CN113065555A - Lightweight improved target detection method and device based on RuiKe micro platform - Google Patents
Lightweight improved target detection method and device based on RuiKe micro platform Download PDFInfo
- Publication number
- CN113065555A CN113065555A CN202110390883.1A CN202110390883A CN113065555A CN 113065555 A CN113065555 A CN 113065555A CN 202110390883 A CN202110390883 A CN 202110390883A CN 113065555 A CN113065555 A CN 113065555A
- Authority
- CN
- China
- Prior art keywords
- model
- data
- target detection
- algorithm
- lightweight
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 87
- 238000012805 post-processing Methods 0.000 claims abstract description 11
- 230000004913 activation Effects 0.000 claims abstract description 9
- 238000006243 chemical reaction Methods 0.000 claims abstract description 9
- 238000000034 method Methods 0.000 claims abstract description 8
- 238000013139 quantization Methods 0.000 claims abstract description 8
- 238000012545 processing Methods 0.000 claims description 12
- 238000012549 training Methods 0.000 claims description 9
- 230000001133 acceleration Effects 0.000 claims description 7
- 238000013500 data storage Methods 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 5
- 230000003213 activating effect Effects 0.000 claims description 4
- 238000007405 data analysis Methods 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 230000003993 interaction Effects 0.000 claims description 2
- 238000005457 optimization Methods 0.000 abstract description 5
- 230000006872 improvement Effects 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 11
- 238000012360 testing method Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 4
- 238000012795 verification Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000013401 experimental design Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (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)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a light-weight improved target detection method and device based on a Rui-core micro platform. The model is based on Tinyyolov4 algorithm, and prior frame improvement is carried out on the self-built data set by using k-means algorithm; improving the number of convolution output channels and the activation function in a backbone network according to the characteristics of the transplanted and deployed Rui core micro platform hardware unit; and obtaining a lightweight target detection model through an asymmetric 8-bit quantization optimization conversion mode. The device takes an improved algorithm as a core, and an RKNN tool chain is used for converting the improved algorithm into an RKNN model which can be deployed on a Rui-core micro platform; scheduling NPU loading through a CPU, using an rknn model, and accelerating reasoning of image data preprocessed by a GPU to realize forward reasoning; and the CPU schedules the GPU to perform data post-processing, and the processed data is transmitted to the display to be displayed. The method and the device of the invention realize that: when the target detection model after being lightened slightly loses accuracy, the weight parameter quantity is reduced and the detection speed is increased; real-time robust image detection can be achieved.
Description
Technical Field
The invention relates to the field of target detection, in particular to a light-weight improved target detection method and device based on a Rui-core micro platform.
Background
Target detection is a hotspot research topic in computer vision and image processing, and plays an important role in scenes such as photoelectric detection, automatic driving, social life and the like. It mainly comprises two different detection tasks: object position and object class detection, which is intended to accurately locate the position and class of an object of interest appearing in an image. With the rise of neural networks and the improvement of computer performance, the target detection algorithm based on deep learning replaces the traditional target detection algorithm with higher performance indexes to become the mainstream direction of research.
The deep learning algorithm for target detection performs training reasoning work through a Graphics Processing Unit (GPU), and makes a great breakthrough in detection precision and robustness. However, the GPU is limited by the problems of interface flexibility, price and energy consumption, cannot meet the requirements of some detection devices on indexes such as power consumption, portability and the like, and is not suitable for being used as an application platform of an algorithm; meanwhile, the edge device with limited hardware resources cannot directly run the unoptimized model. At present, the academic world mainly researches from three aspects of lightweight model structure design, model compression and high-performance chip architecture. How to realize the detection precision loss on the terminal equipment with limited computing and storing resources is not great, and the difficulty of ensuring the real-time performance of the detection is achieved.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: in view of the above-described problems, an object of the present invention is to provide a light-weight target detection method and apparatus for achieving a detection speed increase with a slight loss of detection accuracy in a core-end micro-terminal device.
The technical scheme adopted by the invention for solving the technical problems is as follows:
according to one aspect of the present invention, there is provided a light-weight target detection method for improving detection accuracy and increasing detection speed in a core-end micro-terminal device, comprising:
1) selecting the Tiny yolov4 as a model structure basis, carrying out clustering processing on a detection target prior frame of an actual application scene data set by using a K-means iterative algorithm, namely formula 1, and replacing the prior frame which is originally trained by a COCO data set in the Tiny yolov4 model with an obtained result.
D(B,C)＝1-IOU(B,C)
In the formula (1), A is the size of a data set clustering box, B is the size of a data set sample box, C is a data set clustering center, IOU represents the interaction ratio of A, B, and D is used as a threshold value of algorithm iteration to judge whether a proper prior box value is obtained.
2) Combining RK3399pro chip architecture resources, utilizing a multiply-add unit (MAC) in the NPU to perform parallel calculation acceleration on the convolutional layer, and changing the number of convolutional output channels in the backbone network into a multiple of 3;
3) replacing the activating function from Leaky Relu to Relu activating function; realizing the fusion of the convolution layer, the Relu activation function and the maximum Pooling layer (MAX Pooling layer) during hardware acceleration; the three are fused and optimized into a neural network basic unit (op operator), so that the calculation bandwidth is reduced;
4) for the algorithm trained by the steps, the floating point tensor of FP32 is converted into an 8-bit unsigned integer (uint8) tensor by asymmetric 8-bit quantization (asymmetric _ quantized-u8), namely formula 2,
in the formula (2), min represents the minimum value of the parameter, max represents the maximum value of the parameter, and xfRepresenting float32 type data, n representing the number of bits of quantized data, xqRepresenting the final quantized value, round represents the data range of the final quantized value.
Preferably, the light-weight improved target detection method based on the Rui-core micro platform provided by the invention comprises the following steps:
1) converting the improved algorithm network into an RKNN model through an RKNN tool chain;
the improved algorithm network takes the Tiny yolov4 as a model base, replaces a prior frame of a self-built data set, changes the number of convolution output channels in a backbone network into a multiple of 3, and takes an activation function as a Relu function; training is carried out through a self-building data set, and an obtained model is converted into a target detection model through asymmetric 8-bit quantization optimization;
2) loading a rknn model on a RuiKe micro platform, and compiling a data acquisition, data analysis and format conversion and data post-processing function to form an algorithm module;
3) after the camera module acquires an input image, transmitting data to the algorithm module;
4) the algorithm module performs data analysis, format conversion and data post-processing on the transmitted image data by using the improved model so as to realize target detection;
5) and outputting the target detection result and the processed image on a display module.
Preferably, the present invention provides a lightweight improved target detection device based on a rayleigh core micro platform, comprising:
the acquisition module is used for configuring Rayleigh core micro RK3399pro platform parameters, and initializing a media interface by using a Rayleigh core micro hardware coding and decoding application program (MPP) middleware, so that a USB camera can acquire surrounding environment information and transmit image data to a GPU;
the processing module is used for scheduling the NPU to load by using a load-RKNN interface by using the CPU, initializing an RKNN Software Development Kit (SDK) environment by using an initialization operation (init _ runtime) interface, configuring model parameters and applying for a data storage space;
processing each frame of image data acquired by a camera through a Rayleigh kernel micro hardware coding and decoding application program (MPP) middleware to ensure that the format, the size and the channel number of the image all meet the condition of algorithm input;
the CPU schedules the NPU to use a hardware inference (inference) interface to infer each frame of image after data processing;
the data post-processing module is used for analyzing the data of the result obtained by inference, marking the detected target type and position information on the image and then displaying the target type and position information on the display through an output interface;
and the ending detection module is used for releasing the video data storage space and the RKNN object through a release resource (release) interface.
Another object of the present invention is to provide a detection apparatus for operating the above light-weight improved target detection method based on a rui-core micro-platform.
Another object of the present invention is to provide a mobile terminal for object detection, which is equipped with the above detection device.
Compared with the prior art, the invention has the following advantages:
by combining all the technical schemes, the invention has the positive effects that:
on the basis of a Tiny yolov4 model network, the invention improves the prior frame according to the data set of the actual scene to be applied; improving the number of convolution output channels and the activation function in a backbone network according to the structural characteristics of a Rui core micro platform hardware unit to be transplanted and deployed; a lightweight improved target detection method based on a Rui-core micro platform is obtained through an asymmetric 8-bit quantization optimization conversion mode; the invention also provides a real-time target detection device which takes the lightweight improved target detection method as a core algorithm. Through the innovation, the invention has the following effects:
1) the weight parameter quantity of the target detection model after being lightened is reduced, and the detection speed is increased;
2) the lightweight target detection model can realize real-time robust image detection on embedded or mobile terminal equipment with limited resources;
the technical results and experimental results include:
the accuracy of the lightweight target detection model on the self-built data set is slightly lower than that of the Tiny yolov4 model on the self-built data set; a large increase in detection speed; the quantized model weight is reduced to 1/4; the lightweight improved target detection device based on the RuiKe micro is characterized in that on an RK3399pro platform, the rapid transmission processing capacity of image data is realized through RuiKe micro hardware coding and decoding application program (MPP) middleware, and the rapid hardware acceleration reasoning capacity is realized on a model through an NPU; the detection speed of the system is close to 60 frames/second in real-time detection, and the detection requirement of real-time robustness is met.
Drawings
FIG. 1 is a network architecture diagram of the Tiny yolov4 model provided by an embodiment of the present invention;
FIG. 2 is a network architecture diagram of an improved model provided by an embodiment of the present invention;
FIG. 3 is a schematic diagram of the NPU fusion op operator acceleration provided by the embodiment of the present invention;
FIG. 4 is a flow chart of an application of development of a Rayleigh core micro platform model according to an embodiment of the present invention;
FIG. 5 is a diagram showing the detection results provided by the embodiment of the present invention;
fig. 6 is a schematic view of a multithread workflow of the lightweight improved target detection apparatus according to an embodiment of the present invention.
Detailed Description
In order to make the technical scheme of the invention more obvious, the invention is further explained in detail by combining the embodiment and the attached drawings
The invention provides a light-weight improved target detection method based on a Rui-core micro platform, which comprises the following steps:
the original prior frame in the Tiny yolov4 is trained according to a COCO data set, is not suitable for being applied to a specific or special scene, and is used for clustering a self-built data set of an actual scene to be applied by using a k-means algorithm to obtain data of 6 prior frames for replacement;
the self-built data set is derived from picture data collected in a laboratory and public picture data acquired from a network, and mainly comprises two major categories of vehicles and pedestrians. After manual screening, labeling by using a Label image data labeling tool, removing a part with an over-fuzzy target size during labeling, and storing a Label as an xml file by default; 4000 pictures are distributed to the marked pictures for training, and 500 pictures are used for verification and testing;
the NPU in the RuiKe micro RK3399pro platform mainly comprises a multiply-add unit (MAC), and in order to obtain the highest resource utilization rate when hardware acceleration is carried out on a convolutional layer, the number of convolution output channels in a backbone network is changed into a multiple of 3;
in order to reduce parameter redundancy in the hardware reasoning process, replacing an activation function with a Relu function;
the network structure of the Tiny yolov4 model before improvement is shown in figure 1;
the network structure of the modified Tiny yolov4 model is shown in FIG. 2;
obtaining a lightweight target detection model by the improved algorithm network in an asymmetric 8-bit quantization optimization conversion mode;
the invention is further described below in conjunction with specific experimental designs and results.
In order to embody the applicability of the just and improved model, the training and testing are carried out on the Tiny yolov4 model and the model provided by the invention on the KITTI data set by setting a test set and a verification set under the condition of setting the same target detection threshold and NMS threshold, and the experimental results are compared and analyzed.
The KITTI data set is a computer vision algorithm evaluation data set under the current international largest automatic driving scene. The original category labels are required to be processed, and the unnecessary category labels are abandoned. And finally, obtaining five category labels including different types of vehicles and pedestrians. And the processed pictures are divided into 6058 training sets, 674 verification sets and 749 test sets.
Evaluating the analysis indexes mainly by weight parameters of model training; comparing and analyzing the detection result of the model in the verification set with the data of the real labeling frame; obtaining an Average Precision mean value (map for short) of target detection; and the speed of the model Per Second in video detection (Frame Per Second, FPS for short); three aspects were evaluated;
after five experiments, the training error of a single model is basically eliminated; the detection result comparison shown in the table I can be obtained; the results of three evaluation analysis indexes can be seen; compared with the Tiny yolov4 model, the average precision mean (map) of the model of the invention on the KITTI data set is reduced from 81.29% to 79.56%; after sacrificing a little detection accuracy, the method can obtain great detection speed improvement by using less model weight parameters;
TABLE 1 model test results on KITTI datasets
After the validity of the algorithm is verified, the invention provides a lightweight improved target detection device which takes the algorithm of the invention as a core algorithm and is realized based on a Rui-core micro platform, and the lightweight improved target detection device comprises:
the acquisition module is used for configuring Rayleigh core micro RK3399pro platform parameters, initializing a media interface by using a Rayleigh core micro hardware coding and decoding application program (MPP) middleware, and enabling the inserted USB camera to transmit external image information through the Rayleigh core micro hardware coding and decoding application program (MPP) middleware;
the processing module is responsible for scheduling the NPU and the GPU by using the CPU in the RK3399pro to complete: (1) acquiring and preprocessing an image; (2) data reasoning; (3) three functions of data post-processing and displaying are realized, and the whole flow framework of the detection device is shown in FIG. 6;
the CPU is used for scheduling the GPU to preprocess the surrounding environment information acquired by the USB camera, so that the image format, the size and the channel number all meet the condition of algorithm input;
the CPU schedules the NPU to load the image data preprocessed by the GPU, and a load rknn model (load _ rknn) interface is used for loading the rknn model, and the steps of loading the reasoning model on the whole Rui-core micro-platform are shown in FIG. 4; initializing an RKNN Software Development Kit (SDK) environment by using an initialization operation (init _ runtime) interface, configuring model parameters and applying for a data storage space;
the data post-processing module is used for the CPU to schedule the GPU to analyze the data of the result obtained by inference, label important information such as the type and the position of a detected target on an image and then display the information on a display through an output interface;
an ending detection module used for releasing the video data storage space and the RKNN object through a release resource (release) interface;
the NPU focuses on the time consumption of the whole model reasoning process when the reasoning is accelerated, and the image preprocessing and post-processing time is longer. The image data preprocessed by the GPU is reasoned by utilizing the two big cores A72 time-sharing multiplexing NPUs;
the invention relates to a light-weight improved target detection method and device based on a Rui-core micro platform. The method is mainly based on the Tiny Yolov4 and optimized by improving the prior frame, the convolution channel number and the activation function of the data set; training the data set through the self-built data set; and carrying out asymmetrical 8-bit quantization optimization conversion on the trained model to obtain a lightweight target detection model. When the detection device is constructed by means of a RuiEn micro-platform deployment algorithm, the improved lightweight target detection algorithm cannot be directly operated on RK3399 pro; therefore, the improved lightweight target detection algorithm is firstly converted into an RKNN model through an RKNN tool chain. Scheduling NPU loading through a CPU, using an rknn model, and accelerating reasoning of image data preprocessed by a GPU to realize forward reasoning; and the CPU schedules the GPU to perform data post-processing, and the processed data is transmitted to the display to be displayed.
Claims (8)
1. A lightweight improved target detection method based on a Rui-core micro platform is characterized in that: the lightweight improved target detection method based on the RuiKe micro platform comprises the following processes:
selecting the Tiny yolov4 as a model structure basis, carrying out clustering processing on a detection target prior frame of an actual application scene data set by using a K-means iterative algorithm, namely a formula (1), replacing the prior frame which is originally trained by a COCO data set in the Tiny yolov4 model with an obtained result,
D(B,C)＝1-IOU(B,C)
in the formula (1), A is the size of a data set clustering box, B is the size of a data set sample box, C is a data set clustering center, IOU represents the interaction ratio of A, B, and D is used as a threshold value of algorithm iteration to judge whether a proper prior box value is obtained.
2. The method for lightweight and improved target detection based on the Rui-Ke micro-platform as claimed in claim 1, wherein the number of convolution output channels in the backbone network is changed to a multiple of 3 by using a multiplication and addition unit (MAC) in the NPU to perform parallel computation acceleration on the convolution layer in combination with RK3399pro chip architecture resources.
3. The lightweight improved target detection method based on the Rui-core micro platform as claimed in claim 2, characterized in that: replacing the activating function from Leaky Relu to Relu activating function; the convolution layer, the Relu activation function and the maximum Pooling layer (MAX Pooling layer) are fused during hardware acceleration, and the fusion of the convolution layer, the Relu activation function and the maximum Pooling layer is optimized into a neural network computing basic unit (op operator), so that the computing bandwidth is reduced.
4. The lightweight improved target detection method based on the Rui-core micro platform as claimed in claim 3, characterized in that: the floating point tensor of FP32 is converted to an 8-bit unsigned integer (uint8) tensor by asymmetric 8-bit quantization (asymmetric _ quantized-u8), i.e., equation (2),
in formula (2), min represents a minimum parameter value, max represents a maximum parameter value, xf represents float32 type data, n represents the number of bits of quantized data, xq represents a final quantized value, and round represents a data range of the final quantized value.
5. The lightweight improved target detection method based on the Rui-core micro platform as claimed in claim 4, wherein: the lightweight improved target detection method based on the Rui-core micro platform further comprises the following steps:
1) converting the improved algorithm network into an RKNN model through an RKNN tool chain;
the improved algorithm network takes the Tiny yolov4 as a model base, replaces a prior frame of a self-built data set, changes the number of convolution output channels in a backbone network into a multiple of 3, and takes an activation function as a Relu function; training through a self-building data set to obtain a model; then, a lightweight improved target detection model is obtained through asymmetric 8-bit quantization conversion;
2) loading a rknn model on a RuiKe micro platform, and compiling a data acquisition, data analysis and format conversion and data post-processing function to form an algorithm module;
3) after the camera module acquires an input image, transmitting data to the algorithm module;
4) the algorithm module performs data analysis, format conversion and data post-processing on the transmitted image data by using the improved model so as to realize target detection;
5) and outputting the target detection result and the processed image on a display module.
6. The utility model provides a target detection device is improved in lightweight based on end core micro platform which characterized in that: target detection device is improved in lightweight based on end core micro platform includes:
the acquisition module is used for configuring Rayleigh core micro RK3399pro platform parameters, and initializing a media interface by using a Rayleigh core micro hardware coding and decoding application program (MPP) middleware, so that a USB camera can acquire surrounding environment information and transmit image data to a GPU;
the processing module is used for scheduling the NPU to load by using a load-RKNN interface by using the CPU, initializing an RKNN Software Development Kit (SDK) environment by using an initialization operation (init _ runtime) interface, configuring model parameters and applying for a data storage space;
processing each frame of image data acquired by a camera through a Rayleigh kernel micro hardware coding and decoding application program (MPP) middleware to ensure that the format, the size and the channel number of the image all meet the condition of algorithm input;
the CPU schedules the NPU to use a hardware inference (inference) interface to infer each frame of image after data processing;
the data post-processing module is used for analyzing the data of the result obtained by inference, marking the detected target type and position information on the image and then displaying the target type and position information on the display through an output interface;
and the ending detection module is used for releasing the video data storage space and the RKNN object through a release resource (release) interface.
7. The utility model provides a detection device based on end core micro platform which characterized in that: the detection device operates the method for detecting a lightweight improved object according to any one of claims 1 to 5.
8. A mobile terminal for object detection, characterized by: a detection device according to claim 7 is mounted.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110390883.1A CN113065555A (en) | 2021-04-12 | 2021-04-12 | Lightweight improved target detection method and device based on RuiKe micro platform |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110390883.1A CN113065555A (en) | 2021-04-12 | 2021-04-12 | Lightweight improved target detection method and device based on RuiKe micro platform |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113065555A true CN113065555A (en) | 2021-07-02 |
Family
ID=76566420
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110390883.1A Pending CN113065555A (en) | 2021-04-12 | 2021-04-12 | Lightweight improved target detection method and device based on RuiKe micro platform |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113065555A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113837144A (en) * | 2021-10-25 | 2021-12-24 | 广州微林软件有限公司 | Intelligent image data acquisition and processing method for refrigerator |
| CN114529798A (en) * | 2022-02-21 | 2022-05-24 | 山东浪潮科学研究院有限公司 | Production line product quality inspection implementation method based on TinyML and auxiliary system |
| WO2024114226A1 (en) * | 2022-11-30 | 2024-06-06 | 中兴通讯股份有限公司 | Processing method, chip and apparatus for forward inference, and medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062282A (en) * | 2019-12-05 | 2020-04-24 | 武汉科技大学 | Transformer substation pointer type instrument identification method based on improved YOLOV3 model |
| CN111914937A (en) * | 2020-08-05 | 2020-11-10 | 湖北工业大学 | Lightweight improved target detection method and detection system |
-
2021
- 2021-04-12 CN CN202110390883.1A patent/CN113065555A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062282A (en) * | 2019-12-05 | 2020-04-24 | 武汉科技大学 | Transformer substation pointer type instrument identification method based on improved YOLOV3 model |
| CN111914937A (en) * | 2020-08-05 | 2020-11-10 | 湖北工业大学 | Lightweight improved target detection method and detection system |
Non-Patent Citations (3)
| Title |
|---|
| JOSEPH REDMON等: "YOLOv3: An Incremental Improvement", 《网页在线公开:HTTPS://ARXIV.ORG/ABS/1804.02767V1》 * |
| 丁月: "下一代网络下辅助驾驶安全预警***研究及其实现", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
| 福州瑞芯微电子股份有限公司: "Rockchip Trouble shooting RKNN-Toolkit CN", 《网页在线公开：REPO.ROCK-CHIPS.COM/RK1808/RKNN-TOOLKIT_DOC/》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113837144A (en) * | 2021-10-25 | 2021-12-24 | 广州微林软件有限公司 | Intelligent image data acquisition and processing method for refrigerator |
| CN114529798A (en) * | 2022-02-21 | 2022-05-24 | 山东浪潮科学研究院有限公司 | Production line product quality inspection implementation method based on TinyML and auxiliary system |
| CN114529798B (en) * | 2022-02-21 | 2024-05-21 | 山东浪潮科学研究院有限公司 | TinyML-based production line product quality inspection implementation method and auxiliary system |
| WO2024114226A1 (en) * | 2022-11-30 | 2024-06-06 | 中兴通讯股份有限公司 | Processing method, chip and apparatus for forward inference, and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113065555A (en) | Lightweight improved target detection method and device based on RuiKe micro platform | |
| CN109086678B (en) | Pedestrian detection method for extracting image multilevel features based on deep supervised learning | |
| CN110232696A (en) | A kind of method of image region segmentation, the method and device of model training | |
| CN110378222B (en) | Method and device for detecting vibration damper target and identifying defect of power transmission line | |
| CN112150821B (en) | Lightweight vehicle detection model construction method, system and device | |
| CN113011282A (en) | Graph data processing method and device, electronic equipment and computer storage medium | |
| CN111144329A (en) | Light-weight rapid crowd counting method based on multiple labels | |
| US11475572B2 (en) | Systems and methods for object detection and recognition | |
| CN110852295B (en) | Video behavior recognition method based on multitasking supervised learning | |
| CN111527501A (en) | Chip adaptation determining method and related product | |
| CN114332666A (en) | Image target detection method and system based on lightweight neural network model | |
| CN113076992A (en) | Household garbage detection method and device | |
| CN115049966A (en) | GhostNet-based lightweight YOLO pet identification method | |
| CN109961095A (en) | Image labeling system and mask method based on non-supervisory deep learning | |
| CN109840561A (en) | A kind of rubbish image automatic generation method can be used for garbage classification | |
| CN111860259A (en) | Training and using method, device, equipment and medium of driving detection model | |
| CN113570689A (en) | Portrait cartoon method, apparatus, medium and computing device | |
| CN115348551A (en) | Lightweight service identification method and device, electronic equipment and storage medium | |
| CN113870846B (en) | Speech recognition method, device and storage medium based on artificial intelligence | |
| Akkas et al. | A fast video image detection using tensorflow mobile networks for racing cars | |
| CN113505640A (en) | Small-scale pedestrian detection method based on multi-scale feature fusion | |
| CN116383634A (en) | Landslide signal identification method and device and electronic equipment | |
| CN113408571B (en) | Image classification method and device based on model distillation, storage medium and terminal | |
| CN113554030B (en) | Multi-type license plate recognition method and system based on single character attention | |
| CN115761390A (en) | Image scene recognition method and device |
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: 20210702 |