CN110956218A - Method for generating target detection football candidate points of Nao robot based on Heatmap - Google Patents
Method for generating target detection football candidate points of Nao robot based on Heatmap Download PDFInfo
- Publication number
- CN110956218A CN110956218A CN201911257253.6A CN201911257253A CN110956218A CN 110956218 A CN110956218 A CN 110956218A CN 201911257253 A CN201911257253 A CN 201911257253A CN 110956218 A CN110956218 A CN 110956218A
- Authority
- CN
- China
- Prior art keywords
- heatmap
- target detection
- candidate points
- football
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 47
- 238000001514 detection method Methods 0.000 title claims abstract description 40
- 238000012549 training Methods 0.000 claims abstract description 42
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 24
- 238000012800 visualization Methods 0.000 claims abstract description 9
- 238000012360 testing method Methods 0.000 claims abstract description 7
- 238000004364 calculation method Methods 0.000 claims abstract description 3
- 230000008569 process Effects 0.000 claims description 14
- 238000004422 calculation algorithm Methods 0.000 claims description 9
- 238000013135 deep learning Methods 0.000 claims description 9
- 238000002474 experimental method Methods 0.000 claims description 7
- 238000004088 simulation Methods 0.000 claims description 5
- 230000002238 attenuated effect Effects 0.000 claims description 3
- 238000004519 manufacturing process Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 2
- 238000000605 extraction Methods 0.000 abstract description 7
- 238000012545 processing Methods 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 10
- 230000001133 acceleration Effects 0.000 description 4
- 238000005286 illumination Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000013031 physical testing Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a method for generating target detection football candidate points of a Nao robot based on Heatmap, which comprises the steps of selecting a convolutional neural network as a target detection model, simulating a competition environment, collecting a plurality of groups of data sets for picture making training and testing, generating the Heatmap, processing to obtain a visualization result of the Heatmap, reconstructing the convolutional neural network to accelerate the network calculation speed, setting a proper threshold value, and finally sending the candidate points into a classifier to obtain a final accurate identification result, wherein the points which are larger than the set threshold value in the Heatmap are candidate points of a ball. The method enhances the adaptability of the Nao robot vision system to the light environment of the match field, can realize high-precision identification of the football under different light environments, utilizes fewer convolutional layers to complete feature extraction, ensures the real-time performance of identification, and greatly improves the accuracy of football identification by generating football candidate points and then entering a classifier for identification.
Description
Technical Field
The invention relates to a method for generating target detection football candidate points of a Nao robot based on Heatmap.
Background
The Nao robot is a humanoid robot with powerful function and extremely high research value, and the application of the humanoid robot is spread throughout the global education and research institutions, while the RoboCup robot world cup game is a robot game which aims at promoting the development of artificial intelligence and robotics, wherein the RoboCup standard platform group is specially set as a group for the Nao robot to carry out humanoid football games. In the match, the recognition of the Nao robot on the football involves the decision and final win and loss of the match, and the importance of the Nao robot is self-evident and is a key research topic of the visual direction of the Nao robot. Limited by the computing power of the robot hardware equipment, the current solutions are:
1. generating ball candidate points by adopting a traditional method, classifying to obtain the positions of the balls, selecting pixel points which meet the condition that large non-green gaps exist among green pixels of vertical scanning lines and the non-green gaps extend in the horizontal direction greatly as candidate points when the candidate points of the balls are generated, widening the candidate conditions along with the distance, sending the obtained regions of the candidate points into a classifier, and finally determining the position information of the balls.
2. The method is characterized in that detection and identification are directly carried out on the football through a deep learning target detection method, a target detector is built through a convolutional neural network, and single target detection is finally realized through training.
However, under the interference of other factors such as variable illumination environment and field robot, the above method has many defects:
1. the traditional method has the characteristics of small calculated amount, high real-time performance and simple realization principle, but has low robustness, is greatly influenced by the change of an external illumination environment, is easy to cause misjudgment to generate a plurality of wrong candidate points, thereby increasing the calculated amount of the subsequent classifier identification and ensuring that the accuracy of the whole football identification process is low.
2. The method for directly detecting the deeply learned target is limited by real-time performance, the accuracy is sacrificed for pursuing the real-time performance, the accuracy cannot be very good, and a large number of false identifications are also generated.
Disclosure of Invention
The invention aims to provide a method for generating target detection football candidate points of a Nao robot based on Heatmap, which is characterized in that feature extraction is carried out on images collected by the Nao robot by utilizing a convolutional neural network to determine candidate points of a ball in a mode of generating the Heatmap, the Heatmap represents a probability distribution diagram of the detected target ball generated after feature extraction of the convolutional neural network, the probability of a certain point is higher, the probability of the point is higher, namely the candidate points are determined in a mode of generating the Heatmap, and then the Nao robot is classified and identified to enhance the adaptability of the Nao robot to the environment, so that the method can adapt to sudden illumination transformation and accurately identify the ball, and has high accuracy.
In order to achieve the purpose, the invention adopts the technical scheme that:
a method for generating Nao robot target detection football candidate points based on Heatmap comprises the following steps:
s1: selecting a target detection model and a training frame, wherein a convolutional neural network is taken as the target detection model, building a plurality of convolutional layers and a target detection model of an output layer for training,
s2: simulating a competition environment, collecting a plurality of groups of picture making training and testing data sets,
s3: performing training and comparison experiments for multiple times to determine final weight parameters,
s4: generating a Heatmap, standardizing the Hetmap to obtain a visualization result of the Heatmap,
s5: reconstructing the convolutional neural network to accelerate the network calculation speed, setting a proper threshold value, wherein the point in the Heatmap, which is larger than the set threshold value, is the candidate point of the ball, and finally sending the candidate point to a classifier to obtain a final accurate identification result.
Preferably, the generating method further includes S6: and storing the trained model parameters, integrating the reconstructed target detection model into a vision system, performing simulation and entity test experiments, and identifying the accuracy of the candidate points of the ball.
Preferably, in S1: and (3) adopting a YoloV3 detection algorithm, and taking DarkNet as a deep learning training framework. In a RoboCup game, the Nao robot has simple site information, needs to constantly position the accurate position of a ball, actually needs to determine the position of the central point of the ball, and then uses an edge detector to fit the size of the ball to accurately position the ball, so a YoloV3 detection algorithm is adopted to build a convolutional neural network, feature extraction is carried out according to a picture obtained from the NAO robot, the position of the ball is accurately positioned, and a network output layer is directly replaced by a Heatmap containing target position information to complete a design task; by adopting DarkNet as a deep learning framework, the problem of limited hardware computing capability of the Nao robot can be solved.
Further preferably, when the DarkNet deep learning framework performs convolution operation, a matrix multiplication mode of im2col and gemm is adopted, and the im2col comprises a large number of for cycles, so that the overall operation time of the network is greatly increased.
Preferably, in S2: and marking by using labelImg software to manufacture the VOC data set.
Preferably, in S2: and acquiring pictures through a camera of the Nao robot.
Preferably, in S3: setting hyper-parameters, including: the method comprises the steps of batch _ size, subdivisions, decay, duration, exposure, hue, learning _ rate, max _ batches and steps, wherein in the training process, pictures of the batch _ size number are read into a buffer area each time, and are divided into subdivisions times of training to be finished, the whole training process needs to read the pictures of the max _ batches times of the batch _ size number and finish the training, and meanwhile, when the training iteration number reaches the value of the steps in the training process, the learning _ rate is attenuated by ten times compared with the previous value.
Preferably, in S4: and generating a Heatmap by normalizing the output characteristic graph of the last convolution layer of the convolutional neural network.
Preferably, in S4: when the Heatmap visualization results were obtained, the processed data were normalized to 0-255 gray scale values.
Preferably, in S5: the SSE instruction set is used for reconstructing the convolutional neural network to accelerate the network computing speed, the SSE instruction set is a technology specially aiming at CPU operation acceleration, the capability of CPU floating point operation can be effectively enhanced, and 4 32-bit floating point numbers can be stored and operated in parallel by utilizing a 128-bit register, so that the whole computing process is accelerated. Rewriting the network by using set, load, store, arithmetic instruction and the like of an SSE instruction set and deploying the network into the Nao robot vision system to realize the acceleration and final deployment of the network.
The method adopts a mode of generating the Heatmap by the deep learning convolutional neural network to obtain candidate points when the Nao robot identifies the football, and sends the candidate points to the classifier to determine the final identification result, thereby realizing the accurate identification of the Nao robot in the RoboCup football match. Firstly, the convolutional neural network has strong capability in the visual field from AlexNet 2012 by depending on the strong feature extraction capability of the convolutional neural network, can realize the extraction and utilization of local and global features of an image through a convolutional layer, thereby accurately classifying the image, and realizes the synchronous classification and positioning of the convolutional neural network along with the efforts of more people; YoloV3 is an outstanding representative of the target detection algorithm of end-to-end training, and the invention also carries out research work on the basis of the YoloV 3. Secondly, although the convolutional neural network target detection algorithm has a good identification and positioning effect, a high hardware platform is configured mostly based on a GPU and the like, the high real-time and high-accuracy effect of the convolutional neural network target detection algorithm at a mobile end with limited hardware resources is still a big difficulty, a balance point is selected between the accuracy and the real-time property to be a normal state, and although some algorithms of a compression acceleration network appear, the requirement of the Nao robot in the football match is still difficult to meet in the accuracy.
Due to the application of the technical scheme, compared with the prior art, the invention has the following advantages and effects:
1. the adaptability of the Nao robot vision system to the light environment of the competition field is enhanced, and the high-precision identification of the ball can be realized under different light environments.
2. The feature extraction is completed by using fewer convolutional layers, the real-time performance of identification is guaranteed, and meanwhile, the accuracy of football identification is greatly improved by a method of generating football candidate points and then entering a classifier for identification.
Drawings
FIG. 1 is a process diagram of the production method of this example;
FIG. 2 is a schematic diagram of the role of Heatmap in the Nao robot football recognition system;
FIG. 3 is a schematic diagram of a convolutional neural network structure;
FIG. 4 is a schematic diagram of the structure of a VOC data set;
FIG. 5a is a diagram illustrating the visualization result of the original image and different layers of Heatmap;
FIG. 5b is a schematic representation of the visualization of the final Heatmap;
FIG. 6 is a flow diagram of an SSE command acceleration network;
FIG. 7a is a diagram illustrating simulation results;
FIG. 7b is a schematic diagram of the results of the physical testing.
Detailed Description
The invention is further described below with reference to the accompanying drawings and embodiments:
as shown in fig. 1 and 2, a method for generating a target detection football candidate point based on a Nao robot of Heatmap includes: s1: a design target and the hardware level of an experiment carrier Nao robot are determined, a convolutional neural network is determined to be a target detection model based on a deep learning method, and a deep learning training frame DarkNet is selected according to a YoloV3 detection algorithm (target detection algorithm). And constructing a target detection model with six convolutional layers and one output layer for training, wherein the network structure is shown in FIG. 3.
S2: simulating a competition environment, obtaining an entity photo from a camera of the Nao robot, collecting a large number of pictures, and labeling and sorting by using LabelImg software to obtain a VOC data set (target detection data set) shown in figure 4 for training.
S3: the super parameters of the network are set as shown in the following table, in the training process, pictures of the quantity of batch _ size are read into a buffer area every time, the pictures are divided into subdivisions for training completion, the whole training process needs to read the pictures of the quantity of batch _ size for max _ batches and is terminated only after the training is completed, in order to avoid that the learning _ rate is too large to cause the result to exceed the optimal value when the training is finished, and to ensure that the learning rate at the initial training stage is not too small to cause the speed to be too slow, therefore, the steps value is set, so that the learning rate is ten times attenuated compared with the previous value when the number of training iterations reaches the values of the steps in the training process. Other hyper-parameters include: the decay weight attenuation coefficient is used to prevent overfitting; saturation, exposure, hue, etc. are used to generate more training samples. After the setting is finished, a network structure is set up to carry out a plurality of training comparison experiments, and the weight parameters of the final structure and the network are determined.
| Hyper-parameter | Value taking |
| batch _ size (number of samples for one training) | 64 |
| subdivisions | 16 |
| Decay (weight attenuation coefficient) | 0.0005 |
| Saturation (Saturation) | 1.5 |
| Exposure (Exposure) | 1.5 |
| Hue (tone) | 0.1 |
| learning _ rate (learning rate) | 0.001 |
| max _ batches (maximum training sample number) | 500200 |
| Steps (training iteration times) | 400000，500000 |
S4: when a training result is tested, normalizing the output characteristic diagram of the last convolutional layer of the convolutional neural network to obtain a Heatmap, wherein the main equation form is as follows:
in order to more intuitively represent the Heatmap, the processed data is further normalized to 0-255 gray values to obtain the visualization result, as shown in fig. 5a and 5b, the main equation form of the visualization process is:
s5: reconstructing the convolutional neural network by using the SSE instruction set to accelerate the network computation speed, and setting a proper threshold, wherein the point in the Heatmap larger than the set threshold is a candidate point of the ball, and finally sending the candidate point to a classifier to obtain a final accurate recognition result, wherein a flow chart of the SSE instruction set reconstructed network is shown in FIG. 6. The whole recognition process meets the real-time requirement, can adapt to the change of the external illumination environment, and improves the accuracy of ball recognition during Nao robot competition.
S6: storing the trained model parameters, integrating the model rewritten by the SSE into the whole visual system, performing simulation and entity test experiments, and identifying the accuracy of the ball through the simulation and the multi-frame test network of the entity, so as to improve the real-time performance and accuracy of the method and ensure the requirement is met.
The above embodiments are merely illustrative of the technical ideas and features of the present invention, and the purpose thereof is to enable those skilled in the art to understand the contents of the present invention and implement the present invention, and not to limit the protection scope of the present invention. All equivalent changes and modifications made according to the spirit of the present invention should be covered within the protection scope of the present invention.
Claims (10)
1. A method for generating target detection football candidate points of a Nao robot based on Heatmap is characterized by comprising the following steps: the method comprises the following steps:
s1: selecting a target detection model and a training frame, wherein a convolutional neural network is taken as the target detection model, building a plurality of convolutional layers and a target detection model of an output layer for training,
s2: simulating a competition environment, collecting a plurality of groups of picture making training and testing data sets,
s3: performing training and comparison experiments for multiple times to determine final weight parameters,
s4: generating a Heatmap, standardizing the Hetmap to obtain a visualization result of the Heatmap,
s5: reconstructing the convolutional neural network to accelerate the network calculation speed, setting a proper threshold value, wherein the point in the Heatmap, which is larger than the set threshold value, is the candidate point of the ball, and finally sending the candidate point to a classifier to obtain a final accurate identification result.
2. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 1, wherein: the generation method further includes S6: and storing the trained model parameters, integrating the reconstructed target detection model into a vision system, performing simulation and entity test experiments, and identifying the accuracy of the candidate points of the ball.
3. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 1, wherein: in S1: and (3) adopting a YoloV3 detection algorithm, and taking DarkNet as a deep learning training framework.
4. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 3, wherein: and adopting a matrix multiplication mode of im2col and gemm when the DarkNet deep learning framework carries out convolution operation.
5. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 1, wherein: in S2: and marking by using labelImg software to manufacture the VOC data set.
6. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 1, wherein: in S2: and acquiring pictures through a camera of the Nao robot.
7. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 1, wherein: in S3: setting hyper-parameters, including: the method comprises the steps of batch _ size, subdivisions, decay, duration, exposure, hue, learning _ rate, max _ batches and steps, wherein in the training process, pictures of the batch _ size number are read into a buffer area each time, and are divided into subdivisions times of training to be finished, the whole training process needs to read the pictures of the max _ batches times of the batch _ size number and finish the training, and meanwhile, when the training iteration number reaches the value of the steps in the training process, the learning _ rate is attenuated by ten times compared with the previous value.
8. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 1, wherein: in S4: and generating a Heatmap by normalizing the output characteristic graph of the last convolution layer of the convolutional neural network.
9. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 1, wherein: in S4: when the Heatmap visualization results were obtained, the processed data were normalized to 0-255 gray scale values.
10. The generation method of target detection football candidate points based on the Nao robot of Heatmap according to claim 1, wherein: in S5: and reconstructing the convolutional neural network by adopting the SSE instruction set to accelerate the network computing speed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911257253.6A CN110956218A (en) | 2019-12-10 | 2019-12-10 | Method for generating target detection football candidate points of Nao robot based on Heatmap |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911257253.6A CN110956218A (en) | 2019-12-10 | 2019-12-10 | Method for generating target detection football candidate points of Nao robot based on Heatmap |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110956218A true CN110956218A (en) | 2020-04-03 |
Family
ID=69980630
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911257253.6A Pending CN110956218A (en) | 2019-12-10 | 2019-12-10 | Method for generating target detection football candidate points of Nao robot based on Heatmap |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110956218A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112749626A (en) * | 2020-12-10 | 2021-05-04 | 同济大学 | DSP platform-oriented rapid face detection and recognition method |
| CN113222805A (en) * | 2021-05-08 | 2021-08-06 | 西北工业大学 | Rapid high-accuracy NAO type football robot vision processing method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107009357A (en) * | 2017-04-05 | 2017-08-04 | 武汉科技大学 | A kind of method that object is captured based on NAO robots |
| CN108549876A (en) * | 2018-04-20 | 2018-09-18 | 重庆邮电大学 | The sitting posture detecting method estimated based on target detection and human body attitude |
| CN109523013A (en) * | 2018-10-15 | 2019-03-26 | 西北大学 | A kind of air particle pollution level estimation method based on shallow-layer convolutional neural networks |
| CN109917818A (en) * | 2019-01-31 | 2019-06-21 | 天津大学 | Collaboratively searching based on ground robot contains method |
| CN110491480A (en) * | 2019-05-22 | 2019-11-22 | 腾讯科技（深圳）有限公司 | A kind of medical image processing method, device, electromedical equipment and storage medium |
-
2019
- 2019-12-10 CN CN201911257253.6A patent/CN110956218A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107009357A (en) * | 2017-04-05 | 2017-08-04 | 武汉科技大学 | A kind of method that object is captured based on NAO robots |
| CN108549876A (en) * | 2018-04-20 | 2018-09-18 | 重庆邮电大学 | The sitting posture detecting method estimated based on target detection and human body attitude |
| CN109523013A (en) * | 2018-10-15 | 2019-03-26 | 西北大学 | A kind of air particle pollution level estimation method based on shallow-layer convolutional neural networks |
| CN109917818A (en) * | 2019-01-31 | 2019-06-21 | 天津大学 | Collaboratively searching based on ground robot contains method |
| CN110491480A (en) * | 2019-05-22 | 2019-11-22 | 腾讯科技（深圳）有限公司 | A kind of medical image processing method, device, electromedical equipment and storage medium |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112749626A (en) * | 2020-12-10 | 2021-05-04 | 同济大学 | DSP platform-oriented rapid face detection and recognition method |
| CN112749626B (en) * | 2020-12-10 | 2022-09-13 | 同济大学 | DSP platform-oriented rapid face detection and recognition method |
| CN113222805A (en) * | 2021-05-08 | 2021-08-06 | 西北工业大学 | Rapid high-accuracy NAO type football robot vision processing method |
| CN113222805B (en) * | 2021-05-08 | 2023-04-07 | 西北工业大学 | Rapid high-accuracy NAO type football robot vision processing method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111126472B (en) | SSD (solid State disk) -based improved target detection method | |
| CN109543606B (en) | Human face recognition method with attention mechanism | |
| CN111767882B (en) | Multi-mode pedestrian detection method based on improved YOLO model | |
| CN111062282B (en) | Substation pointer instrument identification method based on improved YOLOV3 model | |
| CN110136154A (en) | Remote sensing images semantic segmentation method based on full convolutional network and Morphological scale-space | |
| CN109509187B (en) | Efficient inspection algorithm for small defects in large-resolution cloth images | |
| CN106504233A (en) | Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN | |
| CN111833237B (en) | Image registration method based on convolutional neural network and local homography transformation | |
| CN113128335B (en) | Method, system and application for detecting, classifying and finding micro-living ancient fossil image | |
| CN110781882A (en) | License plate positioning and identifying method based on YOLO model | |
| CN114549507B (en) | Improved Scaled-YOLOv fabric flaw detection method | |
| CN111310609B (en) | Video target detection method based on time sequence information and local feature similarity | |
| CN112528845A (en) | Physical circuit diagram identification method based on deep learning and application thereof | |
| CN113160136B (en) | Wood defect identification and segmentation method based on improved Mask R-CNN | |
| CN110956218A (en) | Method for generating target detection football candidate points of Nao robot based on Heatmap | |
| CN112364974B (en) | YOLOv3 algorithm based on activation function improvement | |
| CN110533629A (en) | A kind of detection method and detection device of Bridge Crack | |
| CN109543617A (en) | The detection method of intelligent vehicle movement traffic information based on YOLO target detection technique | |
| CN112329771A (en) | Building material sample identification method based on deep learning | |
| CN113657423A (en) | Target detection method suitable for small-volume parts and stacked parts and application thereof | |
| CN107644203A (en) | A kind of feature point detecting method of form adaptive classification | |
| Peng et al. | Application of deep residual neural network to water meter reading recognition | |
| CN111738099B (en) | Face automatic detection method based on video image scene understanding | |
| Zhao et al. | Object detector based on enhanced multi-scale feature fusion pyramid network | |
| CN109766768A (en) | A kind of sportsman's recognition methods based on deep learning |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200403 |