[go: up one dir, main page]

CN114494872B - An embedded lightweight remote sensing target detection system - Google Patents

An embedded lightweight remote sensing target detection system Download PDF

Info

Publication number
CN114494872B
CN114494872B CN202210081517.2A CN202210081517A CN114494872B CN 114494872 B CN114494872 B CN 114494872B CN 202210081517 A CN202210081517 A CN 202210081517A CN 114494872 B CN114494872 B CN 114494872B
Authority
CN
China
Prior art keywords
target detection
model
convolution
remote sensing
residual
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.)
Active
Application number
CN202210081517.2A
Other languages
Chinese (zh)
Other versions
CN114494872A (en
Inventor
赵丹培
袁智超
苑博
姜志国
史振威
张浩鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beihang University
Original Assignee
Beihang University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beihang University filed Critical Beihang University
Priority to CN202210081517.2A priority Critical patent/CN114494872B/en
Publication of CN114494872A publication Critical patent/CN114494872A/en
Application granted granted Critical
Publication of CN114494872B publication Critical patent/CN114494872B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses an embedded lightweight remote sensing target detection system which comprises embedded equipment and a target detection model deployed on the embedded equipment, wherein the target detection model adopts a deep learning model and is deployed to the embedded equipment after a GPU platform is trained to obtain parameters, the target detection model comprises a structural re-parameterization module used for absorbing redundant residual errors and convolution structures in the deep learning model into a backbone network, and an improved residual error structure module used for realizing the same dimension of input and output through an improved residual error connection mode. The embedded light remote sensing target detection system is based on structural heavy parameterization, increases the use of residual connection, improves the residual connection mode when the input and output dimensions are different, reduces FLOPs of a model, improves the detection precision of the model, keeps the reasoning speed of the model, realizes the real-time performance of remote sensing target detection of embedded equipment under the limited computing capacity, and is convenient to improve the working efficiency.

Description

Embedded lightweight remote sensing target detection system
Technical Field
The invention relates to the fields of intelligent interpretation, deep learning and embedded systems of remote sensing images, in particular to an embedded lightweight remote sensing target detection system.
Background
At present, with the continuous development of high-resolution remote sensing image technology, remote sensing image data are widely applied to various fields, such as safety monitoring, resource exploration, disaster prevention and relief, military reconnaissance and the like. The acquisition of the remote sensing data has the advantages of breaking through time region limitation, large data volume, cost saving and the like. However, only a small part of specific targets in the large-format remote sensing image are often focused objects, so how to detect interesting targets from the large-format remote sensing image is a hot spot problem in the field of digital image processing. A large number of algorithms based on the deep neural network enable the target detection accuracy of the remote sensing image to reach a higher level. But high precision algorithm operation requires a lot of memory space and is time complex and thus highly dependent on high performance graphics processors. In the practical engineering application of remote sensing target detection, the current data processing is concentrated in a ground measurement and control station. On one hand, the processing platform is a more universal platform because a large amount of graphic processors are distributed, and on the other hand, the task also requires higher detection precision and running speed of the algorithm under the condition of limited computing capacity and memory space of the embedded system. Although the traditional target detection method has low calculation consumption and high running speed, the detection precision is far from meeting the requirements of tasks. Therefore, how to migrate the remote sensing target detection algorithm based on deep learning into the embedded system is a research field with application value.
In the field of image object detection, a traditional method is to extract local region features of an image by using an operator manually designed such as SIFT features and HOG features, and then classify the features by using a classifier such as a support vector machine. Along with the development of deep learning technology, a convolutional neural network-based method is continuously emerging, and the detection precision and the running speed are higher than those of the traditional method, so that the method is gradually the main stream and the front edge of research. However, the remote sensing image has its specificity with respect to the more widely studied natural image. In the remote sensing image, people often pay attention to sensitive targets such as airplanes, oil tanks, ships and the like, and also pay attention to sensitive ground objects such as airports, ports and the like. Therefore, in the target detection of the remote sensing image, the problems such as large target size difference, unbalanced target class labeling and the like are particularly prominent, and the problems are also the problems to be solved in the target detection field of the remote sensing image. On the other hand, the remote sensing image has the characteristic of real-time performance, and the image processing is required to be carried out on terminal equipment such as satellites, planes and the like in real time. In practical application, the equipment of the terminal has limited storage resources and computing power, and too complex and huge models cannot be laid, and the real-time requirement is difficult to achieve.
Disclosure of Invention
In view of the above problems, the present invention provides an embedded lightweight remote sensing target detection system, which can solve the problems that the current embedded system has limited computing power for target detection, cannot meet real-time performance, and has low working efficiency.
The embodiment of the invention provides an embedded lightweight remote sensing target detection system, which comprises embedded equipment and a target detection model deployed on the embedded equipment;
The target detection model adopts a deep learning model, is deployed to the embedded equipment after the GPU platform is trained to obtain parameters, and comprises the following components:
the structure re-parameterization module is used for absorbing redundant residual errors and convolution structures in the deep learning model into a backbone network;
And the improved residual structure module is used for realizing the same dimension of input and output through an improved residual connection mode.
Further, the structure of the target detection model is recompiled by using a TensorRT library based on C++, and the trained parameters are converted and deployed on embedded equipment to run.
Further, the structural reparameterization module is specifically configured to convert the residual connection of the 1×1 convolution and the identity mapping in the deep learning model into a 3×3 convolution form.
Further, converting the residual connection of the 1×1 convolution and identity mapping in the deep learning model into a3×3 convolution form, comprising:
the 1 x 1 convolution and identity mapping are converted to a 3 x 3 convolution, and the batch normalization operation is integrated into the convolution parameters by linear transformation, as shown in the following formula:
Wherein bn (·) is a batch normalization operation, M is a feature map extracted from the remote sensing image, μ, σ, γ, β is a batch normalization parameter;
for convolution with weight matrix W, let Then:
bn (m×w, μ, σ, γ, β) =m×w '+b'; batch normalization is absorbed into the convolution kernel.
Further, the improved residual structure module is specifically configured to directly intercept a feature map identical to the output dimension and add the feature map to the convolution result when the output dimension is smaller than the input dimension.
Furthermore, the improved residual structure module is further specifically configured to, when the output dimension is greater than the input dimension, perform the residual connection by copying and overlapping the input feature map along the dimension to obtain a feature map with the same dimension as the convolution output, and then add the feature map with the convolution output.
The technical scheme provided by the embodiment of the invention has the beneficial effects that at least:
The embedded light remote sensing target detection system comprises embedded equipment and a target detection model deployed on the embedded equipment, wherein the target detection model adopts a deep learning model and is deployed to the embedded equipment after a GPU platform is trained to obtain parameters, the target detection model comprises a structural re-parameterization module used for absorbing redundant residual errors and convolution structures in the deep learning model into a backbone network, and an improved residual error structure module used for realizing the same dimension of input and output in an improved residual error connection mode. According to the embedded light remote sensing target detection system, a deep learning model is designed in a light manner, and target detection is carried out on a remote sensing image by using the embedded system. The network design concept based on depth separable convolution and structure re-parameterization is explored, the influence of factors such as floating point calculation number FLOPs and memory access amount on the model reasoning speed is considered, based on the structure re-parameterization, the use of residual error connection is increased, the residual error connection mode when input and output dimensions are different is improved, FLOPs of the model is reduced, the model detection precision is improved, the reasoning speed of the model is kept, a lightweight target detection model with excellent comprehensive performance is obtained, the real-time performance and accuracy of remote sensing target detection are met under the condition of limited computing capacity of embedded equipment, and the work efficiency is improved.
Furthermore, in order to meet the real-time requirement of the on-board intelligent processing system, a lightweight target detection model can be deployed on the terminal embedded system by utilizing TensorRT, so that real-time embedded system target detection is realized.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings.
The technical scheme of the invention is further described in detail through the drawings and the embodiments.
Drawings
The accompanying drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate the invention and together with the embodiments of the invention, serve to explain the invention. In the drawings:
fig. 1 is a schematic diagram of an embedded lightweight remote sensing target detection system provided by an embodiment of the invention for detecting targets of an oil tank and an aircraft;
FIG. 2 is a diagram of an overall architecture of a target detection model according to an embodiment of the present invention;
Fig. 3a is a schematic diagram of a training phase residual structure according to an embodiment of the present invention;
FIG. 3b is a schematic diagram of a residual structure of a test phase according to an embodiment of the present invention;
FIG. 4a is a schematic diagram of a residual connection scheme with the same dimensions of input and output;
FIG. 4b is a schematic diagram of a residual connection mode when the input is larger than the output dimension according to the embodiment of the present invention;
FIG. 4c is a schematic diagram of a residual connection mode when input is smaller than output dimension according to the embodiment of the present invention;
fig. 5 is a schematic diagram of a deployment process TensorRT according to an embodiment of the present invention.
Detailed Description
Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
Referring to fig. 1, an embodiment of the present invention provides an embedded lightweight remote sensing target detection system, which is designed to lightweight a deep learning model, and uses an embedded system to perform target detection on a remote sensing image, such as identification of a storage equipment oil tank, an aircraft, a ship, and a dam. By exploring a network design concept based on depth separable convolution and structural re-parameterization, the influence of factors such as floating point calculation FLOPs and memory access amount on the model reasoning speed is considered, and a lightweight target detection model capable of detecting in real time in a terminal embedded system is designed.
The embedded lightweight remote sensing target detection system comprises embedded equipment and a target detection model deployed on the embedded equipment, wherein the target detection model adopts a deep learning model and is deployed to the embedded equipment after a GPU platform is trained to obtain parameters, and the target detection model comprises the following components:
the structure re-parameterization module is used for absorbing redundant residual errors and convolution structures in the deep learning model into a backbone network;
And the improved residual structure module is used for realizing the same dimension of input and output through an improved residual connection mode.
As shown in fig. 2, the overall architecture diagram of the complete target detection model is that the target detection model adopts a complex multi-convolution kernel multi-residual error mode to fully model the characteristics in a training stage, absorbs redundant residual errors and convolution structures into a backbone network through structural re-parameterization in a testing stage, and performs single-channel reasoning in the testing stage so as to improve the reasoning speed of the network.
The specific implementation mode and the structural flow are as follows, in the training stage, the network module adopts a multi-branch structure, simultaneously uses 3×3 convolution, 1×1 convolution and identity transformation to extract the characteristics, and adopts residual connection to improve the characteristic expression capability of the model. In the test stage, the structure of the training model is adjusted by utilizing the structure re-parameterization, and the 1X 1 convolution and the identity transformation operation are absorbed into the 3X 3 convolution under the condition of keeping the mathematical operation unchanged, so that the test model presents a single-channel information flow form. Such a structure can maintain a fast reasoning speed. Meanwhile, the residual connection mode when the input and output dimensions are different is modified, 1 multiplied by 1 convolution is omitted, the mode of stacking feature layers is adopted, and the model reasoning speed is improved. Aiming at the embedded system, the model structure is recompiled by utilizing a TensorRT library based on C++, and the trained parameters are converted, so that the running speed of the model structure on the embedded system is further improved, and real-time detection is realized.
In the embodiment, the target detection model is based on structural heavy parameterization, so that the use of residual connection is increased, the residual connection mode when the input and output dimensions are different is improved, FLOPs of the model is reduced, the accuracy of the model for detecting remote sensing targets is improved, and the reasoning speed of the model is kept. In the specific implementation, for example, in order to meet the real-time requirement of the on-board intelligent processing system, the target detection model can be deployed on the terminal embedded system by utilizing TensorRT, so that the real-time target detection of the embedded system is realized.
The technical scheme of the invention is specifically described by three aspects:
1. lightweight design of model based on structural heavy parameterization
In practical application, the reasoning speed of the model is also influenced by factors such as memory access quantity, hardware optimization and the like. In the most popular residual structure of the convolutional neural network at present, the instantaneous memory access amount of the model is larger than that of the channel model because intermediate results need to be temporarily reserved in the memory when calculating the residual. In addition, in the current mainstream neural network framework, such as PyTorch, optimization for convolution results in broken convolution structure in depth separable convolution, which is rather unfavorable for parallel execution of convolution calculation, resulting in a reduction of operation speed. Because the network structure and the parameters have a one-to-one correspondence, if the parameters of multiple layers can be combined, the network structure can respond to the change, the macroscopic process of the structural reparameterization is to convert the residual connection of the 1X 1 convolution and the identity mapping into a form of 3X 3 convolution, then combine 3 branches into one to realize a model of a single branch, and the 1X 1 convolution and the identity mapping can be easily regarded as the 3X 3 convolution at the microscopic level, and then integrate the batch normalization operation into the convolution parameters through linear transformation to enhance the generalization of the model, wherein the batch normalization operation is shown in the following formula:
Wherein bn (·) is a batch normalization operation, M is a feature map extracted from the remote sensing image, μ, σ, γ, β is a batch normalization parameter;
for convolution with weight matrix W, let Then:
bn (m×w, μ, σ, γ, β) =m×w '+b'. After the batch normalization is absorbed into the convolution kernel, the residual connection can be fused into a branch by simple addition.
2. Improving residual structure to promote inference speed
Starting from ResNet with the use of residual connections in deep neural networks for the first time, this architecture is widely used in various network architectures. The residual structure can reserve the gradient of the shallow layer, and prevent the gradient from disappearing in the deep network training. Meanwhile, more nonlinear transformation can be introduced through the ReLU activation function, so that the expression capability of the characteristics is improved. However, the residual structure used in the structure re-parameterization is a "pseudo residual" structure without an implicit feature layer, and if the network depth is too large, gradient disappearance may still occur in the gradient back propagation process.
Referring to fig. 3a, the input is connected with the input by residual after passing through two parallel structures, and then the output is obtained through a ReLU activation function. In the reasoning phase model structure shown in fig. 3b, the 1 x1 convolution and identity residual connection of the parallel structure are absorbed by the structure reparameterization into a 3 x 3 convolution, thus becoming a single-branch structure. But the residual connection of the outermost layer is preserved and added to the convolved result before activating the function. This structure preserves the residual connection during reasoning and the ability of the network to resist gradient extinction through ReLU nonlinear changes after residual connection.
In addition, the improved residual connection mode can also cope with the residual connection requirements of different input and output dimensions. As shown in fig. 4a-4c, the manner in which the residuals are connected is also different for different input-output dimensions. Fig. 4a shows the case where the input and output dimensions are the same. In this case, the residual is connected in the same manner as in most networks, and the residual is mapped by identity mapping and added to the convolved feature map. For the case shown in fig. 4b where the output dimension is smaller than the input dimension (which is common in the downsampling stage), the characteristic dimension of the residual branch is not changed by 1×1 convolution, but the characteristic map which is the same as the output dimension is directly truncated and added to the convolution result. For the case shown in fig. 4c where the output dimension is greater than the input dimension (usually in the up-sampling stage), the residual connection is completed by copying and overlapping the input feature map along the dimension to obtain a feature map with the same dimension as the convolution output and adding the feature map to the convolution output. For the case of different input-output resolutions after downsampling, the maximum pooling is performed in the residual branch first, and then the identity mapping, truncation, or stacking in fig. 4a-4c is performed. Through the residual connection mode, the residual branch does not consume any FLOPs, and meanwhile, the running speed is improved through reducing parallel convolution calculation.
3. Deployment of lightweight models in embedded systems
Current neural network models mostly rely on a deep learning framework, such as TensorFlow, pyTorch, for high performance servers. On one hand, the frames are based on Python language, are simpler and more friendly in development and design, but have lower running efficiency, and on the other hand, the frames are developed based on CUDA, are optimized for a GPU platform and are not suitable for the calculation mode of an embedded system, so that the running speed on the embedded system is not good. Therefore, the embodiment of the invention also utilizes TensorRT libraries specially designed for the quantization of the embedded system to re-write the model structure, so that the model structure can be deployed on the embedded platform, and the reasoning speed of the model is further improved.
Because the computing capability of the embedded platform is poor and training cannot be performed, the embedded platform needs to be deployed to the embedded platform for reasoning after training on the GPU in advance to obtain parameters. Using the TensorRT library, the PyTorch model for the GPU may be converted to a model for the embedded platform, the deployment process is shown in fig. 5. Firstly, constructing PyTorch model at a server end by utilizing a PyTorch framework based on Python according to a designed model structure, inputting training images for iterative training, and finally obtaining model parameters. The model is then re-built using the C++ based TensorRT library, whose structure needs to be consistent with the PyTorch model to ensure correctness of the parameter loading. After the embedded terminal compiles TensorRT the model and loads the model parameters, a model engine for reasoning can be obtained. As shown in FIG. 5, the test image is input to the inference engine to infer in the embedded system, thereby outputting accurate detection of the target in the test image.
For example, the current on-board remote sensing processing system has poor large-scale operation capability and intelligent degree, and the embedded light-weight remote sensing target detection system provided by the embodiment of the invention can be used for designing a light-weight target detection model. According to the embodiment of the invention, the remote sensing target detection model is designed in a light-weight manner, the structure heavy parameterization module designs a light-weight target detection network based on the structure heavy parameterization method, and the residual structure module is improved to analyze the influence of residual branches on the reasoning speed, and improve the residual connection mode when the input dimension and the output dimension are different, so that FLOPs of the model is reduced, the model detection precision is improved, the reasoning speed of the model is kept, the real-time performance and the accuracy of the remote sensing target detection are met under the condition of limited computing capacity of the embedded equipment, and the work efficiency is improved.
It will be apparent to those skilled in the art that various modifications and variations can be made to the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.

Claims (1)

1.一种嵌入式轻量化遥感目标检测系统,其特征在于,包括:嵌入式设备以及部署在所述嵌入式设备上的目标检测模型;所述嵌入式设备为星上遥感处理系统;1. An embedded lightweight remote sensing target detection system, comprising: an embedded device and a target detection model deployed on the embedded device; the embedded device is an on-board remote sensing processing system; 1)所述目标检测模型采用深度学习模型,并在GPU平台进行训练得到参数后部署到所述嵌入式设备;该目标检测模型包括:1) The target detection model adopts a deep learning model and is deployed to the embedded device after training on a GPU platform to obtain parameters; the target detection model includes: 结构重参数化模块,用于将深度学习模型中多余的残差和卷积结构吸收到主干网络中;The structure reparameterization module is used to absorb the redundant residual and convolutional structures in the deep learning model into the backbone network; 改进残差结构模块,用于通过改进的残差连接方式,实现输入输出相同维度;Improved residual structure module, used to achieve the same dimension of input and output through improved residual connection method; 其中,所述改进残差结构模块,具体用于当输出维度小于输入维度时,直接截取和输出维度相同的特征图与卷积结果相加;还具体用于当输出维度大于输入维度时,通过将输入特征图沿维度复制叠加,得到和卷积输出同等维度的特征图后再与卷积输出相加的方式完成残差连接;The improved residual structure module is specifically used for directly intercepting a feature map with the same dimension as the output and adding it to the convolution result when the output dimension is smaller than the input dimension; and is also specifically used for completing the residual connection by copying and superimposing the input feature map along the dimension to obtain a feature map with the same dimension as the convolution output and then adding it to the convolution output when the output dimension is larger than the input dimension; 2)利用基于C++的TensorRT库对所述目标检测模型的结构进行重新编译,并将训练好的参数进行转换,部署在嵌入式设备上运行;2) Recompile the structure of the target detection model using the C++-based TensorRT library, convert the trained parameters, and deploy them on an embedded device; 所述目标检测模型在训练阶段,网络模块采用多分支结构,同时使用3×3卷积、1×1卷积和恒等变换对特征进行提取,并采用残差连接提升模型的特征表达能力,输入通过两个并行结构后与输入进行残差连接,然后经过ReLU激活函数得到输出;在测试阶段,利用结构重参数化,对训练模型的结构进行调整,在保持数学运算不变的条件下将1×1卷积和恒等变换操作吸收到3×3卷积当中,使测试模型呈现单通道信息流;最外层的残差连接得以保留,并在激活函数之前与卷积后的结果相加;In the training phase of the target detection model, the network module adopts a multi-branch structure, and uses 3×3 convolution, 1×1 convolution and identity transformation to extract features, and uses residual connection to improve the feature expression ability of the model. The input passes through two parallel structures and is residually connected with the input, and then the output is obtained through the ReLU activation function. In the test phase, the structure of the training model is adjusted by structural reparameterization, and the 1×1 convolution and identity transformation operations are absorbed into the 3×3 convolution while keeping the mathematical operation unchanged, so that the test model presents a single-channel information flow; the outermost residual connection is retained and added to the result after convolution before the activation function; 通过线性变换将批归一化操作整合到卷积参数当中,批归一化操作如下式所示:The batch normalization operation is integrated into the convolution parameters through linear transformation. The batch normalization operation is shown as follows: 其中,bn(·)为批归一化操作;M为遥感图像提取的特征图;μ,σ,γ,β为批归一化参数;Where bn(·) is the batch normalization operation; M is the feature map extracted from the remote sensing image; μ, σ, γ, β are batch normalization parameters; 对于权重矩阵为W的卷积,令则:For the convolution with weight matrix W, let but: bn(M*W,μ,σ,γ,β)=M*W'+b';将批归一化吸收到卷积核当中。bn(M*W,μ,σ,γ,β)=M*W'+b'; absorb batch normalization into the convolution kernel.
CN202210081517.2A 2022-01-24 2022-01-24 An embedded lightweight remote sensing target detection system Active CN114494872B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210081517.2A CN114494872B (en) 2022-01-24 2022-01-24 An embedded lightweight remote sensing target detection system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210081517.2A CN114494872B (en) 2022-01-24 2022-01-24 An embedded lightweight remote sensing target detection system

Publications (2)

Publication Number Publication Date
CN114494872A CN114494872A (en) 2022-05-13
CN114494872B true CN114494872B (en) 2024-12-13

Family

ID=81474487

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210081517.2A Active CN114494872B (en) 2022-01-24 2022-01-24 An embedded lightweight remote sensing target detection system

Country Status (1)

Country Link
CN (1) CN114494872B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115512324B (en) * 2022-10-13 2024-07-12 中国矿业大学 Pavement disease detection method based on edge symmetrical filling and large receptive field
CN115546659B (en) * 2022-10-17 2025-05-13 江苏思远集成电路与智能技术研究院有限公司 A method for airport target damage detection based on satellite images
CN116597301B (en) * 2023-04-18 2025-09-12 北京科技大学 An urban remote sensing embedded multi-target detection and positioning method and system

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109190491A (en) * 2018-08-08 2019-01-11 上海海洋大学 Residual error convolutional neural networks SAR image sea ice classification method
CN113762479A (en) * 2021-09-10 2021-12-07 深圳朴生智能科技有限公司 A kind of neural network optimization method and device

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110599401A (en) * 2019-08-19 2019-12-20 中国科学院电子学研究所 Remote sensing image super-resolution reconstruction method, processing device and readable storage medium
CN113297843B (en) * 2020-02-24 2023-01-13 华为技术有限公司 Reference resolution method and device and electronic equipment
CN113408423B (en) * 2021-06-21 2023-09-05 西安工业大学 Aquatic product target real-time detection method suitable for TX2 embedded platform

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109190491A (en) * 2018-08-08 2019-01-11 上海海洋大学 Residual error convolutional neural networks SAR image sea ice classification method
CN113762479A (en) * 2021-09-10 2021-12-07 深圳朴生智能科技有限公司 A kind of neural network optimization method and device

Also Published As

Publication number Publication date
CN114494872A (en) 2022-05-13

Similar Documents

Publication Publication Date Title
CN114494872B (en) An embedded lightweight remote sensing target detection system
Cheng et al. Feature enhancement network for object detection in optical remote sensing images
Zhang et al. PVT: Point‐voxel transformer for point cloud learning
US11341369B2 (en) Distributed batch normalization using partial populations
Yang et al. An effective and lightweight hybrid network for object detection in remote sensing images
Yin et al. An enhanced lightweight convolutional neural network for ship detection in maritime surveillance system
CN112445497B (en) Remote sensing image processing system based on plug-in extensible architecture
Ieracitano et al. An explainable embedded neural system for on-board ship detection from optical satellite imagery
Xu et al. On-board ship detection in SAR images based on L-YOLO
CN117786848A (en) A GPU-based efficient simulation method, storage medium and computer for spacecraft systems
Zhang et al. DSP-based traffic target detection for intelligent transportation
Yu et al. Optimizing FPGA-based convolutional encoder-decoder architecture for semantic segmentation
Liu et al. OAB‐YOLOv5: One‐Anchor‐Based YOLOv5 for Rotated Object Detection in Remote Sensing Images
Qin et al. A CNN hardware accelerator designed for YOLO algorithm based on RISC-V SoC
Du et al. Adaptive anchor-based detector with constrained RIRConv for oriented vehicles in SAR images
Yu et al. Spaceborne synthetic aperture radar imaging mapping methodology based on FPGA‐DSP hybrid heterogeneous architecture
Xie et al. Lightweight and anchor-free frame detection strategy based on improved CenterNet for multiscale ships in SAR images
Hu et al. MFF‐YOLOv8: Small Object Detection Based on Multi‐Scale Feature Fusion for UAV Remote Sensing Images
Gonzalez-Arjona et al. Analysis and implementation of space avionics co-processor for deep learning acceleration
Gao et al. The Aeroplane and Undercarriage Detection Based on Attention Mechanism and Multi‐Scale Features Processing
Tiwari et al. A novel YOLOv4‐modified approach for efficient object detection in satellite imagery
Zhang et al. Model-architecture Codesign for High-performance and Energy-efficient SAR ATR on FPGA
Zhang et al. Application of Lightweight Target Detection Algorithm Based on YOLOv8 for Police Intelligent Moving Targets
Lu et al. Lightweight One-Stage Ship Detection Based on SAR Images for Embedded Platform
Jiang et al. Oriented ship target detection based on improved CenterNet in synthetic aperture radar images

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