CN116416468B - SAR target detection method based on neural architecture search - Google Patents
SAR target detection method based on neural architecture search Download PDFInfo
- Publication number
- CN116416468B CN116416468B CN202310376226.0A CN202310376226A CN116416468B CN 116416468 B CN116416468 B CN 116416468B CN 202310376226 A CN202310376226 A CN 202310376226A CN 116416468 B CN116416468 B CN 116416468B
- Authority
- CN
- China
- Prior art keywords
- search
- target detection
- calculation
- layer
- sar
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a SAR target detection method based on neural architecture search, which relates to the technical field of target detection and comprises the following steps: inputting SAR image data into an SAR target detection model based on neural architecture search, and outputting an SAR target detection result; the method comprises the steps of extracting features of SAR image data based on a plurality of search layers of a backbone network searcher; integrating the extracted features based on a Neck module; and predicting the integrated features based on the Head module to obtain SAR target detection results. The invention provides a SAR target detection model based on neural architecture search, which can automatically design a high-precision and lightweight model suitable for the current SAR target detection task, and greatly reduces the artificial design burden.
Description
Technical Field
The invention relates to the technical field of target detection, in particular to a SAR target detection method based on neural architecture search.
Background
Synthetic Aperture Radar (SAR) is an active earth observation system that achieves earth observation by emitting microwaves and collecting reflected information. Due to the unique imaging mechanism of SAR, the SAR has stronger observation capability compared with optical remote sensing, and can realize all-day and all-weather observation. Meanwhile, the SAR also has certain penetrating capacity and very strong anti-interference and camouflage identification capacities. Based on the above advantages, SAR systems are widely used in civilian and military applications.
When the SAR image is manually processed, massive data information and image characteristics different from optical imaging greatly influence the processing speed and accuracy. Meanwhile, in many SAR image processing tasks, particularly in military tasks, high requirements are put on the speed and accuracy of SAR image recognition. In this context, researchers have begun to study SAR image target detection algorithms, hopefully replacing manual work to automatically detect targets contained in SAR images.
In recent years, with the increasing computational power of hardware devices, a convolutional neural network (convolutional neural network, CNN) model based on deep learning starts to gradually bring about a brand-new angle in an optical image target detection task. The CNN model is a deep learning model capable of effectively extracting two-dimensional image data features, and has the advantages of being capable of replacing manual feature design, automatically performing active learning and extraction on the image features, and meanwhile, compared with a traditional image processing algorithm, the CNN model can form a more complex image recognition function due to the fact that the CNN network comprises multiple layers of convolution layers, and therefore image target features to be learned can be fitted better. CNN networks have achieved very good results in the task of target detection on optical images. With reference to the excellent performance of the CNN network on the optical image detection task, the CNN network is gradually introduced in the research of SAR image target detection so as to solve the problems of low detection rate and poor robustness in the traditional image processing method. The main idea is to improve the optical image detection CNN model and combine the characteristic properties of SAR images to obtain a new SAR image target detection CNN model.
Although CNN network has achieved good effect on SAR image target detection task, researchers have met new difficulties, and the main reasons for these difficulties are that the neural network has complex structure, large parameter quantity and large design difficulty. The existing SAR image target detection algorithm based on the CNN mainly faces the following difficulties:
(1) The existing deep neural network needs to manually design a network architecture according to expert experience, and the deep neural network meeting SAR target detection application is constructed by continuously testing the wrong tuning parameters. This process is cumbersome, time consuming, and is typically a locally optimal model.
(2) The deployment of the hardware platform under the limited computing resource condition facing the edge application needs to perform a customized model compression strategy aiming at the computing resource of the hardware platform, and is generally difficult to ensure the balance and unification of the precision and the computing quantity.
Disclosure of Invention
The invention provides a SAR target detection method based on neural architecture search, which can automatically find a calculation module suitable for each convolution calculation layer in an SAR target detection CNN network from a search space (a set of all candidate calculation modules) under the condition of considering the model detection precision and indexes such as calculation complexity, parameter quantity, reasoning speed and the like of a model through an automatic structure search algorithm, and finally obtain an SAR target detection model network structure which has optimal comprehensive performance and is suitable for being deployed on edge segment hardware equipment with limited calculation resources.
The invention provides a SAR target detection method based on neural architecture search, which comprises the following steps:
SAR image data are obtained;
inputting SAR image data into an SAR target detection model based on neural architecture search, and outputting an SAR target detection result;
obtaining SAR target type and position information according to the SAR target detection result;
the step of inputting the SAR image data into the SAR target detection model based on the neural architecture search and outputting the SAR target detection result comprises the following steps:
performing feature extraction on SAR image data based on a plurality of search layers of a backbone network searcher;
integrating the extracted features based on a Neck module;
and predicting the integrated features based on the Head module to obtain SAR target detection results.
Preferably, each search layer corresponds to a corresponding convolution calculation layer of the CNN network, and a plurality of candidate calculation modules are added in parallel to each search layer.
Preferably, the feature extraction is performed on the SAR image data by the plurality of search layers based on the backbone network searcher, and the method specifically comprises the following steps:
computing a Block for each candidate of a plurality of search layers i Corresponding coefficient theta is distributed i ；
Calculating each candidate calculation module Block through Gumbel Softmax function i Corresponding actual weighted calculation coefficients
Block is calculated for each candidate of the current search layer 1～i And the corresponding actual weighted calculation coefficientMultiplying and adding to obtain the calculation result of the current search layer;
and transmitting the calculation result of the current search layer to the next search layer, transmitting the calculation result of the current search layer to all candidate calculation blocks by the next search layer for parallel calculation, and obtaining a final output characteristic diagram of the backbone network searcher after calculation of all the search layers.
Preferably, the calculation formula of the gummel Softmax function is as follows:
in the formula ,gi Gumbel (0, 1) is random Gumbel noise, τ is a temperature parameter, k is the total number of calculation modules of the search layer, and exp is an exponential function.
Preferably, the calculation formula of the search layer is as follows:
where LayerOutput is the output of the search layer.
Preferably, a loss function based on neural architecture search is constructed, and the weight parameters in the backbone network searcher are iteratively updated through the loss function.
Preferably, the formula of the loss function is as follows:
wherein ,
in precision loss For the current precision loss value of the model, the latency is loss To model reasoning time loss values, flow loss For model calculation loss, param loss The layer is a calculation layer, alpha is a parameter for adjusting the calculated loss, beta is a parameter for adjusting the loss value, and l is a parameter for adjusting the loss value i and pi Respectively is Block i FLPs (m) is the calculated amount in the current search network, T is the upper limit of the calculated amount, and gamma is a number much larger than 1.
Preferably, after the iterative updating of the weight parameters in the backbone network searcher is completed, selecting a value θ with the largest coefficient θ of each layer of search layer max Corresponding candidate calculation moduleAs a result of the search of the current search layer, the CNN convolution calculation layer corresponding to the current search layer will employ a candidate calculation module +.>As a convolution calculation method of this layer.
Compared with the prior art, the invention has the beneficial effects that:
the invention designs a SAR target detection model based on neural architecture search, a backbone network searcher based on neural architecture search is constructed, the backbone network searcher is used for searching the optimal backbone structure of a target detection network, and the optimal backbone network is combined with a Neck module and a Head module of the target detection network to finally obtain the SAR target detection model. By applying the method provided by the invention, a high-precision and lightweight model suitable for the current SAR target detection task can be automatically designed, and the artificial design burden is greatly reduced.
Drawings
In order to more clearly illustrate the embodiments of the invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a SAR target detection method based on neural architecture search of the present invention;
FIG. 2 is a schematic structural diagram of a SAR target detection model based on neural architecture search according to the present invention;
FIG. 3 is a schematic diagram of a backbone structure searcher of the present invention;
FIG. 4 is a schematic diagram of the internal computing architecture of the search layer of the present invention;
FIG. 5 is a schematic diagram showing the visual comparison of the detection effect of the experimental result and the current other SAR target detection models;
fig. 6 is a visual display diagram of the detection effect of the experimental result of the invention on a large scene SAR picture.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Referring to fig. 1, neural architecture searches (Neural Architecture Search, NAS) can efficiently evaluate structural performance and find the best network structure in a large set of neural network structures. The structural layout of the CNN obtained by searching by the NAS method is often quite different from that of the CNN which is designed manually, but the performance effect which is not inferior to that of the CNN which is designed manually can be still obtained. Therefore, in order to solve the problems faced by the current SAR image target detection algorithm based on deep learning CNN network, the invention provides an SAR target detection method based on neural architecture search. The method specifically comprises the following steps:
the first step: SAR image data is acquired.
And a second step of: and constructing a SAR target detection model SARNas based on neural architecture search.
Referring to fig. 2, the general flow of SARNas is similar to that of a general object detection framework, and is mainly divided into 3 parts including SARNas-BS (SARNas-backbone searcher), neg and Head parts, and SARMI-Loss. After SAR image data is input into SARNas, firstly, a trunk searching part SARNas-BS performs feature extraction, then, the extracted features are input into a subsequent Neck and Head part to be converted into model training results, finally, the model training results enter into a Loss function calculating part SARNas-Loss, and Loss values are transmitted to the SARNas-BS according to a Loss function SARMI-Loss (SAR multi-indication Loss) to guide the searching of a trunk network structure.
SARNas, like the common target detection network model, can be trained directly with the target detection data set. Therefore, the additional work of searching or manufacturing the classification data set related to the SAR image by oneself can be omitted, the SAR target detection data set can be directly used for training, meanwhile, the trunk which is finally sampled and obtained by SARNas is also suitable for the trunk structure of the target detection Neck and Head module which is adopted at present, and the searched trunk (Backbone) network can be directly combined with the follow-up modules to form a target detection network to execute the SAR image target detection training task without being used as additional modification on the structure.
A super network SARNas-BS responsible for performing the trunk NAS search task is constructed, and combined with the Neck and Head parts in the target detection network to form a search super network for trunk structure search. In order to reduce the NAS searching operation amount and searching time, the data set is directly transmitted to train the super network in the searching process, and the super network is sampled after the training is finished, so that the optimal sub network is obtained.
Referring to fig. 3, the backbone structure searcher SARNas-BS is composed of a plurality ofA stack of search layers (searchlayers) results, each corresponding to a respective convolutional computational layer of the CNN network, with multiple candidate computational blocks being added in parallel per layer. In each search layer, the present invention assigns a coefficient θ to each candidate calculation module. Each candidate Block i Corresponding to different coefficients theta i Is assigned to distinguish the importance of different candidate computing blocks in the current search layer. In the process of super-network training, back propagation calculation is carried out according to a Loss value obtained by calculation of a Loss function SARMI-Loss, and iterative updating is carried out on the value of the coefficient theta in the super-network.
Referring to fig. 4, in the calculation process of the search layer, after the calculation result of the previous layer is transferred to the current search layer, the current search layer transfers the input to all candidate calculation blocks in the search layer for parallel calculation. Meanwhile, each candidate calculation Block is calculated through Gumbel Softmax function i Corresponding coefficient theta i To achieve the effect of normalizing the parameters. Obtaining an actual weighted calculation coefficient corresponding to each candidate calculation module in the search layer through calculation of Gumbel SoftmaxGumbel Softmax is calculated as follows:
in the formula (1), g i Gumbel (0, 1) is a random Gumbel noise, τ is called a temperature parameter, used to adjust the p- θ i Obtained after re-parameterizationThe density among the coefficients, k, is the total number of calculation modules of the search layer.
By introducing Gumbel Softmax, the coefficient θ can be scaled i Back-propagating and updating them. Finally, each candidate computing Block is blocked 1～i And the corresponding actual weighted calculation coefficientMultiplied and added. And then, obtaining a final output characteristic diagram of the search layer. The calculation method of the search layer has the following general formula:
as the super network training proceeds, the calculation coefficient θ of each search layer is updated according to the loss function. After the super network training is finished, the coefficient theta finally determined in the search layer i To distinguish the importance of each candidate calculation block to the search layer in the calculation process, and the coefficient theta corresponding to the candidate calculation block contained in each search layer is extracted. At this time, SARNas will select the value θ with the largest coefficient θ for each search layer max Corresponding candidate calculation moduleAs search results for the search layer. The search result represents the convolution calculation mode of the CNN convolution calculation layer corresponding to the current search layer, which takes the calculation module as the layer.
For the SAR target detection algorithm model of actual deployment, besides good detection precision, the comprehensive performance of the model needs to be optimized by combining the performance of actual hardware equipment. Therefore, when the search algorithm is designed, the accuracy is taken as an important index for evaluating and searching the network model, and a plurality of main attributes of the CNN model, such as reasoning speed, parameter quantity and calculated quantity (the measurement unit is FLPs), are introduced to be taken as reference indexes for designing the SARNas search direction.
According to the deep learning technical principle and the training characteristics of the neural network model, the invention redesigns the Loss function, adds the network model precision, the reasoning speed, the parameter quantity and the calculated quantity into the calculation of the Loss function, thereby providing a new Loss function SARMI-Loss (SAR Multi-Index Loss), and guides the searching direction of SARNas by carrying out back propagation on the Loss value obtained by the calculation of the Loss function in the super-network training process. SARMI-Loss formula is as follows:
wherein precision is loss And representing the current precision loss value of the model, and reflecting the detection precision of the current search model. labensy of loss The model reasoning time loss value is represented, and the reasoning time of the current search model is reflected. flow-type loss Representing the computational loss of the current search model, which reflects the computational loss of the current search model. Param loss The parameter loss value representing the current search model reflects the size of the model parameter.
For each loss value in the loss function, taking the reasoning time loss value as an example, the calculation thought of the design of the invention is as follows: multiplying the calculation modules of each layer by the module coefficients thereof respectively and accumulating to obtain the reasoning time loss of the search layer, accumulating the reasoning time loss of all the search layers, and finally obtaining the reasoning time loss latency of the current search model loss . Calculation loss flow loss And parameter quantity loss param loss The same calculation idea design is followed. The calculation formula is listed below:
the parameter beta is used for adjusting the magnitude of the loss value, so that the magnitude of the loss value of each part is kept coordinated in training, and unbalance of the NAS searching direction caused by overlarge difference of the loss values is avoided.Representing corresponding candidate computing modules Block i Weight coefficient of (c) in the above-mentioned formula (c). l (L) i and pi Respectively represent Block i The delay time and the number of parameters.
In the actual landing of the SAR image algorithm, the calculation power of target hardware equipment is often limited, and the calculation amount of the model algorithm needs to be optimally designed. To achieve automation, architecture searches are performed within the allowable range of computing power according to the conditions of hardware devices. Thus, for flows in the Loss function SARMI-Loss loss The invention provides a specific implementation idea, namely, flow loss The loss term is set as a piecewise function, and a new limiting coefficient is added in the piecewise function, and the formula is as follows:
wherein FLPs (m) represents the calculated amount (Floating Point Operations, FLPs, floating point operands) in the network currently searched, the larger the FLPs value is, the larger the calculated amount representing the algorithm is, the higher the calculation complexity is, and T represents the upper limit of the calculated amount. Alpha is a parameter whose value corresponds to the flow of the currently searched model for adjusting the magnitude of the computation loss at this time. When FLPs (m) > T, the value of α is γ. In practical applications, the gamma value can be adjusted as the case may be, and the gamma value is typically set to a number much greater than 1. Setting the value of alpha in a segmentation way, so that the search direction is constrained according to the situation in the NAS search process: when the network operand exceeds the expected design upper limit, focusing the searching on the model combination for reducing the model operand; when the network operand is within the expected design upper limit, the search algorithm will focus more on other performance information of the model.
And a third step of: and inputting SAR image data into an SAR target detection model based on neural architecture search to obtain a target detection result.
Referring to fig. 5 and 6, in order to verify the effectiveness of the present invention, the SARNas algorithm proposed by the present invention is experimentally verified on two current mainstream SAR image target detection data sets SSDD and HRSID. SAR target detection based on the Yolo series algorithm is selected as the algorithm improvement base model.
Currently, YOLOX and YOLOX 5 are two most representative deep learning target detectors with excellent comprehensive performance in the YOLO series, and are widely applied in engineering application. Therefore, the verification experiment of the invention selects YOLOX as an anchor-free detection frame representative model and YOLOV5 as an anchor-free detection frame representative model, and respectively uses SARNas algorithm to perform design optimization work of an automatic deep learning neural network model. According to experimental results, the model finally and automatically designed by the invention can effectively balance the factors influencing the actual hardware deployment performance of the model, such as reasoning time, parameter, calculated quantity and the like of the search model while ensuring the effective recognition precision of the SAR target detection task. Experimental specific data are shown in table 1 and table 2 below, and corresponding model visual recognition results are shown in fig. 5 and 6.
Table 1 results of comparison of detection effects with the currently up-to-date target detection network
Table 2 results of comparison of detection effects with other SAR target detection algorithms
By applying the technical algorithm of the invention, a high-precision and lightweight model suitable for the current SAR target detection task can be automatically designed, and the artificial design burden is greatly reduced.
The technical algorithm provided by the invention can also be used for designing a lightweight SAR target detection model in a customized manner aiming at hardware equipment with limited computing resources, thereby being convenient for practical deployment and application.
The technical algorithm provided by the invention can be used for optimizing the structure of the existing model algorithm, and the structure of the existing model algorithm is secondarily perfected on the basis of the existing model, so that the effects of improving the accuracy of the original model algorithm and lightening the model are achieved.
(1) The invention designs an automatic algorithm SARNas aiming at the SAR target detection model design task, and constructs a flexible and efficient super-network search strategy and a subnet contribution degree evaluation strategy. The SARNas algorithm provided by the invention can autonomously perform end-to-end search, and automatically obtain the SAR target detection model with optimal comprehensive performance (comprising precision, model calculation amount, parameter and inference time).
(2) Aiming at the difficulty that the SAR detection model is often limited in resource constraint and cannot be smoothly applied when the edge-side hardware equipment is deployed, the invention designs a novel NAS Loss function SARMI-Loss which is used for guiding and learning the SAR target detector with more balanced performance by taking target detection precision and model calculation complexity (comprising reasoning speed, parameters, model calculation amount and the like) as combined guiding targets.
(3) The SARNas algorithm provided by the invention can be used as an aided design optimization tool of a designer, any SAR target detection model based on a deep learning technology can be automatically optimally designed according to the computing resource capacity of edge end deployment equipment, and the SARNas algorithm has extremely strong flexibility and customization space.
While preferred embodiments of the present invention have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. It is therefore intended that the following claims be interpreted as including the preferred embodiments and all such alterations and modifications as fall within the scope of the invention.
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 (6)
1. The SAR target detection method based on the neural architecture search is characterized by comprising the following steps of:
SAR image data are obtained;
inputting SAR image data into an SAR target detection model based on neural architecture search, and outputting an SAR target detection result;
obtaining SAR target type and position information according to the SAR target detection result;
the step of inputting the SAR image data into the SAR target detection model based on the neural architecture search and outputting the SAR target detection result comprises the following steps:
performing feature extraction on SAR image data based on a plurality of search layers of a backbone network searcher; each search layer corresponds to a corresponding convolution calculation layer of the CNN network, and a plurality of candidate calculation modules are added in parallel to each search layer;
integrating the extracted features based on a Neck module;
predicting the integrated features based on the Head module to obtain SAR target detection results;
the multiple search layers based on the backbone network searcher perform feature extraction on SAR image data, and specifically comprise the following steps:
computing a Block for each candidate of each search layer i Corresponding coefficient theta is distributed i ；
Calculating each candidate calculation module Block through Gumbel Softmax function i Corresponding actual weighted calculation coefficients
Block a plurality of candidate calculation blocks of the current search layer 1 i And the corresponding actual weighted calculation coefficientMultiplying and adding to obtain the calculation result of the current search layer;
and transmitting the calculation result of the current search layer to the next search layer, transmitting the calculation result of the current search layer to all candidate calculation blocks by the next search layer for parallel calculation, and obtaining a final output characteristic diagram of the backbone network searcher after calculation of all the search layers.
2. The SAR target detection method based on neural architecture search of claim 1, wherein the gummel Softmax function has the following calculation formula:
in the formula ,gi Gumbel (0, 1) is random Gumbel noise, τ is a temperature parameter, k is the total number of calculation modules of the search layer, and exp is an exponential function.
3. The SAR target detection method based on neural architecture search of claim 2, wherein the calculation formula of the search layer is as follows:
where LayerOutput is the output of the search layer.
4. The SAR target detection method of claim 1, wherein a neural architecture search-based loss function is constructed, and the weight parameters in the backbone network searcher are iteratively updated by the loss function.
5. The SAR target detection method based on neural architecture search of claim 4, wherein the formula of the loss function is as follows:
wherein ,
in precision loss For the current precision loss value of the model, the latency is loss To model reasoning time loss values, flow loss For model calculation loss, param loss The layer is a calculation layer, alpha is a parameter for adjusting the calculated loss, beta is a parameter for adjusting the loss value, and l is a parameter for adjusting the loss value i and pi Respectively is Block i FLPs (m) is the calculated amount in the current search network, T is the upper limit of the calculated amount, and gamma is a number much larger than 1.
6. The SAR target detection method based on neural architecture search as set forth in claim 4, wherein after the iterative updating of the weight parameters in the backbone network searcher is completed, selecting a value θ with the largest coefficient θ of each layer of search layer max Corresponding candidate calculation moduleAs a result of the search of the current search layer, the CNN convolution calculation layer corresponding to the current search layer will employ a candidate calculation module +.>As a convolution calculation method of this layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310376226.0A CN116416468B (en) | 2023-04-11 | 2023-04-11 | SAR target detection method based on neural architecture search |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310376226.0A CN116416468B (en) | 2023-04-11 | 2023-04-11 | SAR target detection method based on neural architecture search |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116416468A CN116416468A (en) | 2023-07-11 |
| CN116416468B true CN116416468B (en) | 2023-10-03 |
Family
ID=87052719
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310376226.0A Active CN116416468B (en) | 2023-04-11 | 2023-04-11 | SAR target detection method based on neural architecture search |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116416468B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117010447B (en) * | 2023-10-07 | 2024-01-23 | 成都理工大学 | End-to-end based microarchitecturable search method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110488288A (en) * | 2019-08-09 | 2019-11-22 | 西安电子科技大学 | A kind of carried SAR high-resolution chromatography imaging method |
| WO2021018163A1 (en) * | 2019-07-30 | 2021-02-04 | 华为技术有限公司 | Neural network search method and apparatus |
| US10976429B1 (en) * | 2017-10-16 | 2021-04-13 | Hrl Laboratories, Llc | System and method for synthetic aperture radar target recognition utilizing spiking neuromorphic networks |
| CN113076938A (en) * | 2021-05-06 | 2021-07-06 | 广西师范大学 | Deep learning target detection method combined with embedded hardware information |
| CN113780152A (en) * | 2021-09-07 | 2021-12-10 | 北京航空航天大学 | A small target detection method for ships in remote sensing images based on target perception |
| CN115546555A (en) * | 2022-10-18 | 2022-12-30 | 安徽大学 | Lightweight SAR target detection method based on hybrid characterization learning enhancement |
| CN115688908A (en) * | 2022-09-28 | 2023-02-03 | 东南大学 | An Efficient Neural Network Search and Training Method Based on Pruning Technology |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US12112112B2 (en) * | 2020-11-12 | 2024-10-08 | Samsung Electronics Co., Ltd. | Method for co-design of hardware and neural network architectures using coarse-to-fine search, two-phased block distillation and neural hardware predictor |
-
2023
- 2023-04-11 CN CN202310376226.0A patent/CN116416468B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10976429B1 (en) * | 2017-10-16 | 2021-04-13 | Hrl Laboratories, Llc | System and method for synthetic aperture radar target recognition utilizing spiking neuromorphic networks |
| WO2021018163A1 (en) * | 2019-07-30 | 2021-02-04 | 华为技术有限公司 | Neural network search method and apparatus |
| CN110488288A (en) * | 2019-08-09 | 2019-11-22 | 西安电子科技大学 | A kind of carried SAR high-resolution chromatography imaging method |
| CN113076938A (en) * | 2021-05-06 | 2021-07-06 | 广西师范大学 | Deep learning target detection method combined with embedded hardware information |
| CN113780152A (en) * | 2021-09-07 | 2021-12-10 | 北京航空航天大学 | A small target detection method for ships in remote sensing images based on target perception |
| CN115688908A (en) * | 2022-09-28 | 2023-02-03 | 东南大学 | An Efficient Neural Network Search and Training Method Based on Pruning Technology |
| CN115546555A (en) * | 2022-10-18 | 2022-12-30 | 安徽大学 | Lightweight SAR target detection method based on hybrid characterization learning enhancement |
Non-Patent Citations (2)
| Title |
|---|
| 基于多结构卷积神经网络的高分遥感影像飞机目标检测;姚相坤;万里红;霍宏;方涛;;计算机工程(第01期);全文 * |
| 基于特征融合卷积神经网络的SAR图像目标检测方法;刘志宏;李玉峰;;微处理机(第02期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116416468A (en) | 2023-07-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108647655B (en) | Low-altitude aerial image power line foreign object detection method based on light convolutional neural network | |
| CN116740587B (en) | A reliable identification method for drone aerial photography targets based on data and model uncertainty perception | |
| CN116306813B (en) | A method based on YOLOX lightweight and network optimization | |
| CN113408423A (en) | Aquatic product target real-time detection method suitable for TX2 embedded platform | |
| CN116883822B (en) | A weak feature target detection method based on multiple prior guidance | |
| CN117649526A (en) | High-precision semantic segmentation method for automatic driving road scene | |
| CN115661677B (en) | Lightweight satellite image cloud detection method based on dark channel feature guidance | |
| CN116416468B (en) | SAR target detection method based on neural architecture search | |
| CN118570447A (en) | Feature balance-based remote sensing image fine-grained target detection method and system | |
| Li et al. | Channel pruned YOLOv5-based deep learning approach for rapid and accurate outdoor obstacles detection | |
| CN119068016A (en) | A RGBT target tracking method based on modality-aware feature learning | |
| CN113327271B (en) | Decision-level target tracking method and system based on double-optical twin network and storage medium | |
| CN119445086A (en) | Real-time detection method and system for insulator defects based on channel pruning and distillation | |
| CN118505998A (en) | Real-time semantic segmentation system and segmentation method based on three-way segmentation fusion | |
| CN114842214B (en) | Sonar image target recognition method based on reinforcement learning | |
| CN116704378A (en) | Homeland mapping data classification method based on self-growing convolution neural network | |
| CN116975677A (en) | Dynamic spectrum signal identification method, device, electronic equipment and storage medium | |
| CN115147718A (en) | Scene self-adaption system and method for unmanned mobile terminal visual analysis | |
| CN117496119A (en) | Target detection method and device based on light brain-like target detection model | |
| Li et al. | Single object tracking based on OSTrack and NAS head | |
| CN116958199B (en) | A quantization-friendly twin network target tracking method and device | |
| CN116310636B (en) | A Lightweight Neural Network Structure Search Method Based on Neural Network Topology | |
| CN119274064B (en) | Litchi identification and yield estimation method, equipment and medium based on image vision | |
| CN120409567B (en) | Transformer adaptive compression method and system in data-limited scenarios | |
| CN116992944B (en) | Image processing method and device based on leavable importance judging standard pruning |
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 |