CN112347857A - Ship detection device and method of optical remote sensing satellite image - Google Patents
Ship detection device and method of optical remote sensing satellite image Download PDFInfo
- Publication number
- CN112347857A CN112347857A CN202011093587.7A CN202011093587A CN112347857A CN 112347857 A CN112347857 A CN 112347857A CN 202011093587 A CN202011093587 A CN 202011093587A CN 112347857 A CN112347857 A CN 112347857A
- Authority
- CN
- China
- Prior art keywords
- layer
- network
- data
- module
- rows
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F13/00—Interconnection of, or transfer of information or other signals between, memories, input/output devices or central processing units
- G06F13/38—Information transfer, e.g. on bus
- G06F13/42—Bus transfer protocol, e.g. handshake; Synchronisation
- G06F13/4204—Bus transfer protocol, e.g. handshake; Synchronisation on a parallel bus
- G06F13/4221—Bus transfer protocol, e.g. handshake; Synchronisation on a parallel bus being an input/output bus, e.g. ISA bus, EISA bus, PCI bus, SCSI bus
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2213/00—Indexing scheme relating to interconnection of, or transfer of information or other signals between, memories, input/output devices or central processing units
- G06F2213/0026—PCI express
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
Abstract
The invention relates to a ship detection device and method of an optical remote sensing satellite image, belonging to the field of automatic processing and algorithm hardware acceleration of the optical remote sensing satellite image; the device comprises an image processing module, a storage module, a data input module, a bus protocol conversion module and a data output module, and the method comprises the following steps: preprocessing an optical remote sensing satellite image, and deleting the number of input channels of a first-layer network; a fused BN normalization processing module; reordering network parameters; the image processing module is used for ship detection and mainly used for solving the problem of low bandwidth utilization rate of the on-orbit satellite.
Description
Technical Field
The invention relates to the field of satellite remote sensing image automatic processing and algorithm hardware acceleration, in particular to a ship detection device and method of an optical remote sensing satellite image.
Background
The communication bandwidth of the on-orbit remote sensing satellite is a key national strategic resource, the remote sensing satellite utilizes a photoelectric remote sensor, a radar or a radio receiver to detect, monitor or track a target from an orbit so as to obtain electromagnetic wave information radiated, reflected or transmitted by the ground, ocean or aerial target, then the electromagnetic wave information is transmitted to a ground receiving station in a radio transmission mode, and valuable military intelligence is extracted from the information through optical and electronic equipment and computer processing. However, most of the returned data is transmitted in a high-resolution and high-spectrum remote sensing image form, and needs to occupy a very large bandwidth, the proportion of the data containing the monitored target to the total data amount is only between 5% and 25%, most of the bandwidth is wasted by invalid data, and communication resources of other loads are occupied, so that how to improve the transmission proportion of the valid data becomes the key for improving the performance of the on-orbit remote sensing satellite. At present, research in the field is mainly realized on a deep learning server platform based on an nVidia display card on the ground, and the platform has higher power consumption and large volume and cannot be applied to an on-orbit remote sensing satellite.
Therefore, aiming at the problems, an image processing module is designed based on the FPGA platform, the detection of the optical remote sensing satellite image ship target is completed, invalid data are removed, and the satellite bandwidth utilization rate is improved.
Disclosure of Invention
The invention provides a ship detection device and method of an optical remote sensing satellite image, which solve the problem of detection of a ship target of the optical remote sensing satellite image.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows:
a ship detection device of an optical remote sensing satellite image comprises an image processing module, a storage module, a data input module, a bus protocol conversion module and a data output module;
the image processing module adopts an FPGA as a main control chip, the image processing module and the storage module can carry out bidirectional data transmission, the image processing modules work in parallel, and the storage module adopts a DDR3 specification type memory;
the data input module is used for receiving original image information through an CAMERALINK bus protocol, and the data output module is used for transmitting processed images through a PCIE bus protocol;
the bus protocol conversion module comprises AXI4 and CAMERALINK bus protocol conversion and AXI4 and PCIE bus protocol conversion, the AXI4 and CAMERALINK bus protocol conversion transmits image data of the data input module into the storage module, and the AXI4 and PCIE bus protocol conversion transmits the image data in the storage module into the data output module.
A ship detection method of optical remote sensing satellite image comprises the following steps,
(1) the number of channels is reduced: changing the number of input channels of the first-layer network of YOLO V3 to 1, and performing accumulation operation on the weight parameters of the corresponding positions of the original three input channels of the first-layer network to be used as new weight parameters of the first-layer network;
(2) weight parameter pretreatment, namely removing the BN normalization processing module in a network layer containing the BN normalization processing module, and carrying out pretreatment on each network parameter value based on the following formula to obtain a new parameter
Wherein, WnewFor the weight parameter after preprocessing, WoldAs a pre-processing weight parameter, Variance is the pre-processing batch Variance, BiasnewBias parameter after preprocessing, BiasoldThe Mean is a batch processing Mean value before pretreatment;
(3) resetting network weight parameters, rearranging the storage sequence of the network parameters, traversing and sequencing according to the sequence of the ordinal number of the convolutional layers, the total number of output channels of each convolutional layer, the total number of input channels of each convolutional layer, the total number of rows of convolutional kernels of each convolutional layer and the total number of columns of convolutional kernels of each convolutional layer, wherein the ordinal number of the network layers is from 0 to 106, and is 107 layers in total, and the specific mode is as follows:
the total number of output channels of each convolution layer is 64 as a greatest common divisor, parameters required by the first 64 output channels are firstly arranged, and then parameters required by the last 64 output channels are sequentially arranged until all parameters required by all output channels of the current network layer are rearranged; if the output channels are less than 64 channels, rearranging is completed at one time according to the number of the original output channels;
the total number of input channels of each convolutional layer is taken as a greatest common divisor, the parameters required by the first 8 input channels are firstly arranged, and then the parameters required by the last 8 input channels are arranged until all the parameters required by all the input channels of the current network layer are rearranged; if the number of input channels is less than 8, finishing the arrangement at one time according to the number of the original input channels;
the total number of rows of the convolution kernel of each layer takes 3 as the greatest common divisor, the first 3 rows of parameters are arranged firstly, and then the last three rows of parameters are arranged until all the rows of the convolution kernel of the current network layer are rearranged; if the number of rows of the convolution kernel is less than 3, finishing the arrangement at one time according to the number of rows of the original convolution kernel;
taking 3 as the greatest common divisor for the total number of the rows of the convolution kernel of each layer, arranging the first 3 rows of parameters, and then arranging the last three rows of parameters until all the rows of the convolution kernel of the current network layer are rearranged; if the number of the rows of the convolution kernels is less than 3, finishing the arrangement once according to the number of the rows of the original convolution kernels;
loading all the reordered network parameters into a storage module of a specified address in a DMA data transmission mode;
(4) the image processing module is used for carrying out ship detection: presetting the initial addresses of input and output images of each layer of network, calculating the sizes of the input and output images of each layer of network, and placing the convolution layer image in front of the SHORT layer of each layer in a storage module of a relatively low-order address of an input image cache address of the SHORT layer;
deleting the UPSAMPLE network layer, copying each output data value for 1 time in the convolution layer on the upper layer of the UPSAMPLE, placing the output data value behind each copied data, sequentially moving the original subsequent data of the copied data backwards by one bit, copying each row of data once again, placing the copied data in the position of the next row, and sequentially moving the original subsequent data of the copied data downwards by one row;
setting a network operation flag value, and when the network operates to the 80 th layer, starting to operate the 83 th layer network layer; when the network runs to layer 92, layer 95 network starts running.
Due to the adoption of the technical scheme, the invention has the technical progress that:
the FPGA chip is adopted to replace the prior GPU display card, so that the power consumption is greatly reduced, and the stability of remote work is guaranteed; the hardware characteristic of the FPGA is efficiently utilized by adopting a parallel mechanism, the memory capacity of a DDR3 memory is fully utilized, the information transmission efficiency among all modules of the system is increased by the conversion of various bus protocols, the parallel characteristic is exerted by the combination of an algorithm and the system, and the utilization rate of a broadband is improved.
Drawings
FIG. 1 is a schematic diagram of the calculation order among network layers according to the present invention;
FIG. 2 is a schematic view of the structure of the apparatus of the present invention.
Detailed Description
The following specific algorithm steps for realizing the optical remote sensing satellite image ship detection function based on the device are as follows:
and a hardware logic circuit is utilized to realize a YOLO V3 network on an FPGA chip, complete multi-channel data parallel processing and be used for detecting a ship target in the optical remote sensing satellite image. Firstly, a GOOGLE open source satellite image set is utilized to manufacture a satellite image set containing 1000 satellites, ship targets are labeled, a network is trained in an image processor, and finally trained model parameters are obtained.
Preprocessing the model parameters: firstly, since the optical remote sensing satellite image is a single-channel format-free original RAW image, in order to enable the optical remote sensing satellite image to be applied to a YOLO V3 network, the size of a first layer network of the YOLO V3 is changed to 416 × 1, wherein the first layer network 416 represents that the length of an input image is 416 pixels, the second layer network 416 represents that the width of the input image is 416 pixels, and the number of channels of the input image is 1; except for modifying the network size, the trained weight files are preprocessed, and the weight parameters of the same position of different channels in the first layer of network are accumulated to obtain new weight parameters.
Rearranging the storage sequence of the network parameters, and performing traversal sequencing according to the sequence of the ordinal number of the convolutional layers, the total number of output channels of each convolutional layer, the total number of input channels of each convolutional layer, the total number of rows of convolutional kernels of each convolutional layer and the total number of columns of convolutional kernels of each convolutional layer. The total number of output channels of each convolution layer is 64 as a greatest common divisor, parameters required by the first 64 output channels are firstly arranged, and then parameters required by the last 64 output channels are sequentially arranged until all parameters required by all output channels of the current network layer are rearranged; if the output channels are less than 64 channels, the rearrangement is completed at one time according to the number of the original output channels. The total number of input channels of each convolutional layer is taken as a greatest common divisor, the parameters required by the first 8 input channels are firstly arranged, and then the parameters required by the last 8 input channels are arranged until all the parameters required by all the input channels of the current network layer are rearranged; if the number of input channels is less than 8, the arrangement is completed at one time according to the number of the original input channels. The total number of rows of the convolution kernel of each layer takes 3 as the greatest common divisor, the first 3 rows of parameters are arranged firstly, and then the last three rows of parameters are arranged until all the rows of the convolution kernel of the current network layer are rearranged; and if the number of rows of the convolution kernel is less than 3, finishing the arrangement at one time according to the number of rows of the original convolution kernel. Taking 3 as the greatest common divisor for the total number of the rows of the convolution kernel of each layer, arranging the first 3 rows of parameters, and then arranging the last three rows of parameters until all the rows of the convolution kernel of the current network layer are rearranged; and if the number of the columns of the convolution kernel is less than 3, finishing the arrangement at one time according to the number of the rows of the original convolution kernel.
Except for all BN normalization processing modules in the YOLO V3 network, in the YOLO V3 network, except for the 81 st, 83 th and 105 th layers, all other convolution layers contain BN normalization processing modules. In order to reduce the time delay in the network forward inference process, the weight parameters are preprocessed in the weight file in advance. Processing the weight parameters in the convolutional layers except for the 81 st layer, the 83 th layer and the 105 th layer according to the following formula to obtain new parameters:
wherein, WnewFor the weight parameter after preprocessing, WoldAs a pre-processing weight parameter, Variance is the pre-processing batch Variance, BiasnewBias parameter after preprocessing, BiasoldFor the bias parameters before pre-processing, Mean is the batch Mean before pre-processing.
Fig. 2 shows the sequence of 107 network layers running with the time axis as the reference in the process of the operation of the image processor. Where Layerx, stands for the xth layer. The whole network starts to operate from the 0 th layer, and when the network operates to the 80 th layer, the 80 th layer and the 83 th layer simultaneously perform data processing and continue to operate the subsequent network layers. When the layer 92 is operated, the layer 92 and the layer 95 are simultaneously processed, and the subsequent network layer is continuously operated until the layer 106 is operated.
To ensure that data processing is performed at the 80 th and 83 th layers simultaneously, and at the 92 th and 95 th layers simultaneously. Data of 80 th to 82 th layers are required to be separately stored in the storage modules with independent addresses, data of 83 th to 94 th layers are required to be separately stored in the storage modules with independent addresses, and data of 95 th to 106 th layers are required to be separately stored in the storage modules with independent addresses, so that data of a network layer in front is prevented from overwriting data of a network layer behind.
Modifying the network, deleting the UPSAMPLE network layer, directly fusing in the convolution layer, setting a flag value, respectively adding a processing module when the network runs to the 84 th layer and the 85 th layer, copying each output value once, placing the copied data behind the output value, and sequentially shifting the original subsequent data of the copied data backwards by one bit. And copying each row of data once, placing the data at the position of the next row which is close to the data, sequentially moving the original subsequent data of the copied data downwards by one row, converting the modified YOLO V3 network into a logic circuit by using a hardware logic language, and operating the logic circuit in an FPGA chip.
Fig. 2 is a schematic structural diagram of a ship detection device for optical remote sensing satellite images, which includes an image processor, a DDR3 memory, a data input module, a data output module, an AXI4 and CAMERALINK bus protocol conversion module, an AXI4 and PCIE bus protocol conversion module, and 3 image processing modules for parallel processing, where the whole device can process 3 optical remote sensing satellite images simultaneously, and the image processing module has 2 channels inside, and supports the 80 th layer and the 83 th layer to perform computation simultaneously, and supports the 92 th layer and the 95 th layer to perform computation simultaneously.
The data input module mainly adopts CAMERALINK protocol and supports three modes, namely a BASE mode, a MEDIA mode and a FULL mode; the data in the three formats enter an AXI4 and CAMERALINK protocol conversion module from an CAMERALINK interface, then enter a DDR3 memory based on an AXI4 protocol, three optical remote sensing satellite images can be simultaneously received in an input module to be processed in parallel, 3 identical image processing modules perform parallel calculation simultaneously, and the AXI4 bus protocol and the DDR3 memory perform data interaction until ship target detection is completed.
And the output module reads the ship detection result in the DDR3 through an AXI4 bus protocol, and outputs the ship detection result through the AXI4 and the PCIE protocol conversion module by utilizing a PCIE bus protocol through the data output module. Performance pairs based on different hardware platforms and detection methods are shown in table 1, and it can be seen that the ship detection device of the optical remote sensing satellite image provided by the patent has better effects in two aspects of single-core processing capacity and running power consumption.
The following table shows the comparison between the performance of the ship detection device of the optical remote sensing satellite image and the performance of the YOLO V3 ship detection method and device based on nVidia 1050 Ti:
TABLE 1
As can be seen from the experimental results, the detection rate of the device is 3fps, and the running power consumption of the device is 4.3 w. The experiment basically verifies the feasibility of the ship detection device provided by the embodiment of the invention.
Claims (2)
1. A naval vessel detection device of optical remote sensing satellite image which characterized in that: the device comprises an image processing module, a storage module, a data input module, a bus protocol conversion module and a data output module;
the image processing module adopts an FPGA as a main control chip, the image processing module and the storage module can carry out bidirectional data transmission, the image processing modules work in parallel, and the storage module adopts a DDR3 specification type memory; the data input module is used for receiving original image information, and the data output module is used for transmitting a processed image through a PCIE bus protocol;
the bus protocol conversion module comprises AXI4 and CAMERALINK bus protocol conversion and AXI4 and PCIE bus protocol conversion, the AXI4 and CAMERALINK bus protocol conversion transmits image data of the data input module into the storage module, and the AXI4 and PCIE bus protocol conversion transmits the image data in the storage module into the data output module.
2. A ship detection method of an optical remote sensing satellite image is characterized in that: comprises the following steps of (a) carrying out,
(1) the number of channels is reduced, the number of input channels of the first-layer network of the YOLO V3 is changed into 1, and the weight parameters of the positions corresponding to the original three input channels of the first-layer network are accumulated to be used as new weight parameters of the first-layer network;
(2) weight parameter pretreatment, in the network layer containing BN normalization processing module, removing BN normalization processing module, and pretreating each network parameter based on the following formula to obtain new parameter
Wherein, WnewFor the weight parameter after preprocessing, WoldAs a pre-processing weight parameter, Variance is the pre-processing batch Variance, BiasnewBias parameter after preprocessing, BiasoldThe Mean is a batch processing Mean value before pretreatment;
(3) resetting network weight parameters, rearranging the storage sequence of the network parameters, and traversing and sequencing according to the sequence of the ordinal number of the convolutional layers, the total number of output channels of each convolutional layer, the total number of input channels of each convolutional layer, the total number of rows of convolutional kernels of each convolutional layer and the total number of columns of convolutional kernels of each convolutional layer, wherein the ordinal number of the network layers is from 0 to 106, and is 107 layers; the concrete mode is as follows:
the total number of output channels of each convolution layer is 64 as a greatest common divisor, parameters required by the first 64 output channels are firstly arranged, and then parameters required by the last 64 output channels are sequentially arranged until all parameters required by all output channels of the current network layer are rearranged; if the output channels are less than 64 channels, rearranging is completed at one time according to the number of the original output channels;
the total number of input channels of each convolutional layer is taken as a greatest common divisor, the parameters required by the first 8 input channels are firstly arranged, and then the parameters required by the last 8 input channels are arranged until all the parameters required by all the input channels of the current network layer are rearranged; if the number of input channels is less than 8, finishing the arrangement at one time according to the number of the original input channels;
the total number of rows of the convolution kernel of each layer takes 3 as the greatest common divisor, the first 3 rows of parameters are arranged firstly, and then the last three rows of parameters are arranged until all the rows of the convolution kernel of the current network layer are rearranged; if the number of rows of the convolution kernel is less than 3, finishing the arrangement at one time according to the number of rows of the original convolution kernel;
taking 3 as the greatest common divisor for the total number of the rows of the convolution kernel of each layer, arranging the first 3 rows of parameters, and then arranging the last three rows of parameters until all the rows of the convolution kernel of the current network layer are rearranged; if the number of the rows of the convolution kernels is less than 3, finishing the arrangement once according to the number of the rows of the original convolution kernels;
loading all the reordered network parameters into a storage module of a specified address in a DMA data transmission mode;
(4) the image processing module is used for carrying out ship detection: presetting the initial addresses of input and output images of each layer of network, calculating the sizes of the input and output images of each layer of network, and placing the convolution layer image in front of the SHORT layer of each layer in a storage module of a relatively low-order address of an input image cache address of the SHORT layer;
deleting the UPSAMPLE network layer, copying each output data value for 1 time in the convolution layer on the upper layer of the UPSAMPLE, placing the output data value behind each copied data, sequentially moving the original subsequent data of the copied data backwards by one bit, copying each row of data once again, placing the copied data in the position of the next row, and sequentially moving the original subsequent data of the copied data downwards by one row;
and setting a network operation flag value, wherein when the network operates to the 80 th layer, the 83 th layer network layer starts to operate, and when the network operates to the 92 th layer, the 95 th layer network starts to operate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011093587.7A CN112347857B (en) | 2020-10-14 | 2020-10-14 | Ship detection device and method of optical remote sensing satellite image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011093587.7A CN112347857B (en) | 2020-10-14 | 2020-10-14 | Ship detection device and method of optical remote sensing satellite image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112347857A true CN112347857A (en) | 2021-02-09 |
| CN112347857B CN112347857B (en) | 2022-04-19 |
Family
ID=74360771
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011093587.7A Active CN112347857B (en) | 2020-10-14 | 2020-10-14 | Ship detection device and method of optical remote sensing satellite image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112347857B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113920030A (en) * | 2021-10-25 | 2022-01-11 | 自然资源部国土卫星遥感应用中心 | Large-area high-fidelity satellite remote sensing image uniform color mosaic processing method and device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109429043A (en) * | 2017-08-27 | 2019-03-05 | 南京理工大学 | The acquisition system and method for traffic sign video image based on FPGA |
| CN110472529A (en) * | 2019-07-29 | 2019-11-19 | 深圳大学 | Target identification navigation methods and systems |
| CN110515889A (en) * | 2019-07-27 | 2019-11-29 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Embedded FPGA swarm intelligence computing platform hardware frame |
| CN110647988A (en) * | 2019-09-09 | 2020-01-03 | 博云视觉（北京）科技有限公司 | Accelerated calculation method of SSD (solid State disk) target detection convolutional neural network |
| CN111414994A (en) * | 2020-03-03 | 2020-07-14 | 哈尔滨工业大学 | FPGA-based Yolov3 network computing acceleration system and acceleration method thereof |
-
2020
- 2020-10-14 CN CN202011093587.7A patent/CN112347857B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109429043A (en) * | 2017-08-27 | 2019-03-05 | 南京理工大学 | The acquisition system and method for traffic sign video image based on FPGA |
| CN110515889A (en) * | 2019-07-27 | 2019-11-29 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Embedded FPGA swarm intelligence computing platform hardware frame |
| CN110472529A (en) * | 2019-07-29 | 2019-11-19 | 深圳大学 | Target identification navigation methods and systems |
| CN110647988A (en) * | 2019-09-09 | 2020-01-03 | 博云视觉（北京）科技有限公司 | Accelerated calculation method of SSD (solid State disk) target detection convolutional neural network |
| CN111414994A (en) * | 2020-03-03 | 2020-07-14 | 哈尔滨工业大学 | FPGA-based Yolov3 network computing acceleration system and acceleration method thereof |
Non-Patent Citations (2)
| Title |
|---|
| CHAN KIM ET AL.: "Implementation of Yolo-v2 Image Recognition and Other Testbenches for a CNN Accelerator", 《2019 IEEE 9TH INTERNATIONAL CONFERENCE ON CONSUMER ELECTRONICS》 * |
| 陈辰等: "基于Zynq7000 FPGA异构平台的YOLOv2加速器设计与实现", 《计算机科学与探索》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113920030A (en) * | 2021-10-25 | 2022-01-11 | 自然资源部国土卫星遥感应用中心 | Large-area high-fidelity satellite remote sensing image uniform color mosaic processing method and device |
| CN113920030B (en) * | 2021-10-25 | 2022-04-08 | 自然资源部国土卫星遥感应用中心 | Large-area high-fidelity satellite remote sensing image uniform color mosaic processing method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112347857B (en) | 2022-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108416327B (en) | Target detection method and device, computer equipment and readable storage medium | |
| US20210019594A1 (en) | Convolutional neural network accelerating device and method | |
| JP2019036298A (en) | Intelligent high bandwidth memory system and logic dies therefor | |
| CN111047008B (en) | Convolutional neural network accelerator and acceleration method | |
| US11314457B2 (en) | Data processing method for data format conversion, apparatus, device, and system, storage medium, and program product | |
| CN111460999A (en) | Low-altitude aerial image target tracking method based on FPGA | |
| US20220375186A1 (en) | Method and apparatus for generating bounding box, device and storage medium | |
| CN112347857B (en) | Ship detection device and method of optical remote sensing satellite image | |
| CN110717405B (en) | Face feature point positioning method, device, medium and electronic equipment | |
| CN113705803B (en) | Image hardware identification system and deployment method based on convolutional neural network | |
| US20240220541A1 (en) | Fpga-based method and system for accelerating graph construction | |
| CN110826683A (en) | Method and device for predicting user behavior occurrence probability | |
| CN111047037B (en) | Data processing method, device, equipment and storage medium | |
| WO2020118547A1 (en) | Fpga-based acceleration using opencl on fcl in robot motion planning | |
| CN116166185A (en) | Caching method, image transmission method, electronic device and storage medium | |
| CN110309877A (en) | A kind of quantization method, device, electronic equipment and the storage medium of feature diagram data | |
| Le et al. | An opencl-based sift accelerator for image features extraction on fpga in mobile edge computing environment | |
| CN111260046B (en) | Operation method, device and related product | |
| CN113128673B (en) | Data processing method, storage medium, neural network processor and electronic device | |
| CN104156954A (en) | Registration pre-processing system suitable for multispectral image compression | |
| CN113255538A (en) | FPGA-based infrared small and weak target detection tracking device and method | |
| US20230315478A1 (en) | Reducing power consumption by hardware accelerator during generation and transmission of machine learning inferences | |
| CN111860540B (en) | Neural network image feature extraction system based on FPGA | |
| CN111260070A (en) | Operation method, device and related product | |
| CN115391727B (en) | Calculation method, device and equipment of neural network model 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 |