CN117635924A - Low-energy-consumption target detection method based on adaptive knowledge distillation - Google Patents
Low-energy-consumption target detection method based on adaptive knowledge distillation Download PDFInfo
- Publication number
- CN117635924A CN117635924A CN202410104431.6A CN202410104431A CN117635924A CN 117635924 A CN117635924 A CN 117635924A CN 202410104431 A CN202410104431 A CN 202410104431A CN 117635924 A CN117635924 A CN 117635924A
- Authority
- CN
- China
- Prior art keywords
- training
- feature map
- sampling rate
- edge server
- image feature
- 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 86
- 238000005265 energy consumption Methods 0.000 title claims abstract description 32
- 238000013140 knowledge distillation Methods 0.000 title claims abstract description 24
- 230000003044 adaptive effect Effects 0.000 title claims description 19
- 238000012549 training Methods 0.000 claims abstract description 73
- 238000005070 sampling Methods 0.000 claims abstract description 46
- 238000000034 method Methods 0.000 claims abstract description 35
- 239000000284 extract Substances 0.000 claims abstract description 5
- 238000004364 calculation method Methods 0.000 claims description 27
- 230000005540 biological transmission Effects 0.000 claims description 7
- 238000010586 diagram Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 abstract description 7
- 230000009467 reduction Effects 0.000 abstract description 3
- 238000005516 engineering process Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 3
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 238000013139 quantization Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000008094 contradictory effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
- 239000013585 weight reducing agent Substances 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a low-energy-consumption target detection method based on self-adaptive knowledge distillation, and aims to solve the problem that resource-limited equipment is easily affected by data drift in the target detection process. Firstly, the terminal equipment carrying the lightweight target detection algorithm periodically uploads the image, the detection result and the residual electric quantity information which are partially detected to the edge server carrying the high-precision target detection model. Then, the edge server detects the uploaded image, calculates the relative loss value of precision, the average image similarity and the energy consumption proportion, finally obtains the self-adaptive sampling rate and the training trigger instruction, and transmits the self-adaptive sampling rate and the training trigger instruction back to the terminal equipment. And finally, the terminal equipment extracts part of images from the experience pool, forms a training set with the received data, performs model training and updates the sampling rate and the experience pool. The invention can solve the problem of detection precision reduction caused by the influence of data drift on the resource-constrained equipment, improve the detection precision and realize continuous learning with low energy consumption.
Description
Technical Field
The invention belongs to the technical field of video target detection in computer vision, and particularly relates to a low-energy-consumption target detection method based on self-adaptive knowledge distillation under a terminal-side collaborative framework, which realizes a low-energy-consumption continuous learning technology of resource-limited equipment.
Background
Thanks to the rapid development of the deep learning technology, the accuracy of the target detection model is significantly improved. The improvement of the accuracy of the object detection model is derived from more accurate extraction of image features and higher-level semantic understanding of the deep learning model. However, as the complexity of the model increases, the model parameters also exhibit a rapidly increasing trend. Large-scale deep networks may involve hundreds of millions or even billions of parameters, which makes deployment of models on resource-constrained devices extremely difficult. The resource-constrained devices, such as embedded systems, mobile devices or edge computing devices, have limited computing, storage capabilities and power, and cannot effectively support large-scale model parameters. This results in that in a resource-limited environment, the object detection model cannot achieve efficient real-time reasoning, or even run on the device.
To address this challenge, researchers have proposed a series of approaches to model weight reduction and optimization. These methods include techniques such as network pruning, parameter quantization (parameter quantization), model compression (model compression), etc., in order to reduce the number of parameters of the model. In addition, there are some specially designed lightweight object detection models, such as MobileNet, efficientDet, which reduce the parameters and computational complexity of the model while maintaining high accuracy through well-designed structures and strategies. While real-time reasoning can be implemented on resource constrained devices by a method that reduces the number of parameters, the reduction in the number of parameters also means that the model is more sensitive to data drift (data drift). When there is a discrepancy between the deployed scene and the data set used in model training, a dramatic decrease in the accuracy of target detection may result.
Knowledge distillation is expected to solve this problem. As one of the continuous learning techniques, a strategy of knowledge distillation is to solve the problem of data drift sensitivity caused by the reduction of the parameter amount by transferring the knowledge of a complex teacher model to a light-weight student model. The basic idea of knowledge distillation is to take the prediction result of the teacher model as an auxiliary target of the student model and learn by minimizing the distance between the prediction distribution of the teacher model and the student model. However, for the terminal device with limited resources, it is not practical to frequently upload and download the model parameters, which causes high time delay and energy consumption.
Therefore, the invention discloses a low-energy-consumption target detection method based on adaptive knowledge distillation. The knowledge distillation technology under the cooperation of the end edges is designed, and the energy consumption of the terminal equipment is reduced by adaptively adjusting the sampling rate and the training trigger instruction.
Disclosure of Invention
The invention aims to solve the problems of the prior art, and provides a low-energy-consumption target detection method based on adaptive knowledge distillation, which solves the problem that resource-limited equipment is easily influenced by data drift in the target detection process by decoupling a training and labeling part of the knowledge distillation and adaptively adjusting a sampling rate and a training trigger instruction, and reduces the energy consumption in the knowledge distillation process.
The technical solution for realizing the purpose of the invention is as follows: a low energy consumption target detection method based on adaptive knowledge distillation, the method comprising the steps of:
step 1: the terminal equipment carrying the lightweight target detection algorithm selects partial images from the detected images according to the sampling rate, and stores the characteristic images into a buffer pool; meanwhile, uploading the image feature map, the detection result and the current battery energy remaining proportion information in the buffer pool to an edge server carrying a high-precision and high-generalization capacity target detection model;
step 2: the edge server carries out high-precision detection on the uploaded image feature images, calculates a precision relative loss value, average image similarity and energy consumption proportion based on the detection result and the uploaded information, and obtains an adaptive sampling rate and a training trigger instruction; then, the edge server selects to transmit the target detection result, the self-adaptive sampling rate and the training trigger instruction data to the terminal equipment according to the calculation result;
step 3: the terminal equipment randomly extracts the image feature images with the same number from the experience pool, and forms a training set with the downloaded image feature images and the labels thereof, and is used for training the carried model and updating the sampling rate; after training is completed, the experience pool is updated, and the extracted image feature map is replaced by the downloaded image feature map.
Further, step 1: the terminal equipment carrying the lightweight target detection algorithm selects partial image feature images from the detected images according to the sampling rate and stores the partial image feature images into a buffer pool; meanwhile, the image feature map, the detection result and the current battery energy remaining proportion information in the buffer pool are uploaded to an edge server carrying a high-precision and high-generalization capacity target detection model, and the method specifically comprises the following steps:
step 1-1: the terminal equipment collects images in real time to detect targets and samples the images according to the sampling rateRandomly sampling from the detected image, and storing the characteristic diagram and the detection result into a buffer pool;
step 1-2: every other time slotThe terminal device will buffer all image feature maps +.>Corresponding target detection results->Composing the feature map set->In total->Stretch and assemble the feature map->Device batteryInformation on the energy left->Uploading to an edge server;
further, step 2: the purpose of self-adaptive adjustment of the sampling rate is to measure the training requirement on the terminal equipment according to the current system state, and when the requirement is lower, the sampling rate is reduced, so that the energy consumption of transmission and training can be reduced; when the demand is higher, increase sampling rate, can increase training sample, improve the precision fast. If the self-adaptive method is not adopted, the sampling rate cannot be reduced when the training requirement is low, so that resource waste is generated, and the sampling rate cannot be quickly improved when the requirement is high, so that the accuracy cannot be quickly improved. And the edge server carries out high-precision detection on the uploaded image feature map, calculates a precision relative loss value, average image similarity and energy consumption proportion based on the detection result and the uploaded information, and obtains an adaptive sampling rate and a training trigger instruction. Then, the edge server selects to transmit the target detection result, the self-adaptive sampling rate and the training trigger instruction data to the terminal equipment according to the calculation result; the method specifically comprises the following steps:
step 2-1: the edge server uploads the image feature mapInputting into high-precision detection network to obtain detection result +.>；
Step 2-2: the detection result is processedAs a real tag, the detection result of the terminal device is calculated +.>mAP value of each image in (2) and then taking the average value as the calculation accuracy relative loss value +.>The method comprises the steps of carrying out a first treatment on the surface of the Computing feature map set->Middle->Average similarity of image feature map>The calculation mode is as follows:
，
wherein,、/>and->Is constant (I)>And->Feature map set->Middle feature map->Mean and variance of output vector obtained after inputting object detection model on edge server, ++>Is in the range of +.>The method comprises the steps of carrying out a first treatment on the surface of the Calculation ofRatio of energy consumption to battery remaining energy required for image transmission and training on end devices +.>The calculation mode is as follows:
，
wherein,bandwidth allocated to the edge server to the terminal device, < >>For the transmission power of the terminal device,for the channel gain +.>Is channel noise>For the data volume of an image, +.>The energy consumption required for locally training an image;
step 2-3: calculating an adaptive sampling rateThe calculation mode is as follows:
，
wherein,、/>and->Parameters (respectively)>、/>And->The accuracy relative loss threshold value, the feature map similarity threshold value and the energy ratio threshold value are respectively +.>And->For the lower and upper bounds of the sampling rate, +.>Indicating that the calculation result is not out of limits; computing training trigger instruction +.>The calculation mode is as follows:
，
step 2-4: according to training trigger instruction, the edge serverThe calculation result judges the data needing to be returned, if +.>Then the high-precision model detection result on the edge server is +.>、/>Time sample rate->Training trigger instruction +.>Issuing to terminal equipment; if->Issuing training trigger instruction +.>Andtime sample rate->；
Further, step 3: the terminal equipment randomly extracts the image feature images with the same number from the experience pool, and forms a training set with the downloaded image feature images and the labels thereof, and is used for training the carried model and updating the sampling rate; after training is completed, updating the experience pool, and replacing the extracted image feature map with the downloaded image feature map; the method specifically comprises the following steps:
step 3-1: the terminal equipment receives the data returned from the edge server; if training trigger instructionThe device update sampling rate is +.>The method comprises the steps of carrying out a first treatment on the surface of the If training trigger instruction->Performing the training operation of the step 3-2;
step 3-2: if training trigger instructionRandomly extracting the number +.>Is associated with the feature map set stored in the last time slot +.>And detection result downloaded from edge server +.>Form training set +.>Then training a target detection model on the terminal equipment; after training is completed, the feature images are gatheredDetection result->Replacing the extracted feature map, storing the feature map in an experience pool, and updating the sampling rate of the device to be +.>The method comprises the steps of carrying out a first treatment on the surface of the After completion, steps 1-1 to 3-2 are repeated.
Compared with the prior art, the invention has the remarkable advantages that: 1) The target detection method is based on a knowledge distillation technology, solves the problem that the resource-limited equipment is easily affected by data drift in the target detection process, and effectively improves the target detection precision; 2) The training and labeling parts in the knowledge distillation frame with the coordinated end edges are decoupled and respectively deployed on the end equipment and the edge server, and the self-adaptive sampling rate and the training trigger instruction are designed, so that the continuous learning process with low energy consumption can be realized.
The invention is described in further detail below with reference to the accompanying drawings.
Drawings
FIG. 1 shows a schematic overall framework of a low energy consumption target detection method based on adaptive knowledge distillation of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the present application.
It should be noted that, if there is a description of "first", "second", etc. in the embodiments of the present invention, the description of "first", "second", etc. is only for descriptive purposes, and is not to be construed as indicating or implying relative importance or implying that the number of technical features indicated is indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include at least one such feature. In addition, the technical solutions of the embodiments may be combined with each other, but it is necessary to base that the technical solutions can be realized by those skilled in the art, and when the technical solutions are contradictory or cannot be realized, the combination of the technical solutions should be considered to be absent and not within the scope of protection claimed in the present invention.
The invention aims to solve the problems of the prior art, and provides a low-energy-consumption target detection method based on adaptive knowledge distillation, which solves the problem that resource-limited equipment is easily influenced by data drift in the target detection process by decoupling a training and labeling part of the knowledge distillation and adaptively adjusting a sampling rate and a training trigger instruction, and reduces the energy consumption in the knowledge distillation process.
FIG. 1 shows a schematic overall framework of a low energy consumption target detection method based on adaptive knowledge distillation of the present invention. Referring to fig. 1, the implementation steps of the method of the present embodiment are described in detail below.
Step 1-1: the terminal equipment is carried with a lightweight target detection model MobileNet, acquires images in real time through a camera to detect targets, and samples the targets according to a sampling rateRandomly sampling from the detected image, and storing the characteristic diagram and the detection result into a buffer pool;
step 1-2: every other time slotThe terminal device will buffer all image feature maps +.>Corresponding target detection results->Composing the feature map set->In total->Stretch and assemble the feature map->And the current remaining energy information of the device battery +.>Uploading to an edge server;
step 2-1: the edge server uploads the image feature mapInputting high-precision detection network Fast-R-CNN to obtain detection result +.>；
Step 2-2: the detection result is processedAs a real tag, the detection result of the terminal device is calculated +.>mAP value of each image in (2) and then taking the average value as the calculation accuracy relative loss value +.>The method comprises the steps of carrying out a first treatment on the surface of the Computing feature map setsMiddle->Average similarity of image feature map>The calculation mode is as follows:
，
wherein,、/>and->Is constant (I)>And->Feature map set->Middle feature map->Mean and variance of output vector obtained after inputting object detection model on edge server, ++>Is in the range of +.>The method comprises the steps of carrying out a first treatment on the surface of the Calculation ofRatio of energy consumption to battery remaining energy required for image transmission and training on end devices +.>The calculation mode is as follows:
，
wherein,bandwidth allocated to the edge server to the terminal device, < >>For the transmission power of the terminal device,for the channel gain +.>Is channel noise>The data amount for one image;
step 2-3: calculating an adaptive sampling rateThe calculation mode is as follows:
，
wherein,、/>and->Parameters (respectively)>、/>And->The accuracy relative loss threshold value, the feature map similarity threshold value and the energy ratio threshold value are respectively +.>And->For the lower and upper bounds of the sampling rate, +.>Indicating that the calculation result is not out of limits; computing training trigger instruction +.>The calculation mode is as follows:
，
step 2-4: according to training trigger instruction, the edge serverThe calculation result judges the data needing to be returned, if +.>Then the high-precision model detection result on the edge server is +.>、/>Time sample rate->Training trigger instruction +.>Issuing to terminal equipment; if->Issuing training trigger instruction +.>Andtime sample rate->。
Step 3-1: the terminal device receives the data transmitted back from the edge server. If training trigger instructionThe device update sampling rate is +.>The method comprises the steps of carrying out a first treatment on the surface of the If training trigger instruction->Performing the training operation of the step 3-2;
step 3-2: if training trigger instructionRandomly extracting the number +.>Is associated with the feature map set stored in the last time slot +.>And detection result downloaded from edge server +.>Form training set +.>And then training the target detection model on the terminal equipment. After training, the feature map is assembled +.>Detection result->Replacing the extracted characteristic diagram and storing the characteristic diagram in an experience pool. Device update sampling rate is +.>. Then enter the next time slot, repeat step 1-1 to step 3-2.
The foregoing has outlined and described the basic principles, features, and advantages of the present invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (4)
1. A low energy consumption target detection method based on adaptive knowledge distillation, characterized in that the method comprises the following steps:
step 1: the terminal equipment carrying the lightweight target detection algorithm selects partial images from the detected images according to the sampling rate, and stores the characteristic images into a buffer pool; meanwhile, uploading the image feature map, the detection result and the current battery energy remaining proportion information in the buffer pool to an edge server carrying a high-precision and high-generalization capacity target detection model;
step 2: the edge server carries out high-precision detection on the uploaded image feature images, calculates a precision relative loss value, average image similarity and energy consumption proportion based on the detection result and the uploaded information, and obtains an adaptive sampling rate and a training trigger instruction; then, the edge server selects to transmit the target detection result, the self-adaptive sampling rate and the training trigger instruction data to the terminal equipment according to the calculation result;
step 3: the terminal equipment randomly extracts the image feature images with the same number from the experience pool, and forms a training set with the downloaded image feature images and the labels thereof, and is used for training the carried model and updating the sampling rate; after training is completed, the experience pool is updated, and the extracted image feature map is replaced by the downloaded image feature map.
2. The method for detecting the low-energy-consumption target based on the adaptive knowledge distillation according to claim 1, wherein in step 1, the terminal device carrying the lightweight target detection algorithm selects a part of image feature images from the detected images according to a sampling rate, and stores the part of image feature images in a buffer pool; meanwhile, the image feature map, the detection result and the current battery energy remaining proportion information in the buffer pool are uploaded to an edge server carrying a high-precision and high-generalization capacity target detection model, and the method specifically comprises the following steps:
step 1-1: the terminal equipment collects images in real time to detect targets and samples the images according to the sampling rateRandomly sampling from the detected image, and storing the characteristic diagram and the detection result into a buffer pool;
step 1-2: every other time slotThe terminal device will buffer all image feature maps +.>Corresponding target detection results->Composing the feature map set->In total->Stretch and assemble the feature map->And the current remaining energy information of the device battery +.>Uploading to an edge server.
3. The method for detecting the low-energy-consumption target based on the adaptive knowledge distillation according to claim 1, wherein the edge server performs high-precision detection on the uploaded image feature map in step 2, calculates a precision relative loss value, average image similarity and energy consumption proportion based on a detection result and uploaded information, and obtains an adaptive sampling rate and a training trigger instruction; then, the edge server selects to transmit the target detection result, the self-adaptive sampling rate and the training trigger instruction data to the terminal equipment according to the calculation result; the method specifically comprises the following steps:
step 2-1: the edge server uploads the image feature mapInputting into high-precision detection network to obtain detection result；
Step 2-2: the detection result is processedAs a real tag, the detection result of the terminal device is calculated +.>mAP value of each image in (2) and then taking the average value as the calculation accuracy relative loss value +.>The method comprises the steps of carrying out a first treatment on the surface of the Computing feature map set->Middle->Average similarity of image feature map>The calculation mode is as follows:
，
wherein,、/>and->Is constant (I)>And->Feature map set->Middle feature map->Mean and variance of output vector obtained after inputting object detection model on edge server, ++>Is in the range of +.>The method comprises the steps of carrying out a first treatment on the surface of the Calculate->Ratio of energy consumption to battery remaining energy required for image transmission and training on end devices +.>The calculation mode is as follows:
，
wherein,bandwidth allocated to the edge server to the terminal device, < >>For the transmission power of the terminal device, +.>For the channel gain +.>Is channel noise>For the data volume of an image, +.>The energy consumption required for locally training an image;
step 2-3: calculating an adaptive sampling rateThe calculation mode is as follows:
，
wherein,、/>and->Parameters (respectively)>、/>And->The accuracy relative loss threshold value, the feature map similarity threshold value and the energy ratio threshold value are respectively +.>And->For the lower and upper bounds of the sampling rate, +.>Indicating that the calculation result is not out of limits; computing training trigger instruction +.>The calculation mode is as follows:
，
step 2-4: according to the training trigger instruction of the edge serverJudging the data needing to be returned according to the calculation result, ifThen the high-precision model detection result on the edge server is +.>、/>Time sample rate->Training trigger instruction +.>Issuing to terminal equipment; if->Only issuing training trigger instruction +.>Andtime sample rate->。
4. The method for detecting the low-energy-consumption target based on the adaptive knowledge distillation according to claim 1, wherein in the step 3, the terminal equipment randomly extracts the same number of image feature images from the experience pool, and forms a training set with the downloaded image feature images and labels thereof, so as to train the carried model and update the sampling rate; after training is completed, updating the experience pool, and replacing the extracted image feature map with the downloaded image feature map; the method specifically comprises the following steps:
step 3-1: the terminal equipment receives the data returned from the edge server; if training trigger instructionThe device update sampling rate is +.>The method comprises the steps of carrying out a first treatment on the surface of the If training trigger instruction->Performing the training operation of the step 3-2;
step 3-2: if training trigger instructionRandomly extracting the number +.>Is associated with the feature map set stored in the last time slot +.>And detection result downloaded from edge server +.>Form training setThen training a target detection model on the terminal equipment; after training, the feature map is assembled +.>Detection result->Replacing the extracted feature map, storing the feature map in an experience pool, and updating the sampling rate of the device to be +.>The method comprises the steps of carrying out a first treatment on the surface of the After completion, steps 1-1 to 3-2 are repeated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410104431.6A CN117635924B (en) | 2024-01-25 | 2024-01-25 | Low-energy-consumption target detection method based on adaptive knowledge distillation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410104431.6A CN117635924B (en) | 2024-01-25 | 2024-01-25 | Low-energy-consumption target detection method based on adaptive knowledge distillation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117635924A true CN117635924A (en) | 2024-03-01 |
| CN117635924B CN117635924B (en) | 2024-05-07 |
Family
ID=90030763
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410104431.6A Active CN117635924B (en) | 2024-01-25 | 2024-01-25 | Low-energy-consumption target detection method based on adaptive knowledge distillation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117635924B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109507292A (en) * | 2018-12-26 | 2019-03-22 | 西安科技大学 | A kind of method for extracting signal |
| CN110096362A (en) * | 2019-04-24 | 2019-08-06 | 重庆邮电大学 | A kind of multitask discharging method based on Edge Server cooperation |
| CN113177367A (en) * | 2021-05-28 | 2021-07-27 | 北京邮电大学 | High-energy-efficiency federal learning method and device, edge server and user equipment |
| CN117079132A (en) * | 2023-08-24 | 2023-11-17 | 西安理工大学 | Remote sensing image target detection method based on Gaussian distance loss |
-
2024
- 2024-01-25 CN CN202410104431.6A patent/CN117635924B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109507292A (en) * | 2018-12-26 | 2019-03-22 | 西安科技大学 | A kind of method for extracting signal |
| CN110096362A (en) * | 2019-04-24 | 2019-08-06 | 重庆邮电大学 | A kind of multitask discharging method based on Edge Server cooperation |
| CN113177367A (en) * | 2021-05-28 | 2021-07-27 | 北京邮电大学 | High-energy-efficiency federal learning method and device, edge server and user equipment |
| CN117079132A (en) * | 2023-08-24 | 2023-11-17 | 西安理工大学 | Remote sensing image target detection method based on Gaussian distance loss |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117635924B (en) | 2024-05-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107909108B (en) | Edge cache system and method based on content popularity prediction | |
| CN110968426B (en) | Edge cloud collaborative k-means clustering model optimization method based on online learning | |
| CN109990790B (en) | Unmanned aerial vehicle path planning method and device | |
| WO2023040510A1 (en) | Image anomaly detection model training method and apparatus, and image anomaly detection method and apparatus | |
| CN110705684A (en) | Environment self-adaptive learning method and system based on end cloud cooperation | |
| CN110830807B (en) | Image compression method, device and storage medium | |
| KR102592402B1 (en) | Method of generating deep learning model for user characteristic analysis using federated learning | |
| CN113469325A (en) | Layered federated learning method, computer equipment and storage medium for edge aggregation interval adaptive control | |
| CN110798718A (en) | Video recommendation method and device | |
| CN115809147B (en) | Multi-edge collaborative cache scheduling optimization method, system and model training method | |
| CN109375999A (en) | A kind of MEC Random Task moving method based on Bayesian network | |
| CN111787280A (en) | Video real-time target tracking method and device based on edge calculation | |
| CN114116061B (en) | Workflow task unloading method and system in mobile edge computing environment | |
| CN114925720A (en) | Small sample modulation signal identification method based on space-time mixed feature extraction network | |
| CN117635924B (en) | Low-energy-consumption target detection method based on adaptive knowledge distillation | |
| CN113901931A (en) | Knowledge distillation model-based behavior recognition method for infrared and visible light videos | |
| CN112836822A (en) | Federal learning strategy optimization method and device based on width learning | |
| CN111291663B (en) | Method for quickly segmenting video target object by using space-time information | |
| CN113763420A (en) | Target tracking method, system, storage medium and terminal equipment | |
| CN113191183A (en) | Unsupervised domain false label correction method and unsupervised domain false label correction device in personnel re-identification | |
| CN112465856A (en) | Unmanned aerial vehicle-based ship track correction method and device and electronic equipment | |
| CN114640669A (en) | Edge calculation method and device | |
| CN115577797A (en) | Local noise perception-based federated learning optimization method and system | |
| CN116259057A (en) | Method for solving data heterogeneity problem in federal learning based on alliance game | |
| CN110728217A (en) | SAR image recognition method, device, 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 |