform one-time forward propagation on the data in the library to obtain a feature matrix of the logo.
Fig. 2 is a schematic block structure diagram of a logo detection method based on target detection and metric learning according to the present invention.
The module 1 is a candidate logo detection module and is used for detecting whether a logo possibly exists in a picture, realizing redundant detection by reducing confidence of a candidate frame, sampling a plurality of candidate logo areas and outputting the candidate logo as a coordinate of the candidate logo in an original picture.
The module 2 is a logo vector generation module and is used for generating a corresponding logo retrieval picture vector library for a logo retrieval picture, generating 1792-bit feature vectors for each picture, and for candidate logo areas, intercepting the candidate logos in an original drawing through coordinates of the candidate logos in the original drawing, zooming to 256 x 256, and generating 1792-dimensional feature vectors. The module outputs 1792-dimensional floating-point feature vectors for picture conversion.
The module 3 is a logo retrieval and identification module and is used for calculating cosine similarity between candidate logo feature vectors and logo retrieval picture vectors, each candidate logo feature vector can obtain 10 retrieval library picture vectors with the nearest cosine distance so as to obtain 10 nearest retrieval picture targets, and specific categories of the candidate logos are determined according to the rule that the number of the same categories is not less than 7 and the maximum similarity is greater than 0.6.
The invention achieves the following beneficial effects:
1. the method can efficiently detect the logo position in the commodity poster data, determine the candidate area of the logo, and realize identification of newly added brands through mixed training of multiple brands of logos without distinguishing specific brands, thereby effectively improving the model recall rate and realizing redundant detection.
2. The feature extraction model constructed based on metric learning is beneficial to extraction of the feature vectors of the newly added brand logo, retraining of the model is not needed, and the accuracy rate of extraction of the brand logo features can be improved more effectively based on the metric learning method.
3. Compared with the traditional method for directly detecting and identifying the logo based on target detection, the method can increase different brands of logos at will without retraining models, greatly simplifies the complexity of the system, improves the recall rate through logo detection, improves the accuracy rate through logo feature extraction and retrieval, and has better identification effect compared with the method for detecting a single target.
The above embodiments do not limit the present invention in any way, and all other modifications and applications that can be made to the above embodiments in equivalent ways are within the scope of the present invention.
Claims (6)
1. A logo detection method based on target detection and metric learning is characterized in that a target detection technology is used for detecting candidate logo targets in a picture, a metric learning technology is used for carrying out feature extraction on the candidate logo targets, and finally, the candidate logo is distinguished through logo retrieval to determine the type of the candidate logo targets, and the method specifically comprises the following steps:
step S1, constructing and training a logo detection model, specifically comprising:
step S1-1, constructing a Logo detection data set, wherein commodity pictures in the Logo detection data set are taken from a Logo database, and the Logo database is a Logo Det-3K;
s1-2, constructing and training a logo detection model, wherein the logo detection model is a two-classification model developed based on yolov5 target detection algorithm and used for detecting whether a picture contains a logo or not;
step S2, constructing and training a logo feature extraction model, specifically comprising:
step S2-3, constructing a logo classification data set;
step S2-4, constructing and training a logo feature extraction model, wherein the logo feature extraction model is constructed based on metric learning, a logo classification data set is input for training, and after the logo feature extraction model is trained, the features of the last convolutional layer are taken out to serve as picture features;
step S2-5, constructing a logo retrieval picture library for candidate logo retrieval judgment;
step S2-6, extracting the feature vector of the logo picture in the logo retrieval picture library, inputting the logo picture into a logo feature extraction model, outputting and storing the corresponding feature vector for retrieval;
step S3, detecting whether the candidate logo target exists in the picture to be detected, if so, further extracting logo picture characteristics, comparing the logo picture characteristics with a logo retrieval picture library, and determining whether the logo is the logo category in the logo retrieval picture library;
step S3-7, detecting the picture to be detected by using the trained logo detection model to obtain the position of the candidate logo, and intercepting the candidate logo from the original picture;
step S3-8, scaling the picture resolution of the intercepted candidate logo to 256 × 256, inputting a logo feature extraction model to perform feature extraction to obtain feature vectors, calculating cosine similarity between the feature vectors of the candidate logo and the feature vectors stored in the logo search picture library in the step S2-6, and returning 10 nearest samples according to cosine distance, wherein the cosine distance is defined as follows:
a, B are feature vectors of vectors A and B, respectively, and the category of the logo picture is determined through sample voting.
2. The logo detection method based on object detection and metric learning as claimed in claim 1, wherein: and step S1-2, when the commodity pictures in the logo detection data set are sent to the logo detection model for training, the brand categories are all changed into a single category, the labels are set to be the logos, and the category confidence coefficient parameter of the logo detection model is set to be 0.2.
3. The logo detection method based on object detection and metric learning as claimed in claim 2, wherein: and constructing a logo classification data set in the step S2-3, selecting 500 brands, intercepting 200 logo pictures of each brand, and resetting the size of each logo picture to 256 × 256.
4. The logo detection method based on object detection and metric learning as claimed in claim 3, wherein: in step S2-5, a logo feature extraction model is constructed based on an efficientnet algorithm, for the feature of the last convolutional layer, the feature is converted into a 1792 dimensional feature by avgpoling, the training loss is based on an ArcFace algorithm, and the mathematical expression thereof is as follows:
5. the logo detection method based on object detection and metric learning as claimed in claim 4, wherein: in step S3-8, the sample voting rule is: and the cosine distance returns the nearest 10 samples, if the same type belongs to the type, the number is not less than 7, and the maximum similarity value is greater than 0.6, the type of the logo to be detected is determined as the voting type, otherwise, the logo to be detected has no specific type.
6. The logo detection method based on object detection and metric learning as claimed in claim 5, wherein: in step S2-5, the hyper-parameter of the ArcFace algorithm is configured as: the weight s is 30, the margin is 0.5, and the initial learning rate of model training is 1 e-4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111090918.6A CN113780284B (en) | 2021-09-17 | Logo detection method based on target detection and metric learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111090918.6A CN113780284B (en) | 2021-09-17 | Logo detection method based on target detection and metric learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113780284A true CN113780284A (en) | 2021-12-10 |
| CN113780284B CN113780284B (en) | 2024-04-19 |
Family
ID=
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117351242A (en) * | 2023-10-24 | 2024-01-05 | 北京云上曲率科技有限公司 | Identification recognition method and system based on similarity |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108647682A (en) * | 2018-05-17 | 2018-10-12 | 电子科技大学 | A kind of brand Logo detections and recognition methods based on region convolutional neural networks model |
| CN108764313A (en) * | 2018-05-17 | 2018-11-06 | 西安电子科技大学 | Supermarket's commodity recognition method based on deep learning |
| CN109961089A (en) * | 2019-02-26 | 2019-07-02 | 中山大学 | Small sample and zero sample image classification method based on metric learning and meta learning |
| CN112115906A (en) * | 2020-09-25 | 2020-12-22 | 广州市派客朴食信息科技有限责任公司 | Open dish identification method based on deep learning target detection and metric learning |
| CN112668445A (en) * | 2020-12-24 | 2021-04-16 | 南京泓图人工智能技术研究院有限公司 | Vegetable type detection and identification method based on yolov5 |
| CN113076927A (en) * | 2021-04-25 | 2021-07-06 | 华南理工大学 | Finger vein identification method and system based on multi-source domain migration |
| WO2021136027A1 (en) * | 2019-12-30 | 2021-07-08 | Oppo广东移动通信有限公司 | Similar image detection method and apparatus, device and storage medium |
| CN113269127A (en) * | 2021-06-10 | 2021-08-17 | 北京睿芯高通量科技有限公司 | Face recognition and pedestrian re-recognition monitoring method and system for real-time automatic database building |
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108647682A (en) * | 2018-05-17 | 2018-10-12 | 电子科技大学 | A kind of brand Logo detections and recognition methods based on region convolutional neural networks model |
| CN108764313A (en) * | 2018-05-17 | 2018-11-06 | 西安电子科技大学 | Supermarket's commodity recognition method based on deep learning |
| CN109961089A (en) * | 2019-02-26 | 2019-07-02 | 中山大学 | Small sample and zero sample image classification method based on metric learning and meta learning |
| WO2021136027A1 (en) * | 2019-12-30 | 2021-07-08 | Oppo广东移动通信有限公司 | Similar image detection method and apparatus, device and storage medium |
| CN112115906A (en) * | 2020-09-25 | 2020-12-22 | 广州市派客朴食信息科技有限责任公司 | Open dish identification method based on deep learning target detection and metric learning |
| CN112668445A (en) * | 2020-12-24 | 2021-04-16 | 南京泓图人工智能技术研究院有限公司 | Vegetable type detection and identification method based on yolov5 |
| CN113076927A (en) * | 2021-04-25 | 2021-07-06 | 华南理工大学 | Finger vein identification method and system based on multi-source domain migration |
| CN113269127A (en) * | 2021-06-10 | 2021-08-17 | 北京睿芯高通量科技有限公司 | Face recognition and pedestrian re-recognition monitoring method and system for real-time automatic database building |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117351242A (en) * | 2023-10-24 | 2024-01-05 | 北京云上曲率科技有限公司 | Identification recognition method and system based on similarity |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108985334B (en) | General object detection system and method for improving active learning based on self-supervision process | |
| Pham et al. | The random cluster model for robust geometric fitting | |
| Dong et al. | Tablesense: Spreadsheet table detection with convolutional neural networks | |
| JP6005837B2 (en) | Image analysis apparatus, image analysis system, and image analysis method | |
| CN109165540B (en) | Pedestrian searching method and device based on prior candidate box selection strategy | |
| CN102385592B (en) | Image concept detection method and device | |
| JP2001256244A (en) | Device and method for sorting image data | |
| CN111444939A (en) | Small-scale equipment component detection method based on weak supervision cooperative learning in open scene of power field | |
| CN110728307A (en) | Method for realizing small sample character recognition of X-ray image by self-generating data set and label | |
| CN112115291B (en) | Three-dimensional indoor model retrieval method based on deep learning | |
| CN114694178A (en) | Method and system for monitoring safety helmet in power operation based on fast-RCNN algorithm | |
| CN115147380A (en) | Small transparent plastic product defect detection method based on YOLOv5 | |
| CN115601307A (en) | Automatic cell detection method | |
| CN109614512B (en) | Deep learning-based power equipment retrieval method | |
| CN115344733A (en) | Digital collection retrieval method and system based on deep hash | |
| CN113032613B (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network | |
| CN110083724A (en) | A kind of method for retrieving similar images, apparatus and system | |
| WO2024021321A1 (en) | Model generation method and apparatus, electronic device, and storage medium | |
| WO2023273572A1 (en) | Feature extraction model construction method and target detection method, and device therefor | |
| Zihao et al. | Printed Surface Defect Detection Model Based on Positive Samples. | |
| CN113780284B (en) | Logo detection method based on target detection and metric learning | |
| CN113780284A (en) | Logo detection method based on target detection and metric learning | |
| JP2004171316A (en) | Ocr device, document retrieval system and document retrieval program | |
| CN104484869A (en) | Image matching method and system for ordinal measure features | |
| CN114972880A (en) | Label identification method and device, electronic equipment and storage medium |
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 |