CN111507952A - Embedded-end diabetic retinopathy screening solution - Google Patents
Embedded-end diabetic retinopathy screening solution Download PDFInfo
- Publication number
- CN111507952A CN111507952A CN202010283610.2A CN202010283610A CN111507952A CN 111507952 A CN111507952 A CN 111507952A CN 202010283610 A CN202010283610 A CN 202010283610A CN 111507952 A CN111507952 A CN 111507952A
- Authority
- CN
- China
- Prior art keywords
- embedded
- deep learning
- learning model
- neural network
- diabetic retinopathy
- 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
- 206010012689 Diabetic retinopathy Diseases 0.000 title claims abstract description 33
- 238000012216 screening Methods 0.000 title claims abstract description 30
- 238000013136 deep learning model Methods 0.000 claims abstract description 41
- 238000000034 method Methods 0.000 claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000012545 processing Methods 0.000 claims abstract description 15
- 238000013528 artificial neural network Methods 0.000 claims abstract description 13
- 238000013527 convolutional neural network Methods 0.000 claims description 30
- 238000013139 quantization Methods 0.000 claims description 24
- 238000004364 calculation method Methods 0.000 claims description 18
- 238000009826 distribution Methods 0.000 claims description 11
- 230000018109 developmental process Effects 0.000 claims description 9
- 238000011161 development Methods 0.000 claims description 6
- 230000000694 effects Effects 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 230000011218 segmentation Effects 0.000 claims description 6
- 238000013461 design Methods 0.000 claims description 4
- 238000004891 communication Methods 0.000 claims description 3
- 238000007667 floating Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 230000002452 interceptive effect Effects 0.000 claims description 3
- 230000003287 optical effect Effects 0.000 claims description 3
- 238000005728 strengthening Methods 0.000 claims 1
- 238000009440 infrastructure construction Methods 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 7
- 206010012601 diabetes mellitus Diseases 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 201000004569 Blindness Diseases 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000009877 rendering Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20092—Interactive image processing based on input by user
- G06T2207/20104—Interactive definition of region of interest [ROI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30041—Eye; Retina; Ophthalmic
Abstract
The invention discloses a method for screening and solving diabetic retinopathy at an embedded end, which comprises the following steps of 1) designing an embedded system with a heterogeneous microprocessor as a core; 2) developing a deep learning model, and training a deep convolution neural network by using a large number of fundus pictures; 3) quantifying parameters of the neural network trained in the step (2); 4) the image acquisition software, the image processing algorithm and the deep learning model reasoning are linked in the embedded platform; the embedded-end diabetic retinopathy screening solution provided by the invention is a low-cost and high-efficiency solution, belongs to an edge computing solution, has the advantages that the data flow direction of the fundus photo is completely mastered by a user, the data security is higher, the cloud computing and the internet are not depended on, and the AI sugar network screening can be performed in remote areas with weak infrastructure construction and primary hospitals.
Description
Technical Field
The invention relates to an artificial intelligence technology product combining deep learning and embedded system technology, which is applied to screening of diabetic retinopathy aiming at fundus pictures. Specifically, the method is a software and hardware integrated solution combining a deep learning model and an AI chip, and the NPU (neural-network processing units) of the AI chip is used for accelerating the operation of the deep learning model at the embedded end so as to achieve the effect of real-time reasoning.
Background
Diabetic Retinopathy (DR) is one of the most common complications of diabetes, and can lead to severe blindness. The proportion of diabetic retinopathy patients to diabetic patients is 24.7% -37%, about one diabetic retinopathy patient is in every three diabetic patients, and based on the huge base of diabetic patients (more than 1 hundred million diabetic patients exist in China), the potential diabetic retinopathy patient population has a very large scale. Therefore, the screening of the diabetic retinopathy has great significance for promoting the social health.
The invention provides a software and hardware integrated solution combining a deep learning technology and an embedded system. The invention can complete a whole set of work of fundus image processing, deep learning model reasoning, screening report generation and the like in real time at the embedded terminal. Due to the technical scheme, the method does not need to depend on a server, a cloud and a network, and is very suitable for large-scale screening work of the diabetic retinopathy in primary hospitals and underdeveloped areas with weak infrastructure conditions. The solution provided by the invention has the advantages of high efficiency (the actual measurement time of a single screening case is less than 1 second), low cost (the cost is only 1/50 of a low-end deep learning server), safer data and the like, and is a diabetic retinopathy screening solution with large-scale commercial potential.
Disclosure of Invention
The invention relates to an embedded diabetic retinopathy solution, which can complete image processing, deep learning model reasoning, screening report generation and other tasks aiming at fundus photos in real time at an embedded terminal.
In order to realize the technical effects, the adopted specific technical scheme is as follows:
the technical process is as follows:
1) designing an embedded system with a heterogeneous microprocessor as a core;
2) developing a deep learning model, and training a deep convolution neural network by using a large number of fundus pictures;
3) quantizing parameters of the neural network trained in the step (2), quantizing original 32 floating point type parameters of the deep learning model into 8-bit integer type parameters, solidifying the deep learning model into an NPU drive recognizable model in an AI chip after confirming that the loss caused by quantization operation on the model precision is acceptable, and deploying;
4) the image acquisition software, the image processing algorithm and the deep learning model reasoning are linked in the embedded platform.
The embedded-end diabetic retinopathy screening solution comprises: 1) in the middle, the development of the embedded system comprises two blocks of hardware development and software development;
in the aspect of hardware, the main control chip selects a heterogeneous microprocessor carrying an NPU, a hardware module surrounding the processor is provided with an RAM chip as a memory and a ROM chip as a storage structure of a hard disk, and the hardware simultaneously comprises a camera interface and an Ethernet interface which are respectively used for fundus photo data acquisition and communication of a local area network;
the software hierarchical design comprises a bottommost layer of driver → a customized linux kernel → an interactive framework of an embedded end, wherein the bottommost layer of driver comprises an NPU driver and a USB driver;
based on the software and hardware environment, the functions of image acquisition and report generation are developed.
The embedded-end diabetic retinopathy screening solution comprises: 2) in the method, a deep learning model is trained by utilizing a large number of marked fundus pictures, and a deep convolutional neural network is used as a main network for classifying diabetic retinopathy; firstly, image preprocessing is carried out on an original fundus picture to obtain a fundus picture with enhanced characteristics, the purpose of the step is to highlight data characteristics and improve the training efficiency and the training effect of a convolutional neural network, the fundus picture with enhanced characteristics is input into the convolutional neural network, and meanwhile, a label corresponding to the fundus picture is input, wherein the label is a target to be approached by the convolutional neural network; the training of the convolutional neural network is a process of iterative forward calculation and backward calculation, the forward calculation calculates the difference between the output of the convolutional neural network and the target, and the backward calculation updates the weight of the convolutional neural network layer by layer according to the difference until the difference between the output of the convolutional neural network and the target is small enough, and the training process is finished.
The embedded-end diabetic retinopathy screening solution comprises: 3) in the method, the weight of the deep learning model is quantized before the deep learning model is deployed on an embedded platform so as to achieve the purposes of reducing the calculation amount and memory access; meanwhile, in the trained deep convolutional neural network, the weight numerical range of each layer is relatively determined and the fluctuation is small, so that the dynamic range of the deep convolutional neural network is suitable for quantization; the mathematical principle of weight quantization operation is also clear, the weight is mapped from one value range to the other value range on the premise that the weight distribution information is reserved due to the mapping of the weight in the two value ranges, errors caused by the weight of the deep learning model due to weight quantization are considered as noise, and the deep convolutional neural network is just insensitive to the noise, so that the accuracy loss finally brought by quantization is within an acceptable range.
The embedded-end diabetic retinopathy screening solution comprises: 4) in the method, an embedded multimedia frame gstreamer is adopted as a self-defined camera to finish the acquisition of fundus images, and an optical device is needed for assisting a fundus picture acquisition camera; after the data acquisition is finished, processing the pictures by aiming at an image processing algorithm of the fundus picture to strengthen the characteristics of the pictures, then inputting the processed images into a deep learning model for reasoning, finally pushing a reasoning result and an original picture outwards in a data stream mode, and adopting a UDP protocol to carry out local area network flow pushing.
The embedded-end diabetic retinopathy screening solution comprises: 2) in the method, the input original fundus picture preprocessing flow is as follows:
a, performing threshold segmentation on an original fundus picture;
b, obtaining an ROI (region of interest) region of the fundus picture after threshold segmentation in the step a, and performing crop operation according to the ROI;
c, performing resize operation on the result of the step b, because the deep learning model of the invention requires that the size of the input picture is fixed;
d, carrying out image contrast enhancement operation on the result of the step c.
The embedded-end diabetic retinopathy screening solution comprises: 3) in the method, the weight quantization is performed on the trained deep learning model, and the principle is as follows:
intercepting a proper boundary +/-T on the original weight distribution, wherein the proper boundary enables the weight distribution in the boundary to be more balanced, the quantization error is smaller after value domain mapping is carried out, and using a tool tensorrT for weight quantization, wherein the tool is used for optimizing K L-divergence at the selected boundary, the formula of K L-divergence is shown as formula (1), wherein p (i) and q (i) represent elements of two distributions, and the meaning of the formula is the difference of the two distributions:
compared with the prior art, the invention has the following beneficial effects:
1. the invention focuses on the embedded end to complete the screening of the diabetic retinopathy of the fundus picture, and is a solution with low cost and high efficiency.
2. The invention belongs to an edge calculation solution, the data flow direction of the fundus picture is completely mastered by a user, and the data security is higher.
3, the solution of the invention does not depend on cloud computing and the Internet, and can play the advantage of AI sugar network screening in remote areas with weak infrastructure construction and primary hospitals.
Drawings
Fig. 1(a) is a schematic diagram of the overall hardware architecture.
FIG. 1(b) is a schematic diagram of a software hierarchy design.
Fig. 2 is a schematic diagram of a training process.
Fig. 3 is a diagram illustrating the mathematical principle of the weight quantization operation.
Fig. 4 is a schematic diagram of the overall process.
Fig. 5(a) is a schematic diagram of weight (parameter) quantization of the trained deep learning model.
Fig. 5(b) is a schematic diagram of weight (parameter) quantization of the trained deep learning model.
Detailed Description
The invention is further described below with reference to the figures and examples.
The solution provided by the invention adopts a deep convolutional neural network (deep convolutional neural-network) as a core AI module to realize high-precision screening (sensitivity is 96%, specificity is 94%). An AI chip with NPU as a core is used as the operation acceleration hardware of the neural network. Meanwhile, in the solution of the invention, a plurality of image processing tool libraries, embedded end multimedia frames and the like are transplanted and cut in the embedded system. The embedded end of the invention can well meet the requirements of color fundus image processing, rendering and the like.
The invention relates to an embedded diabetic retinopathy solution, which can complete image processing, deep learning model reasoning, screening report generation and other tasks aiming at fundus photos in real time at an embedded terminal.
In order to realize the technical effects, the adopted specific technical scheme is as follows:
the technical process is as follows:
1) designing an embedded system taking a heterogeneous microprocessor (multi-core CPU + NPU) as a core;
2) developing a deep learning model, and training a deep convolution neural network by using a large number of fundus pictures;
3) quantizing parameters of the neural network trained in the step (2), quantizing original 32 floating point type parameters of the deep learning model into 8-bit integer type parameters, solidifying the deep learning model into an NPU drive recognizable model in an AI chip after confirming that the loss caused by quantization operation on the model precision is acceptable, and deploying;
4) the image acquisition software, the image processing algorithm and the deep learning model reasoning are linked in the embedded platform.
Wherein the content of the first and second substances,
1) in the middle, the development of the embedded system comprises two blocks of hardware development and software development;
in the aspect of hardware, because the huge operation requirement of a deep neural network is to be met, a heterogeneous microprocessor carrying an NPU is selected as a main control chip, a hardware module surrounding the processor is provided with a RAM (random access Memory) chip as a Memory and a ROM (Read-Only Memory) chip as a storage structure of a hard disk, the hardware simultaneously comprises a camera interface and an Ethernet interface which are respectively used for fundus photo data acquisition and communication, and the whole hardware architecture is shown in fig. 1 (a);
the software hierarchical design comprises a bottommost driver, wherein the bottommost driver comprises an NPU driver, a USBdriver and the like → a customized linux kernel → an interactive framework of an embedded end, and image acquisition and report generation functions are developed based on the software and hardware environment, as shown in fig. 1 (b).
2) In the method, a Deep learning model is trained by utilizing a large number of labeled fundus pictures, and a Deep convolution neural network (Deep convolution neural network) is adopted as a main network for classifying the diabetic retinopathy; the training process is as shown in figure 2, firstly, image preprocessing is carried out on an original fundus picture to obtain a fundus picture with enhanced characteristics, the purpose of the step is to highlight data characteristics and improve the training efficiency and the training effect of a convolutional neural network, the fundus picture with enhanced characteristics is input into the convolutional neural network, and meanwhile, a label corresponding to the fundus picture is input, wherein the label is a target (target) to which the convolutional neural network needs to approach; the training of the convolutional neural network is a process of iterative forward calculation and backward calculation, the forward calculation calculates the difference (loss) between the output of the convolutional neural network and the target, and according to the loss, the backward calculation updates the weight of the convolutional neural network layer by layer until the loss of the output of the convolutional neural network and the target is small enough, and the training process is finished.
3) In the method, because the deep convolutional neural network adopted by the invention needs a large amount of calculation for one-time inference and prediction, and simultaneously because of the limitation of the calculation power of the hardware of the embedded platform, the weight (parameter) of the deep learning model needs to be quantized before the deep learning model is deployed to the embedded platform, so as to achieve the purposes of reducing the calculation amount and memory access; meanwhile, in the trained deep convolutional neural network, the weight numerical range of each layer is relatively determined and the fluctuation is small, so that the dynamic range of the deep convolutional neural network is suitable for quantization; the mathematical principle of the weight quantization operation is relatively clear, and as shown in fig. 3, the weights are mapped from one value range to another value range on the premise that the mapping of the two value ranges keeps weight distribution information as much as possible. The error caused by weight (parameter) quantization to the weight of the deep learning model can be considered as noise, and the deep convolutional neural network is just insensitive to the noise, so that the accuracy loss finally brought by quantization is within an acceptable range.
4) In the invention, an embedded multimedia frame gstreamer is adopted as a self-defined camera to finish the collection of fundus images (the fundus photo collection camera needs the assistance of optical equipment); after the data acquisition is finished, the pictures are processed by aiming at an image processing algorithm of the fundus pictures, the characteristics of the pictures are enhanced, the processed images are input into a deep learning model for reasoning, and finally the reasoning result and the original pictures are pushed outwards in a data stream mode.
The embedded type end diabetic retinopathy screening solution comprises the following steps:
2) in the above, the flow of preprocessing (feature enhancement) of the input original fundus picture is as follows:
a, performing threshold segmentation on the original fundus picture.
b, obtaining an ROI (region of interest) region of the fundus picture after threshold segmentation in the step a, and performing crop operation according to the ROI.
c resize the result of step b, since the deep learning model of the present invention requires the input picture size to be fixed.
d, carrying out image contrast enhancement operation on the result of the step c.
3) In the method, the trained deep learning model is subjected to weight (parameter) quantization, and the principle is as follows:
the method is characterized in that if the weight value is unevenly distributed in an original data domain, a part of the value domain is wasted, quantization in the way brings great errors, and precision loss is great as shown in figure 5 (a). therefore, a better scheme is that a scheme as shown in figure 5(b) is adopted, a proper boundary +/-T is cut out on the original weight distribution, the weight distribution in the boundary can be more balanced by the proper boundary, and the quantization error is smaller after value domain mapping is completed, the method adopts a tool tensorrT of NVIDIA company to carry out weight quantization, the tool is used for selecting K L-divergence to carry out optimization, and the formula of K L-divergence is shown in formula (1):
although the present invention has been described with respect to the preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (7)
1. An embedded-end diabetic retinopathy screening solution is characterized in that:
1) designing an embedded system with a heterogeneous microprocessor as a core;
2) developing a deep learning model, and training a deep convolution neural network by using the fundus picture;
3) quantizing parameters of the neural network trained in the step (2), quantizing original 32 floating point type parameters of the deep learning model into 8-bit integer type parameters, solidifying the deep learning model into an NPU drive recognizable model in an AI chip after confirming that the loss caused by quantization operation on the model precision is acceptable, and deploying;
4) and 3, the linkage of the image acquisition software, the image processing algorithm and the deep learning model in the embedded platform is inferred.
2. The embedded-end diabetic retinopathy screening solution of claim 1, wherein: 1) in the middle, the development of the embedded system comprises two blocks of hardware development and software development;
in the aspect of hardware, a main control chip selects a heterogeneous microprocessor carrying an NPU, a hardware module surrounding the processor utilizes an RAM chip as a memory and a ROM chip as a storage structure of a hard disk, and the hardware simultaneously comprises a camera interface and an Ethernet interface which are respectively used for fundus photo data acquisition and communication;
the software hierarchical design comprises the steps of a bottommost layer of driver → a customized linux kernel → an interactive framework of an embedded end, wherein the bottommost layer of driver comprises an NPU driver and a USB driver;
based on the software and hardware environments, image acquisition and report generation functions are developed.
3. The embedded-end diabetic retinopathy screening solution of claim 2, wherein: 2) in the method, a deep learning model is trained by utilizing a large number of marked fundus pictures, and a deep convolutional neural network is used as a main network for classifying diabetic retinopathy; firstly, preprocessing an original fundus picture to obtain a fundus picture with enhanced characteristics, wherein the purpose of the step is to highlight data characteristics and improve the training efficiency and the training effect of a convolutional neural network; the training of the convolutional neural network is a process of repeatedly iterating forward calculation and backward calculation, the forward calculation calculates the difference between the output of the convolutional neural network and the target, and the backward calculation updates the weight of the convolutional neural network layer by layer according to the difference until the difference between the output of the convolutional neural network and the target is small enough, and the training process is finished.
4. The embedded-end diabetic retinopathy screening solution of claim 3, wherein: 3) in the method, the weight of the deep learning model is quantized before the deep learning model is deployed on the embedded platform, so that the aims of reducing the calculation amount and memory access are fulfilled; the weight is mapped from one value range to the other value range on the premise of keeping weight distribution information through the mapping of the two value ranges.
5. The embedded-end diabetic retinopathy screening solution of claim 4, wherein: 4) in the method, an embedded multimedia frame gstreamer is adopted as a self-defined camera to finish the acquisition of fundus images, and an optical device is needed for assisting a fundus picture acquisition camera; after the data acquisition is finished, processing the pictures by aiming at an image processing algorithm of the fundus picture, strengthening the characteristics of the pictures, inputting the processed images into a deep learning model for reasoning, and finally pushing a reasoning result and an original picture outwards in a data stream mode by adopting a UDP protocol.
6. The embedded-end diabetic retinopathy screening solution of claim 5, wherein: 2) in the method, the input original fundus picture preprocessing flow is as follows:
a, performing threshold segmentation on an original fundus picture;
b, obtaining an ROI (region of interest) region of the fundus picture after threshold segmentation in the step a, and performing crop operation according to the ROI;
c, performing resize operation on the result of the step b, because the deep learning model requires that the size of the input picture is fixed;
d, carrying out image contrast enhancement operation on the result of the step c.
7. The embedded-end diabetic retinopathy screening solution of claim 6, wherein: 3) in the method, the weight quantization is performed on the trained deep learning model, and the principle is as follows:
intercepting a proper boundary +/-T on the original weight distribution, enabling the weight distribution in the boundary to be more balanced by the proper boundary, after value domain mapping is carried out, the quantization error is smaller, adopting a tool tensorrT to carry out weight quantization, and optimizing the tool by selecting K L-divergence on the boundary, wherein the formula of K L-divergence is shown as a formula (1):
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010283610.2A CN111507952A (en) | 2020-04-13 | 2020-04-13 | Embedded-end diabetic retinopathy screening solution |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010283610.2A CN111507952A (en) | 2020-04-13 | 2020-04-13 | Embedded-end diabetic retinopathy screening solution |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111507952A true CN111507952A (en) | 2020-08-07 |
Family
ID=71876025
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010283610.2A Pending CN111507952A (en) | 2020-04-13 | 2020-04-13 | Embedded-end diabetic retinopathy screening solution |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111507952A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11562231B2 (en) * | 2018-09-03 | 2023-01-24 | Tesla, Inc. | Neural networks for embedded devices |
US11983630B2 (en) | 2023-01-19 | 2024-05-14 | Tesla, Inc. | Neural networks for embedded devices |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180018451A1 (en) * | 2016-07-14 | 2018-01-18 | Magic Leap, Inc. | Deep neural network for iris identification |
CN108021916A (en) * | 2017-12-31 | 2018-05-11 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on notice mechanism |
CN108470359A (en) * | 2018-02-11 | 2018-08-31 | 艾视医疗科技成都有限公司 | A kind of diabetic retinal eye fundus image lesion detection method |
CN108734285A (en) * | 2017-04-24 | 2018-11-02 | 英特尔公司 | The calculation optimization of neural network |
CN109598294A (en) * | 2018-11-23 | 2019-04-09 | 哈尔滨工程大学 | Cloud retina OCT identification intelligent diagnostic system and its application method based on hardware and software platform |
CN109993074A (en) * | 2019-03-14 | 2019-07-09 | 杭州飞步科技有限公司 | Assist processing method, device, equipment and the storage medium driven |
-
2020
- 2020-04-13 CN CN202010283610.2A patent/CN111507952A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180018451A1 (en) * | 2016-07-14 | 2018-01-18 | Magic Leap, Inc. | Deep neural network for iris identification |
CN108734285A (en) * | 2017-04-24 | 2018-11-02 | 英特尔公司 | The calculation optimization of neural network |
CN108021916A (en) * | 2017-12-31 | 2018-05-11 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on notice mechanism |
CN108470359A (en) * | 2018-02-11 | 2018-08-31 | 艾视医疗科技成都有限公司 | A kind of diabetic retinal eye fundus image lesion detection method |
CN109598294A (en) * | 2018-11-23 | 2019-04-09 | 哈尔滨工程大学 | Cloud retina OCT identification intelligent diagnostic system and its application method based on hardware and software platform |
CN109993074A (en) * | 2019-03-14 | 2019-07-09 | 杭州飞步科技有限公司 | Assist processing method, device, equipment and the storage medium driven |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11562231B2 (en) * | 2018-09-03 | 2023-01-24 | Tesla, Inc. | Neural networks for embedded devices |
US11983630B2 (en) | 2023-01-19 | 2024-05-14 | Tesla, Inc. | Neural networks for embedded devices |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11798132B2 (en) | Image inpainting method and apparatus, computer device, and storage medium | |
US20220309674A1 (en) | Medical image segmentation method based on u-net | |
CN110287849B (en) | Lightweight depth network image target detection method suitable for raspberry pi | |
CN107798381B (en) | Image identification method based on convolutional neural network | |
WO2022227913A1 (en) | Double-feature fusion semantic segmentation system and method based on internet of things perception | |
WO2021249255A1 (en) | Grabbing detection method based on rp-resnet | |
WO2023185785A1 (en) | Image processing method, model training method, and related apparatuses | |
CN111767922B (en) | Image semantic segmentation method and network based on convolutional neural network | |
WO2022156561A1 (en) | Method and device for natural language processing | |
WO2021051987A1 (en) | Method and apparatus for training neural network model | |
WO2023236977A1 (en) | Data processing method and related device | |
WO2022242127A1 (en) | Image feature extraction method and apparatus, and electronic device and storage medium | |
US20240019777A1 (en) | Training method and apparatus for lithographic mask generation model, device and storage medium | |
WO2021169366A1 (en) | Data enhancement method and apparatus | |
Wu | Image retrieval method based on deep learning semantic feature extraction and regularization softmax | |
CN114863539A (en) | Portrait key point detection method and system based on feature fusion | |
CN111815593A (en) | Lung nodule domain adaptive segmentation method and device based on counterstudy and storage medium | |
CN111507952A (en) | Embedded-end diabetic retinopathy screening solution | |
CN113920379A (en) | Zero sample image classification method based on knowledge assistance | |
CN113837290A (en) | Unsupervised unpaired image translation method based on attention generator network | |
CN111723934B (en) | Image processing method and system, electronic device and storage medium | |
CN116051699B (en) | Dynamic capture data processing method, device, equipment and storage medium | |
CN110991279B (en) | Document Image Analysis and Recognition Method and System | |
Liu et al. | Deep neural networks with attention mechanism for monocular depth estimation on embedded devices | |
CN112668543A (en) | Isolated word sign language recognition method based on hand model perception |
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 | ||
TA01 | Transfer of patent application right | ||
TA01 | Transfer of patent application right |
Effective date of registration: 20210708 Address after: 215000 room d303, building g-1, shazhouhu science and Technology Innovation Park, Huachang Road, yangshe Town, Zhangjiagang City, Suzhou City, Jiangsu Province Applicant after: Suzhou aikor Intelligent Technology Co.,Ltd. Address before: 201601 building 6, 351 sizhuan Road, Sijing Town, Songjiang District, Shanghai Applicant before: Shanghai Sike Intelligent Technology Co.,Ltd. |
|
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20200807 |