CN115861997A - License plate detection and identification method for guiding knowledge distillation by key foreground features - Google Patents
License plate detection and identification method for guiding knowledge distillation by key foreground features Download PDFInfo
- Publication number
- CN115861997A CN115861997A CN202310167743.7A CN202310167743A CN115861997A CN 115861997 A CN115861997 A CN 115861997A CN 202310167743 A CN202310167743 A CN 202310167743A CN 115861997 A CN115861997 A CN 115861997A
- Authority
- CN
- China
- Prior art keywords
- license plate
- network
- model
- training
- convolution
- 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
Images
Classifications
-
- 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
Abstract
The invention belongs to the technical field of license plate detection and identification, and relates to a license plate detection and identification method for guiding knowledge distillation by key foreground characteristics.
Description
Technical Field
The invention belongs to the technical field of license plate detection and identification, and particularly relates to a license plate detection and identification method for guiding knowledge distillation by key foreground features.
Background
With the development of economy, the number of automobiles is increased sharply, and a lot of challenges are brought to urban traffic management. The license plate is used as the identification of the vehicle identity, the unified management of the vehicle can be realized by detecting and identifying the license plate, and the problem of vehicle management can be relieved to a great extent. With the increase of the number of different types of license plates in different scenes in actual use scenes, in order to obtain better detection capability, a larger network is generally used for learning the characteristics of the license plates, but the deployment of the license plates on edge equipment is limited due to the need for stronger computing capability and slower reasoning speed.
In order to overcome the above problems, the existing methods generally use knowledge distillation, which is a method for inheriting information in a large teacher network into a compact student network and obtaining strong performance without adding extra cost in the reasoning process, so that higher accuracy can be obtained. However, most distillation methods are designed for image classification, which results in a slight improvement in target detection, because the foreground and background features are viewed equally during distillation, and knowledge of the background area, which results in a number of absolute dominant background areas, will result in far better distillation than the foreground features. Existing methods improve upon the above problem primarily from two aspects, one assigning weights to suppress background and the other proposing distilling students' regions of interest, however these methods do not make it clear which features in the target box are key features.
Therefore, an effective method is still lacked in the license plate detection based on knowledge distillation to solve the problem of poor distillation effect caused by unbalanced foreground and background, so that an effective means is urgently needed to improve the license plate detection and recognition performance based on knowledge distillation so as to improve the forward reasoning speed.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a license plate detection and identification method with key foreground characteristics guiding knowledge distillation, which is used for solving the problem of poor distillation effect caused by unbalanced foreground and background in the license plate detection and identification method based on knowledge distillation, can be used for license plate detection and identification tasks in any scene, and can efficiently realize the precision and the model generalization performance of license plate detection and identification.
In order to achieve the above object, the present invention designs a teacher model and a student model, wherein the teacher model uses a large backbone network, the student model uses a light backbone network, trains the teacher network, freezes the teacher network after the teacher network converges, and enables the teacher network to guide the learning of the student network, and simultaneously learns spatial attention in the backbone networks of the teacher model and the student model, and encourages the spatial attention of the student model to imitate the spatial attention map of the teacher model, and also encourages each element in the learning map to represent the importance of the foreground, thereby alleviating the influence of the imbalance of the foreground and the background, emphasizing the learning of the student model to the foreground object, suppressing the learning of the student model to the background pixel, and enabling the student model to know which pixel information should be learned, and improving the learning efficiency and precision of the model, specifically comprising the following steps:
(1) Collecting license plate images under different scenes, constructing a license plate data set, and dividing the license plate data set into a training set, a verification set and a test set;
(2) Constructing a teacher model with ResNeXt101 as a backbone network and training the teacher model;
(3) Constructing a student model taking ResNet18 as a backbone network, and initializing network parameters of the student model by using random initialization;
(4) Simultaneously learning attention maps of characteristics of a third stage, a fourth stage and a fifth stage in a teacher model and student model backbone network, so that the attention maps of the student model and the teacher model are similar as much as possible;
(5) Training a student model to obtain trained license plate detection and recognition network parameters;
(6) The network is tested using the student model as a network for forward reasoning and the license plate number and location are output.
As a further technical scheme of the invention, the license plate data set in the step (1) needs to be marked with a license plate position and a license plate character type, wherein the license plate position comprises four corner points of a license plate, the length, the width and the center point of a horizontal frame of the license plate position are calculated and represented through the positions of the four corner points, and the license plate character type comprises provinces, english letters, numbers and special characters.
As a further technical scheme of the invention, the specific process of the step (2) is as follows:
(21) Taking the license plate images in the training set in the step (1) as input, initializing the size and the numerical range of the license plate images, inputting the processed license plate images into a backbone network for convolution feature extraction, wherein the backbone network uses ResNeXt101 pre-trained in ImageNet data set as a feature extraction network, and adds a feature pyramid network behind ResNeXt101 to enhance multi-scale features in ResNeXt101 to obtain a multi-scale license plate convolution feature map set, so that different sizes of license plates in a short distance and a long distance are captured better, and the extracted shared features are applied to image enhancement, license plate detection and license plate recognition;
(22) After a multi-scale license plate convolution feature map set is obtained, two convolution layers with convolution kernels of 3 x 3 and one average pooling layer are connected after different multi-scale convolution layers, each feature point of the feature map is provided with an anchor frame, two full-connection layer sub-networks with the same structure but without shared parameters are used for learning classification and position deviation information respectively, wherein the position deviation information is the deviation of a minimum external rectangular frame of a license plate and an anchor frame corresponding to the feature point, so that the task of classification and position regression of a target frame is completed, whether the classification is the license plate is judged, and the position information is the minimum external moment (facing to the frame) of the license plate;
(23) Acquiring convolution characteristics corresponding to license plate positions from multi-scale license plate convolution characteristics, using two cascaded convolution layers after the convolution characteristics, further extracting the license plate characteristics through a full connection layer (FC), then positioning character positions in a license plate, classifying character categories by using a Softmax function, wherein the character categories mainly comprise provinces, letters, numbers and special characters (military, police, school and the like), and finally outputting the recognized characters according to a position relation to obtain a final license plate number;
(24) And (3) training the teacher model by using the training set data in the step (1), wherein after 50 times of iterative training, the teacher network converges and the loss tends to be stable, and the final teacher model is stored for guiding the learning of the student models.
As a further technical scheme of the invention, the specific process of the step (5) is as follows: the method comprises the steps that training of a student model is supervised by a real label and a teacher model, wherein the real label comprises the position of a license plate and a real license plate number, the supervision of the teacher model is supervised by a space attention of learned foreground importance, the space attention of the student model is enabled to be similar as much as possible, an L2 function is used as a loss function, license plate detection is achieved by using smooth L1 loss to restrict a regression boundary frame output by the teacher model and a regression boundary frame output by the student model, character positioning and character classification in license plate recognition are achieved by using smooth L1 loss and cross entropy loss respectively to restrict errors between predicted values and true values, parameters are updated through back propagation, after 50 rounds of complete training set training iteration, model parameters with the best results in a verification set are stored and serve as parameters trained by a final model, and accordingly trained license plate detection recognition network parameters are obtained.
As a further technical scheme of the invention, in the testing process, the long edge of the image is zoomed (resize) to 512 under the condition that the proportion of the long edge and the short edge of the image is kept unchanged, then the short edge of the image is filled to ensure that the size of the image is 512 multiplied by 512, the image is used as the input of a network, namely the classification confidence coefficient of the license plate and the coordinate position of the license plate can be output, the license plate with low confidence coefficient is filtered by setting a threshold value, a redundant frame output by the network is deleted by using non-maximum suppression (NMS), and finally the license plate characteristics corresponding to the license plate detection frame are input to a license plate recognition module to obtain the license plate number.
Compared with the prior art, the invention has the beneficial effects that:
the method overcomes the problem of unbalance of foreground and background when knowledge is extracted by target detection based on knowledge distillation, each element in the attention is represented by importance probability of foreground by learning the space attention of the characteristics of a main network, then the space attention is learned simultaneously in the main networks of a teacher model and a student model, and the space attention of the student model is encouraged to imitate the space attention of the teacher model, the method emphasizes the learning of the student model to foreground objects, inhibits the learning of the student model to background pixels, enables the student model to know the information of which pixels should be learned, improves the efficiency and the precision of model learning, can be used for license plate detection and identification tasks and can also be used for various target detection tasks with large-scale data, and compared with the existing YOLOv3 method, the precision of the method in CCPD test data can be improved from 94.2% to 95.5%, and the reasoning speed is improved by 1.5 times.
Drawings
FIG. 1 is a schematic diagram of a network architecture framework for implementing license plate detection and recognition according to the present invention.
FIG. 2 is a block diagram of a process for detecting and recognizing a license plate according to the present invention.
Detailed Description
The invention will be further described by way of examples, without in any way limiting the scope of the invention, with reference to the accompanying drawings.
Example (b):
in the embodiment, the network shown in fig. 1 and the flow shown in fig. 2 are adopted to realize license plate detection and recognition of knowledge distillation guided by key foreground features, a teacher model and a student model are designed, the teacher model and the student model are mainly different in structure in that a large backbone network is used in the teacher model, a light backbone network is used in the student model, then the teacher network is trained, the teacher network is frozen after the teacher network converges, and the teacher network guides the student network to learn, specifically, spatial attention is simultaneously learned in the backbone networks of the teacher model and the student model, the spatial attention of the student model is encouraged to imitate spatial attention of the teacher model, and each element in the attention is expressed as importance of foreground, so that the influence of imbalance of foreground and background is relieved. The method emphasizes the learning of the student model on foreground objects, inhibits the learning of the student model on background pixels, enables the student model to know which pixels should be learned, improves the efficiency and the precision of model learning, and comprises the following specific implementation steps:
(1) Data set construction
Collecting license plate images under different scenes, constructing a data set of a license plate, dividing the data set into a training set, a verification set and a test set, and marking the position of the license plate and the character type of the license plate, wherein the position of the license plate mainly refers to the marking of four angular points of the license plate, the length, the width and the central point of a horizontal frame representing the position of the license plate can be calculated through the positions of the four angular points, and the character type of the license plate mainly comprises provinces, english letters, numbers, special characters and the like;
(2) Construction and training of teacher models
(21) Taking the license plate images in the training set in the step (1) as input, initializing the size and the numerical range of the license plate images, inputting the processed license plate images into a backbone network for convolution feature extraction, wherein the backbone network uses ResNeXt101 pre-trained in ImageNet data set as a feature extraction network, and adds a feature pyramid network behind ResNeXt101 to enhance multi-scale features in ResNeXt101 to obtain a multi-scale license plate convolution feature map set, so that different sizes of license plates in a short distance and a long distance are captured better, and the extracted shared features are applied to image enhancement, license plate detection and license plate recognition;
(22) After a multi-scale license plate convolution feature map set is obtained, two convolution layers with convolution kernels of 3 x 3 and one average pooling layer are connected after different multi-scale convolution layers, each feature point of the feature map is provided with an anchor frame, two full-connection layer sub-networks with the same structure but without shared parameters are used for learning classification and position deviation information respectively, wherein the position deviation information is the deviation of a minimum external rectangular frame of a license plate and an anchor frame corresponding to the feature point, so that the task of classification and position regression of a target frame is completed, whether the classification is the license plate is judged, and the position information is the minimum external moment (facing to the frame) of the license plate;
(23) Acquiring convolution characteristics corresponding to license plate positions from multi-scale license plate convolution characteristics, using two cascaded convolution layers after the convolution characteristics, further extracting the license plate characteristics through a full connection layer (FC), then positioning character positions in a license plate, classifying character categories by using a Softmax function, wherein the character categories mainly comprise provinces, letters, numbers and special characters (military, police, school and the like), and finally outputting the recognized characters according to a position relation to obtain a final license plate number;
(24) Training the teacher model by using the training set data in the step (1), wherein after 50 times of iterative training, the teacher network converges and the loss tends to be stable, and the final teacher model is stored for guiding the learning of the student models;
(3) Construction of student models
The network structure of the student model is basically consistent with that of the teacher model, and the main difference is that the lightweight main network ResNet18 is used, has fewer parameters and faster reasoning speed, and initializes the network parameters of the student model by random initialization after the student model is constructed;
(4) Key prospect feature guided knowledge distillation
Firstly, the characteristics of the third stage C3, the fourth stage C4 and the fifth stage C5 in the backbone network are used as input
It is shown that N is 3, 4, 5, corresponding to the characteristics of the third, fourth and fifth stages C3, C4, C5, respectively
Then, the corresponding space attention is learned by using different convolution layers and is used for ^ er>
Indicates that the attention maps outputted are ^ and ^ respectively>
Each element in the attention drawings means the importance of the foreground, attention drawings of characteristics of a third stage C3, a fourth stage C4 and a fifth stage C5 in a teacher model and student model backbone network are learned at the same time, and the attention drawings of the student model attention drawings are similar to the attention drawings of the teacher model as much as possible, wherein C is the number of channels, and H and W respectively represent the height and the width of an image; the teacher model and the student models are represented by the same symbol because the characteristic dimensions of the teacher model and the student models in the third stage C3, the fourth stage C4 and the fifth stage C5 are the same;
(5) Student model training
The training of the student model is supervised by a real label and a teacher model, wherein the real label comprises the position of the license plate
And a real number plate, wherein>
The coordinates of four corner points of a license plate are taken as coordinates, the supervision of a teacher model mainly takes the learned space attention of the foreground importance as supervision to ensure that the space attention of a student model is similar as much as possible, an L2 function is taken as a loss function, the license plate detection uses smooth L1 loss to restrict a regression boundary box output by the teacher model and the student model, the character positioning and character classification in the license plate recognition respectively use smoothL1 loss and cross entropy loss to restrict the error between a predicted value and a true value, parameters are updated through back propagation, after 50 rounds of complete training set training iteration, model parameters with the best result on a verification set are stored and taken as parameters trained by a final model, and the trained license plate detection recognition network parameters are obtained,
(6) Testing network output license plate number and position
In the testing process, only the student model is used as a network for forward reasoning, the image long side is scaled (resize) to 512 under the condition that the proportion of the image long side and the image short side is kept unchanged, and then the short side of the image is filled to enable the image size to be 512 x 512, and the image size is used as the input of the network. The classification confidence of the license plate and the coordinate position of the license plate can be output, the threshold value is set to filter the license plate with low confidence, a redundant frame output by a network is deleted by using non-maximum suppression (NMS), and finally the license plate characteristics corresponding to the license plate detection frame are input to a license plate recognition module to obtain the license plate number.
It is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various substitutions and modifications are possible without departing from the spirit and scope of the invention and appended claims. Therefore, the invention should not be limited by the disclosure of the embodiments, but should be defined by the scope of the appended claims.
Claims (5)
1. A license plate detection method for guiding knowledge distillation by key foreground features is characterized by comprising the following steps:
(1) Collecting license plate images under different scenes, constructing a license plate data set, and dividing the license plate data set into a training set, a verification set and a test set;
(2) Constructing a teacher model with ResNeXt101 as a backbone network and training the teacher model;
(3) Constructing a student model taking ResNet18 as a backbone network, and initializing network parameters of the student model by using random initialization;
(4) Simultaneously learning attention graphs of characteristics of a third stage, a fourth stage and a fifth stage in a teacher model and a student model backbone network;
(5) Training a student model to obtain trained license plate detection and recognition network parameters;
(6) The network is tested using the student model as a network for forward reasoning and the license plate number and location are output.
2. The license plate detection method for guiding knowledge distillation according to the key foreground features of claim 1, wherein in the step (1), the license plate data set is marked with a license plate position and a license plate character type, the license plate position comprises four corner points of the license plate, the length, the width and the center point of a horizontal frame representing the license plate position are calculated according to the positions of the four corner points, and the license plate character type comprises provinces, english letters, numbers and special characters.
3. The license plate detection method for guiding knowledge distillation according to the key foreground features of claim 2, wherein the specific process of the step (2) is as follows:
(21) Taking the license plate image in the training set in the step (1) as input, initializing the size and the numerical range of the license plate image, inputting the processed license plate image into a backbone network for convolution feature extraction, wherein the backbone network uses ResNeXt101 pre-trained in an ImageNet data set as a feature extraction network, and adds a feature pyramid network behind the ResNeXt101 to enhance multi-scale features in the ResNeXt101 to obtain a multi-scale license plate convolution feature map set;
(22) After a multi-scale license plate convolution feature map set is obtained, two convolution layers with convolution kernels of 3 x 3 and one average pooling layer are connected after different multi-scale convolution layers, each feature point of the feature map is provided with an anchor frame, two full-connection layer sub-networks with the same structure but without shared parameters are used for learning classification and position deviation information respectively, wherein the position deviation information is the deviation of a minimum external rectangular frame of a license plate and an anchor frame corresponding to the feature point, so that the task of classification and position regression of a target frame is completed, whether the classification is the license plate or not is judged, and the position information is the minimum external moment of the license plate;
(23) Acquiring convolution characteristics corresponding to license plate positions from multi-scale license plate convolution characteristics, using two cascaded convolution layers after the convolution characteristics, further extracting the license plate characteristics through a full connection layer, then positioning character positions in a license plate, classifying character categories by using a Softmax function, wherein the character categories mainly comprise provinces, letters, numbers and special characters, and finally outputting the recognized characters according to a position relation to obtain a final license plate number;
(24) And (3) training the teacher model by using the training set data in the step (1), wherein after 50 times of iterative training, the teacher network converges and the loss tends to be stable, and the final teacher model is stored for guiding the learning of the student models.
4. The license plate detection method for guiding knowledge distillation according to the key foreground features of claim 3, wherein the specific process of the step (5) is as follows: the method comprises the steps that training of a student model is supervised by a real label and a teacher model, wherein the real label comprises the position of a license plate and a real license plate number, the supervision of the teacher model is supervised by a space attention of learned foreground importance, an L2 function is used as a loss function, smooth L1 loss is used for license plate detection to restrict a regression boundary frame output by the teacher model and a regression boundary frame output by the student model, smooth L1 loss and cross entropy loss are used for restricting errors between a predicted value and a true value respectively in character positioning and character classification in license plate recognition, parameters are updated through back propagation, model parameters with the best results in a verification set are stored after 50 times of training iteration of a complete training set, the model parameters are used as parameters of final model training, and the trained license plate detection recognition network parameters are obtained.
5. The license plate detection method for guiding knowledge distillation according to the key foreground features of claim 4, wherein in the test process, the long sides of the image are scaled to 512 under the condition that the proportion of the long sides and the short sides of the image is kept unchanged, then the short sides of the image are filled to make the size of the image 512 x 512, the image is used as the input of a network, namely the classification confidence coefficient of the license plate and the coordinate position of the license plate can be output, the license plate with low confidence coefficient is filtered by setting a threshold value, a redundant frame output by the network is deleted by using non-maximum inhibition, and finally the license plate features corresponding to the license plate detection frame are input to a license plate recognition module to obtain the license plate number.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310167743.7A CN115861997B (en) | 2023-02-27 | 2023-02-27 | License plate detection and recognition method for key foreground feature guided knowledge distillation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310167743.7A CN115861997B (en) | 2023-02-27 | 2023-02-27 | License plate detection and recognition method for key foreground feature guided knowledge distillation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115861997A true CN115861997A (en) | 2023-03-28 |
| CN115861997B CN115861997B (en) | 2023-05-16 |
Family
ID=85659045
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310167743.7A Active CN115861997B (en) | 2023-02-27 | 2023-02-27 | License plate detection and recognition method for key foreground feature guided knowledge distillation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115861997B (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190095730A1 (en) * | 2017-09-25 | 2019-03-28 | Beijing University Of Posts And Telecommunications | End-To-End Lightweight Method And Apparatus For License Plate Recognition |
| CN112164054A (en) * | 2020-09-30 | 2021-01-01 | 交叉信息核心技术研究院(西安)有限公司 | Knowledge distillation-based image target detection method and detector and training method thereof |
| CN112308066A (en) * | 2020-10-23 | 2021-02-02 | 西安科锐盛创新科技有限公司 | License plate recognition system |
| CN114283402A (en) * | 2021-11-24 | 2022-04-05 | 西北工业大学 | License plate detection method based on knowledge distillation training and space-time combined attention |
| US20220180199A1 (en) * | 2019-09-04 | 2022-06-09 | Huawei Technologies Co., Ltd. | Neural network model compression method and apparatus, storage medium, and chip |
| CN114882490A (en) * | 2022-07-08 | 2022-08-09 | 松立控股集团股份有限公司 | Unlimited scene license plate detection and classification method based on point-guided positioning |
| CN114898352A (en) * | 2022-06-29 | 2022-08-12 | 松立控股集团股份有限公司 | Method for simultaneously realizing image defogging and license plate detection |
-
2023
- 2023-02-27 CN CN202310167743.7A patent/CN115861997B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190095730A1 (en) * | 2017-09-25 | 2019-03-28 | Beijing University Of Posts And Telecommunications | End-To-End Lightweight Method And Apparatus For License Plate Recognition |
| US20220180199A1 (en) * | 2019-09-04 | 2022-06-09 | Huawei Technologies Co., Ltd. | Neural network model compression method and apparatus, storage medium, and chip |
| CN112164054A (en) * | 2020-09-30 | 2021-01-01 | 交叉信息核心技术研究院(西安)有限公司 | Knowledge distillation-based image target detection method and detector and training method thereof |
| CN112308066A (en) * | 2020-10-23 | 2021-02-02 | 西安科锐盛创新科技有限公司 | License plate recognition system |
| CN114283402A (en) * | 2021-11-24 | 2022-04-05 | 西北工业大学 | License plate detection method based on knowledge distillation training and space-time combined attention |
| CN114898352A (en) * | 2022-06-29 | 2022-08-12 | 松立控股集团股份有限公司 | Method for simultaneously realizing image defogging and license plate detection |
| CN114882490A (en) * | 2022-07-08 | 2022-08-09 | 松立控股集团股份有限公司 | Unlimited scene license plate detection and classification method based on point-guided positioning |
Non-Patent Citations (2)
| Title |
|---|
| CONG ZHANG ET AL.: "V-LPDR: Towards a unified framework for license plate detection, tracking, and recognition in real-world traffic videos", NEUROCOMPUTING * |
| 侯佳英: "基于深度学习的图像生成算法研究与实现", 中国优秀硕士学位论文全文数据库信息科技辑 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115861997B (en) | 2023-05-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109902806B (en) | Method for determining target bounding box of noise image based on convolutional neural network | |
| CN108647655B (en) | Low-altitude aerial image power line foreign matter detection method based on light convolutional neural network | |
| CN110569738B (en) | Natural scene text detection method, equipment and medium based on densely connected network | |
| CN108537824B (en) | Feature map enhanced network structure optimization method based on alternating deconvolution and convolution | |
| CN109492596B (en) | Pedestrian detection method and system based on K-means clustering and regional recommendation network | |
| CN110853057B (en) | Aerial image segmentation method based on global and multi-scale full-convolution network | |
| CN112801158A (en) | Deep learning small target detection method and device based on cascade fusion and attention mechanism | |
| CN111612051A (en) | Weak supervision target detection method based on graph convolution neural network | |
| CN111160205A (en) | Embedded multi-class target end-to-end unified detection method for traffic scene | |
| CN110991444A (en) | Complex scene-oriented license plate recognition method and device | |
| CN112990065A (en) | Optimized YOLOv5 model-based vehicle classification detection method | |
| CN112990282B (en) | Classification method and device for fine-granularity small sample images | |
| CN115063786A (en) | High-order distant view fuzzy license plate detection method | |
| CN113850136A (en) | Yolov5 and BCNN-based vehicle orientation identification method and system | |
| CN110852327A (en) | Image processing method, image processing device, electronic equipment and storage medium | |
| CN115019279A (en) | Context feature fusion method based on MobileNet lightweight network | |
| CN114519819B (en) | Remote sensing image target detection method based on global context awareness | |
| CN116524189A (en) | High-resolution remote sensing image semantic segmentation method based on coding and decoding indexing edge characterization | |
| CN115861756A (en) | Earth background small target identification method based on cascade combination network | |
| CN111160282B (en) | Traffic light detection method based on binary Yolov3 network | |
| CN116229406B (en) | Lane line detection method, system, electronic equipment and storage medium | |
| CN111079634B (en) | Method, device and system for detecting obstacle in running process of vehicle and vehicle | |
| CN112288702A (en) | Road image detection method based on Internet of vehicles | |
| CN110334703B (en) | Ship detection and identification method in day and night image | |
| CN112364864A (en) | License plate recognition 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 | ||
| GR01 | Patent grant |