WO2020239140A1 - 基于多尺度特征与宽度学习的sar图像识别方法及装置 - Google Patents

基于多尺度特征与宽度学习的sar图像识别方法及装置 Download PDF

Info

Publication number
WO2020239140A1
WO2020239140A1 PCT/CN2020/102822 CN2020102822W WO2020239140A1 WO 2020239140 A1 WO2020239140 A1 WO 2020239140A1 CN 2020102822 W CN2020102822 W CN 2020102822W WO 2020239140 A1 WO2020239140 A1 WO 2020239140A1
Authority
WO
WIPO (PCT)
Prior art keywords
sar image
feature
lbp
lpq
width learning
Prior art date
Application number
PCT/CN2020/102822
Other languages
English (en)
French (fr)
Inventor
翟懿奎
余翠琳
余忠信
邓文博
甘俊英
应自炉
王天雷
曾军英
Original Assignee
五邑大学
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 五邑大学 filed Critical 五邑大学
Publication of WO2020239140A1 publication Critical patent/WO2020239140A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
    • G01S13/88—Radar or analogous systems specially adapted for specific applications
    • G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
    • G01S13/90—Radar or analogous systems specially adapted for specific applications for mapping or imaging using synthetic aperture techniques, e.g. synthetic aperture radar [SAR] techniques
    • G01S13/9021—SAR image post-processing techniques
    • G01S13/9027—Pattern recognition for feature extraction
    • 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
    • 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
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/60—Rotation of whole images or parts thereof
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/776—Validation; Performance evaluation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • 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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/467—Encoded features or binary features, e.g. local binary patterns [LBP]

Definitions

  • the invention relates to the field of image recognition, in particular to a SAR image recognition method and device based on multi-scale feature and width learning.
  • SAR is currently one of the important methods of earth observation. Due to its all-weather and strong penetrating imaging capabilities, SAR has been widely used in fields such as battlefield reconnaissance and intelligence acquisition. However, because SAR images are more sensitive to the orientation of the imaging, and there are a lot of environmental noises in the target data; traditional recognition and classification methods have achieved a certain degree of classification accuracy, but there are still many defects, such as low performance for SAR image feature extraction, matrix The convergence speed is slow, the training time is longer, the parameter estimation is complicated, and the efficiency is low.
  • the purpose of the present invention is to solve at least one of the technical problems in the prior art, and to provide a SAR image recognition method and device, and improve the performance of SAR image recognition through a multi-scale feature and width learning network.
  • the first aspect of the present invention provides a SAR image recognition method based on multi-scale feature and width learning, including the following steps:
  • the dimensionality reduction is achieved through the principal component analysis method to obtain the fused feature data X m ;
  • the above-mentioned SAR image recognition method based on multi-scale features and width learning has at least the following beneficial effects: using the fusion feature data fused with LBP features and LPQ features as the input of the width learning network, it makes full use of complementary information while reducing redundant information And improve the feature extraction performance; using the width learning network to recognize SAR images, the training speed is fast, the reconstruction cost is small, and the time cost is greatly reduced. At the same time, it reduces the impact of noise in the SAR image on the recognition results; makes the recognition effect more stable and robust. Great and reliable.
  • the extraction of the region of interest from the original SAR image by the centroid positioning method to obtain the SAR image containing the region of interest specifically includes the following steps: determining the centroid coordinates through the centroid formula; obtaining L* with the centroid as the center The region of interest of L, and then the SAR image containing the region of interest is obtained; where L is the side length of the region of interest; the centroid formula is (x c ,y c ) are the centroid coordinates, m 10 and m 01 are the first-order original matrix of the original SAR image, and m 00 is the zero-order original matrix of the original SAR image.
  • extracting LBP features from a multi-scale SAR image to obtain the LBP feature vector X LBP specifically includes: extracting LBP features from the multi-scale SAR image through an LBP feature operator, and cascading the LBP features
  • the LBP feature vector X LBP is obtained by fusion; among them, the LBP feature operator is (x c , y c ) are the coordinates of the centroid, g c is the gray value of the centroid, g p is the pixel points distributed equidistantly on a circle with the centroid as the center and r as the radius, and p is the number of g p .
  • the input fusion feature data X m to the width learning network to perform image recognition and output the recognition result specifically includes:
  • the input training set X 1 to the width learning network, and training the width learning network to optimize connection weights specifically includes:
  • Generate enhanced nodes input the training set X 1 to generate m groups of enhanced nodes, where the m- th group of characteristic nodes is H m ⁇ (Z n W hm + ⁇ hm ), where W hm is a random matrix different from W ei , ⁇ hm is a bias different from ⁇ ei , ⁇ ( ⁇ ) is a nonlinear activation function different from ⁇ ( ⁇ ), and Z n represents all feature nodes;
  • the second aspect of the present invention provides a multi-scale feature and width learning-based SAR image recognition device using the multi-scale feature and width learning-based SAR image recognition method of the first invention of the present invention, including:
  • Image input module used to input the original SAR image
  • the region of interest extraction module is used to extract the region of interest from the original SAR image through the centroid positioning method to obtain a SAR image containing the region of interest;
  • the data enhancement module is used to rotate the SAR image containing the region of interest and add noise to enhance the data volume to obtain a data-enhanced SAR image;
  • the down-sampling module is used to down-sample the data-enhanced SAR image to obtain a multi-scale SAR image
  • the LBP feature extraction module is used to extract LBP features from multi-scale SAR images and obtain the LBP feature vector X LBP ;
  • the LPQ feature extraction module is used to extract LPQ features from multi-scale SAR images and obtain the LPQ feature vector X LPQ ;
  • the fusion feature module is used to cascade the LBP feature vector X LBP and the LPQ feature vector X LPQ to achieve dimensionality reduction through the principal component analysis method to obtain the fused feature data X m ;
  • the width learning network module is used to input the fusion feature data X m for image recognition and output the recognition result.
  • the above-mentioned SAR image recognition device based on multi-scale features and width learning has at least the following beneficial effects: the network architecture is simple and the division of labor is clear; the fusion feature data fused with LBP features and LPQ features are used as the input of the width learning network, making full use of complementary information At the same time, the redundant information is reduced and the feature extraction performance is improved; the width learning network is used to recognize the SAR image, the training speed is fast, the time cost is greatly reduced, and the influence of noise in the SAR image is reduced; the recognition effect is more stable, robust, reliable.
  • FIG. 1 is a flowchart of a SAR image recognition method based on multi-scale feature and width learning according to an embodiment of the present invention
  • FIG. 2 is a specific flowchart of step S800 in FIG. 1;
  • FIG. 3 is a schematic structural diagram of a SAR image recognition device based on multi-scale feature and width learning according to an embodiment of the present invention.
  • an embodiment of the present invention provides a SAR image recognition method based on multi-scale feature and width learning, including the following steps:
  • S200 Extract the region of interest from the original SAR image by using the centroid positioning method to obtain a SAR image containing the region of interest;
  • S300 Rotate the SAR image containing the region of interest and add noise to enhance the data volume to obtain a data-enhanced SAR image;
  • S800 Input the fusion feature data X m to the width learning network for image recognition and output the recognition result.
  • the fusion feature data fused with LBP features and LPQ features are used as the input of the width learning network, which makes full use of complementary information while reducing redundant information and improving feature extraction performance; using the width learning network to recognize SAR images ,
  • the training speed is fast, the reconstruction cost is small, the time cost is greatly reduced, and the influence of the noise in the SAR image on the recognition result is reduced; the recognition effect is more stable, robust and reliable.
  • step S200 the centroid coordinates are determined by the centroid formula; the L*L region of interest is obtained with the centroid as the center, and then the SAR image containing the region of interest is obtained; where L is the side length of the region of interest, and L is more than The length and width of the original SAR image should be small; the centroid formula is (x c ,y c ) are the centroid coordinates, m 10 and m 01 are the first-order original matrix of the original SAR image, and m 00 is the zero-order original matrix of the original SAR image. Step S200 reduces the dimensions of the image and reduces the influence of background noise.
  • step S300 the SAR image containing the region of interest is rotated once every 1 degree, and rotated 360 degrees to achieve 360 times data enhancement; adding random integers as random noise to the rotated SAR image to further enhance the data,
  • the random integer is any integer in [-8,8].
  • Step S300 greatly enhances the data volume of the SAR image, which is beneficial to the improvement of the recognition accuracy.
  • the data-enhanced SAR image is down-sampling processing, and the down-sampling processing is to perform pixel reduction processing on the image to obtain a multi-scale SAR image.
  • the down-sampling processing is to perform pixel reduction processing on the image to obtain a multi-scale SAR image. For example, a SAR image with an original size of 64 pixels*64 pixels is down-sampled to obtain a SAR image of 54 pixels*54 pixels, 45 pixels*45 pixels, and 20 pixels*20 pixels.
  • the multi-scale SAR image is subjected to the LBP feature operator to extract the LBP feature.
  • the LBP feature is a data matrix
  • the cascade fusion is to parallel the data matrix corresponding to the LBP feature horizontally to obtain the LBP feature vector X LBP .
  • the LBP feature operator is (x c , y c ) are the coordinates of the centroid
  • g c is the gray value of the centroid
  • g p is the pixel points distributed equidistantly on a circle with the centroid as the center and r as the radius
  • p is the number of g p .
  • step S600 LPQ features are extracted from the multi-scale SAR image through the LPQ algorithm, and the LPQ feature vector X LPQ is obtained .
  • the LPQ algorithm utilizes the ambiguity invariance of the Fourier transform phase to improve the robustness of fuzzy image recognition; calculates the phase information after the Fourier transform in the neighborhood of each pixel of the SAR image, and calculates the low-frequency component Perform quantization to form LPQ features in the form of a data matrix.
  • the LPQ features are cascaded and merged to obtain the LPQ feature vector X LPQ .
  • step S700 the LBP feature vector X LBP obtained in step S500 and the LPQ feature vector X LPQ obtained in step S600 are concatenated, and then the dimensionality reduction is achieved by the principal component analysis method to obtain the fused feature data X m .
  • the fusion feature data fused with LBP features and LPQ features have stronger data features, which is beneficial to improve the accuracy of SAR image recognition.
  • the principal component analysis method is to map n-dimensional features to k-dimensional features.
  • the k-dimensional feature is a new orthogonal feature, also known as the principal component.
  • the principal component analysis method is to sequentially find a set of mutually orthogonal coordinate axes from the original space.
  • the choice of the new coordinate axis is closely related to the n-dimensional feature data itself.
  • the first new coordinate axis selection is the direction with the largest variance in the original data
  • the second new coordinate axis selection is the plane orthogonal to the first coordinate axis to maximize the variance
  • the third axis is the same as the first
  • the two-axis orthogonal plane has the largest variance.
  • n such coordinate axes can be obtained, and new coordinate axes obtained in this way. Only the first k coordinate axes that contain most of the variance are retained, and the coordinate axes that contain almost zero variance are ignored, so as to realize the dimensionality reduction processing of data features.
  • step S800 the fusion feature data X m is input to the width learning network.
  • the m- th group of feature nodes are H m ⁇ (Z n W hm + ⁇ hm ), where W hm is a random matrix different from We ei , and ⁇ hm is the same as ⁇ ei
  • W hm is a random matrix different from We ei
  • ⁇ hm is the same as ⁇ ei
  • ⁇ ( ⁇ ) is a non-linear activation function different from ⁇ ( ⁇ )
  • Z n represents all feature nodes.
  • H m ]W m ; then the connection weight of the width learning network It is: W m [Z i
  • FIG. 3 another embodiment of the present invention provides a SAR image recognition device applying the above-mentioned SAR image recognition method based on multi-scale feature and width learning, including:
  • the image input module 1 executes step S100 for inputting the original SAR image
  • the region of interest extraction module 2 executes step S200 for extracting the region of interest from the original SAR image by the centroid positioning method to obtain a SAR image containing the region of interest;
  • the data enhancement module 3 performs step S300, which is used to rotate and add noise to the SAR image containing the region of interest to enhance the amount of data to obtain a data-enhanced SAR image;
  • the down-sampling module 4 performs step S400 for down-sampling the data-enhanced SAR image to obtain a multi-scale SAR image;
  • the LBP feature extraction module 5 executes step S500 to extract LBP features from a multi-scale SAR image, and obtain the LBP feature vector X LBP ;
  • the LPQ feature extraction module 6 executes step S600 to extract LPQ features from the multi-scale SAR image, and obtain the LPQ feature vector X LPQ ;
  • the fusion feature module 7 executes step S700, which is used to cascade the LBP feature vector X LBP and the LPQ feature vector X LPQ to achieve dimensionality reduction through the principal component analysis method to obtain the fused feature data X m ;
  • the width learning network module 8 executes steps S810 to S830, and is used to input the fusion feature data X m for image recognition and output the recognition result.
  • the coordinates of the center of mass are first determined by the center of mass formula; then the L*L region of interest is obtained with the center of mass as the center, and then the SAR image containing the region of interest is obtained; where L is the region of interest The side length of L is smaller than the length and width of the original SAR image; the centroid formula is (x c ,y c ) are the centroid coordinates, m 10 and m 01 are the first-order original matrix of the original SAR image, and m 00 is the zero-order original matrix of the original SAR image.
  • the SAR image containing the region of interest is rotated once every 1 degree and rotated 360 degrees to achieve 360 times data enhancement; the rotated SAR image is added with random integers as random noise to further enhance Data, the random integer is any integer in [-8,8].
  • the data-enhanced SAR image is down-sampled, and the down-sampling process is to reduce the pixels of the image to obtain a multi-scale SAR image.
  • the multi-scale SAR image is extracted through the LBP feature operator to extract the LBP feature.
  • the LBP feature is a data matrix
  • the cascade fusion is to parallel the data matrix corresponding to the LBP feature horizontally to obtain the LBP feature vector X LBP .
  • the LBP feature operator is (x c , y c ) are the coordinates of the centroid
  • g c is the gray value of the centroid
  • g p is the pixel points distributed equidistantly on a circle with the centroid as the center and r as the radius
  • p is the number of g p .
  • the LPQ feature is extracted from the multi-scale SAR image through the LPQ algorithm, and then the LPQ feature is cascaded and fused to obtain the LPQ feature vector X LPQ .
  • the LBP feature vector X LBP obtained in the LBP feature extraction module 5 and the LPQ feature vector X LPQ obtained in the LPQ feature extraction module 6 are cascaded, and then the dimensionality reduction is achieved by the principal component analysis method to obtain the fusion Characteristic data X m .
  • the fusion feature data fused with LBP features and LPQ features have stronger data features, which is beneficial to improve the accuracy of SAR image recognition.
  • the width learning network module 8 contains the model of the width learning network. First, input the fusion feature data X m obtained from the fusion feature module 7, and then divide the fusion feature data X m into a training set X 1 and a test set X 2 , and perform z-score standardization on the training set X 1 and the test set X 2 .
  • the m- th group of feature nodes are H m ⁇ (Z n W hm + ⁇ hm ), where W hm is a random matrix different from We ei , and ⁇ hm is the same as ⁇ ei
  • W hm is a random matrix different from We ei
  • ⁇ hm is the same as ⁇ ei
  • ⁇ ( ⁇ ) is a non-linear activation function different from ⁇ ( ⁇ )
  • Z n represents all feature nodes.
  • H m ]W m ; then the connection weight of the width learning network It is: W m [Z i
  • a SAR image recognition device which includes a processor and a memory connected to the processor; the memory stores instructions executable by the processor, and the instructions are The processor executes, so that the processor can execute the SAR image recognition method based on multi-scale feature and width learning as described above.
  • Another embodiment of the present invention provides a storage medium that stores computer-executable instructions, and the computer-executable instructions are used to make a computer execute the above-mentioned multi-scale feature and width learning-based SAR image recognition method.
  • the above-mentioned SAR image recognition device and storage medium using the SAR image recognition method based on multi-scale feature and width learning by using the fusion feature data of the fusion of LBP features and LPQ features as the input of the width learning network, while making full use of complementary information Reduced redundant information and improved feature extraction performance; using width learning network to recognize SAR images, the training speed is fast, the time cost is greatly reduced, and the influence of noise in the SAR image is reduced; the recognition effect is more stable, robust and reliable.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Remote Sensing (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Biology (AREA)
  • Electromagnetism (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Astronomy & Astrophysics (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

本发明公开了基于多尺度特征与宽度学习的SAR图像识别方法及装置，通过质心定位法对原始SAR图像提取感兴趣区域，对图像旋转并添加噪声以增强数据量，对图像下采样处理，提取LBP特征，提取LPQ特征，将LBP特征向量X LBP和LPQ特征向量X LPQ级联后通过主成分分析方法降维得到融合特征数据X m，输入融合特征数据X m到宽度学习网络进行图像识别并输出识别结果；通过融合LBP特征和LPQ特征充分利用互补信息的同时降低了冗余信息，使用宽度学习网络提升训练速度，降低时间成本；使得识别效果更加稳定、鲁棒、可靠。

Description

基于多尺度特征与宽度学习的SAR图像识别方法及装置 技术领域
本发明涉及图像识别领域，特别是基于多尺度特征与宽度学习的SAR图像识别方法及装置。
背景技术
SAR是当前对地观测的重要手段之一。由于具有全天候性和强穿透成像能力，SAR已经被广泛应用于战场侦查和情报获取等领域。但由于SAR图像对成像的方位比较敏感，并且目标数据存在大量环境噪声；传统的识别分类方法在分类准确率上取得一定的高度，但还存在诸多缺陷，例如对于SAR图像特征提取性能低，矩阵收敛速度较慢，训练时间较长，参数估计复杂，效率较低。
发明内容
本发明的目的在于至少解决现有技术中存在的技术问题之一，提供了SAR图像识别方法及装置，通过多尺度特征和宽度学习网络提升SAR图像识别的性能。
本发明解决其问题所采用的技术方案是：
本发明的第一方面，提供了基于多尺度特征与宽度学习的SAR图像识别方法，包括以下步骤：
输入原始SAR图像；
通过质心定位法对原始SAR图像提取感兴趣区域，得到包含感兴 趣区域的SAR图像；
对包含感兴趣区域的SAR图像旋转并添加噪声以增强数据量，得到数据增强的SAR图像；
对数据增强的SAR图像下采样处理，得到多尺度的SAR图像；
对多尺度的SAR图像提取LBP特征，并得到LBP特征向量X LBP；
对多尺度的SAR图像提取LPQ特征，并得到LPQ特征向量X LPQ；
将LBP特征向量X LBP和LPQ特征向量X LPQ级联后通过主成分分析方法实现降维得到融合特征数据X m；
输入融合特征数据X m到宽度学习网络进行图像识别并输出识别结果。
上述基于多尺度特征与宽度学习的SAR图像识别方法至少具有以下的有益效果：用融合LBP特征和LPQ特征的融合特征数据作为宽度学习网络的输入，在充分利用互补信息的同时降低了冗余信息并提高特征提取性能；利用宽度学习网络对SAR图像识别，训练速度快，重构代价小，时间成本大大降低，同时减小SAR图像中的噪声对识别结果的影响；使得识别效果更加稳定、鲁棒、可靠。
根据本发明的第一方面，所述通过质心定位法对原始SAR图像提取感兴趣区域，得到包含感兴趣区域的SAR图像具体包括以下步骤：通过质心公式确定质心坐标；以质心为中心得到L*L的感兴趣区域，进而得到包含感兴趣区域的SAR图像；其中，L为感兴趣区域的边长；质心公式为
Figure PCTCN2020102822-appb-000001
(x c,y c)为质心坐标，m 10和m 01为原始SAR 图像的一阶原始矩阵，m 00为原始SAR图像的零阶原始矩阵。
根据本发明的第一方面，所述对多尺度的SAR图像提取LBP特征，并得到LBP特征向量X LBP具体为：将多尺度的SAR图像经过LBP特征算子提取LBP特征，将LBP特征级联融合得到LBP特征向量X LBP；其中，LBP特征算子为
Figure PCTCN2020102822-appb-000002
(x c,y c)为质心坐标，g c为质心灰度值，g p为等距离分布于以质心为圆心、r为半径的圆周上的像素点，p为g p的数量。
根据本发明的第一方面，所述输入融合特征数据X m到宽度学习网络进行图像识别并输出识别结果具体包括：
将融合特征数据X m分为训练集X 1和测试集X 2；
输入训练集X 1至宽度学习网络，训练宽度学习网络优化连接权重；
输入测试集X 2至训练后的宽度学习网络得到识别结果。
根据本发明的第一方面，所述输入训练集X 1至宽度学习网络，训练宽度学习网络优化连接权重具体包括：
生成特征节点：输入训练集X 1实现投影并产生i组特征节点，其中第i组特征节点为Z i＝φ(X 1W ei+β ei)，式中W ei为随机矩阵，β ei为偏置，φ(·)为非线性激活函数；
生成增强节点：输入训练集X 1产生m组增强节点，其中第m组特征节点为H m≡ξ(Z nW hm+β hm)，式中W hm为与W ei不同的随机矩阵，β hm为与β ei不同的偏置，ξ(·)为与φ(·)不同的非线性激活函数，Z n表示所有 特征节点；
优化连接权重：将i组特征节点和m组增强节点相互连接得到合并矩阵作为宽度学习网络的实际输入，根据输出矩阵Y计算宽度学习网络的连接权重为：W m＝[Z i|H m] +Y；通过训练集X 1对宽度学习网络不断训练以优化连接权重。
本发明的第二方面，提供了应用本发明第一发明所述的基于多尺度特征与宽度学习的SAR图像识别方法的基于多尺度特征与宽度学习的SAR图像识别装置，包括：
图像输入模块，用于输入原始SAR图像；
感兴趣区域提取模块，用于通过质心定位法对原始SAR图像提取感兴趣区域，得到包含感兴趣区域的SAR图像；
数据增强模块，用于对包含感兴趣区域的SAR图像旋转并添加噪声以增强数据量，得到数据增强的SAR图像；
下采样模块，用于对数据增强的SAR图像下采样处理，得到多尺度的SAR图像；
LBP特征提取模块，用于对多尺度的SAR图像提取LBP特征，并得到LBP特征向量X LBP；
LPQ特征提取模块，用于对多尺度的SAR图像提取LPQ特征，并得到LPQ特征向量X LPQ；
融合特征模块，用于将LBP特征向量X LBP和LPQ特征向量X LPQ级联后通过主成分分析方法实现降维得到融合特征数据X m；
宽度学习网络模块，用于输入融合特征数据X m进行图像识别并输出识别结果。
上述基于多尺度特征与宽度学习的SAR图像识别装置至少具有以下的有益效果：网络架构简单且分工明确；用融合LBP特征和LPQ特征的融合特征数据作为宽度学习网络的输入，在充分利用互补信息的同时降低了冗余信息并提高特征提取性能；利用宽度学习网络对SAR图像识别，训练速度快，时间成本大大降低，同时减小SAR图像中的噪声影响；使得识别效果更加稳定、鲁棒、可靠。
附图说明
下面结合附图和实例对本发明作进一步说明。
图1是本发明实施例基于多尺度特征与宽度学习的SAR图像识别方法的流程图；
图2是图1中步骤S800的具体流程图；
图3是本发明实施例基于多尺度特征与宽度学习的SAR图像识别装置的结构示意图。
具体实施方式
本部分将详细描述本发明的具体实施例，本发明之较佳实施例在附图中示出，附图的作用在于用图形补充说明书文字部分的描述，使人能够直观地、形象地理解本发明的每个技术特征和整体技术方案，但其不能理解为对本发明保护范围的限制。
参照图1，本发明实施例提供了，基于多尺度特征与宽度学习的 SAR图像识别方法，包括以下步骤：
S100、输入原始SAR图像；
S200、通过质心定位法对原始SAR图像提取感兴趣区域，得到包含感兴趣区域的SAR图像；
S300、对包含感兴趣区域的SAR图像旋转并添加噪声以增强数据量，得到数据增强的SAR图像；
S400、对数据增强的SAR图像下采样处理，得到多尺度的SAR图像；
S500、对多尺度的SAR图像提取LBP特征，并得到LBP特征向量X LBP；
S600、对多尺度的SAR图像提取LPQ特征，并得到LPQ特征向量X LPQ；
S700、将LBP特征向量X LBP和LPQ特征向量X LPQ级联后通过主成分分析方法实现降维得到融合特征数据X m；
S800、输入融合特征数据X m到宽度学习网络进行图像识别并输出识别结果。
在该实施例中，用融合LBP特征和LPQ特征的融合特征数据作为宽度学习网络的输入，在充分利用互补信息的同时降低了冗余信息并提高特征提取性能；利用宽度学习网络对SAR图像识别，训练速度快，重构代价小，时间成本大大降低，同时减小SAR图像中的噪声对识别结果的影响；使得识别效果更加稳定、鲁棒、可靠。
进一步，在步骤S200中，通过质心公式确定质心坐标；以质心为中心得到L*L的感兴趣区域，进而得到包含感兴趣区域的SAR图像；其中，L为感兴趣区域的边长，L比原始SAR图像的长和宽都要小；质心公式为
Figure PCTCN2020102822-appb-000003
(x c,y c)为质心坐标，m 10和m 01为原始SAR图像的一阶原始矩阵，m 00为原始SAR图像的零阶原始矩阵。通过步骤S200减少图像的维度，降低背景噪声的影响。
进一步，在步骤S300中，对包含感兴趣区域的SAR图像每隔1度旋转一次，旋转360度，实现360倍的数据增强；对旋转后的SAR图像加入随机整数作为随机噪声再进一步增强数据，随机整数为[-8,8]中的任意整数。步骤S300大大增强了SAR图像的数据量，有利于识别准确率的提高。
进一步，在步骤S400中，对数据增强的SAR图像下采样处理，下采样处理为对图像进行缩小像素处理以得到多尺度的SAR图像。例如，一张原尺寸为64像素*64像素的SAR图像，下采样后得到54像素*54像素、45像素*45像素、20像素*20像素的SAR图像。
进一步，在步骤S500中，将多尺度的SAR图像经过LBP特征算子提取LBP特征。LBP特征为数据矩阵，级联融合即是将LBP特征对应的数据矩阵横向并联，进而得到LBP特征向量X LBP。其中，LBP特征算子为
Figure PCTCN2020102822-appb-000004
(x c,y c)为质心坐标，g c为质心灰度值，g p为等距离分布于以质心为圆心、r为半径的 圆周上的像素点，p为g p的数量。
在步骤S600中，对多尺度的SAR图像通过LPQ算法提取LPQ特征，并得到LPQ特征向量X LPQ。LPQ算法是利用傅里叶变换相位的模糊不变性，提高对模糊图像识别的鲁棒性；在SAR图像的每个像素点的邻域内计算傅里叶变换后的相位信息，并在低频分量上进行量化进而形成数据矩阵形式的LPQ特征。将LPQ特征级联融合，进而得到LPQ特征向量X LPQ。
在步骤S700中，将步骤S500中得到的LBP特征向量X LBP以及步骤S600中得到的LPQ特征向量X LPQ级联后通过主成分分析方法实现降维得到融合特征数据X m。融合LBP特征和LPQ特征的融合特征数据具有更强的数据特征，有利于提高SAR图像识别的准确性。
主成分分析方法是将n维特征映射到k维特征上。k维特征是全新的正交特征，也被称为主成分。主成分分析方法是从原始的空间中顺序地找一组相互正交的坐标轴，新的坐标轴的选择与n维特征的数据本身是密切相关的。其中，第一个新坐标轴选择是原始数据中方差最大的方向，第二个新坐标轴选取是与第一个坐标轴正交的平面中使得方差最大的，第三个轴是与第1、2个轴正交的平面中方差最大的。依次类推，可以得到n个这样的坐标轴，通过这种方式获得的新的坐标轴。只保留前面k个含有绝大部分方差的坐标轴，而忽略包含方差几乎为0的坐标轴，实现对数据特征的降维处理。
参照图2，在步骤S800中，输入融合特征数据X m到宽度学习网 络。
S810、将融合特征数据X m分为训练集X 1和测试集X 2，并对训练集X 1和测试集X 2进行z分数标准化。
S820、输入训练集X 1，对宽度学习网络进行训练；
S821、生成特征节点：
输入训练集X 1实现投影并产生i组特征节点，则第i组特征节点为Z i＝φ(X 1W ei+β ei)，其中W ei为随机矩阵，β ei为偏置，φ(·)为非线性激活函数。
S822、生成增强节点：
输入训练集X 1产生m组增强节点，则第m组特征节点为H m≡ξ(Z nW hm+β hm)，其中W hm为与W ei不同的随机矩阵，β hm为与β ei不同的偏置，ξ(·)为与φ(·)不同的非线性激活函数，Z n表示所有特征节点。
S823、优化连接权重：
将i组特征节点和m组增强节点相互连接得到合并矩阵，作为系统的实际输入，宽度学习网络的输出矩阵为：Y＝＝[Z i|H m]W m；那么宽度学习网络的连接权重为：W m＝[Z i|H m] +Y；通过训练集X 1对宽度学习网络不断训练以优化连接权重。
S830、输入测试集X 2到训练后的宽度学习网络中得到识别结果；利用宽度学习网络的资源需求量少、快速训练和增量学习特性，实现了准确、快速、自适应的SAR图像识别技术。
参照图3，本发明另一个实施例提供了，应用上述的基于多尺度 特征与宽度学习的SAR图像识别方法的SAR图像识别装置，包括：
图像输入模块1，执行步骤S100，用于输入原始SAR图像；
感兴趣区域提取模块2，执行步骤S200，用于通过质心定位法对原始SAR图像提取感兴趣区域，得到包含感兴趣区域的SAR图像；
数据增强模块3，执行步骤S300，用于对包含感兴趣区域的SAR图像旋转并添加噪声以增强数据量，得到数据增强的SAR图像；
下采样模块4，执行步骤S400，用于对数据增强的SAR图像下采样处理，得到多尺度的SAR图像；
LBP特征提取模块5，执行步骤S500，用于对多尺度的SAR图像提取LBP特征，并得到LBP特征向量X LBP；
LPQ特征提取模块6，执行步骤S600，用于对多尺度的SAR图像提取LPQ特征，并得到LPQ特征向量X LPQ；
融合特征模块7，执行步骤S700，用于将LBP特征向量X LBP和LPQ特征向量X LPQ级联后通过主成分分析方法实现降维得到融合特征数据X m；
宽度学习网络模块8，执行步骤S810至步骤S830，用于输入融合特征数据X m进行图像识别并输出识别结果。
进一步，在感兴趣区域提取模块2中，先通过质心公式确定质心坐标；再以质心为中心得到L*L的感兴趣区域，进而得到包含感兴趣区域的SAR图像；其中，L为感兴趣区域的边长，L比原始SAR图像的长和宽都要小；质心公式为
Figure PCTCN2020102822-appb-000005
(x c,y c)为质心坐标，m 10 和m 01为原始SAR图像的一阶原始矩阵，m 00为原始SAR图像的零阶原始矩阵。
进一步，在数据增强模块3中，对包含感兴趣区域的SAR图像每隔1度旋转一次，旋转360度，实现360倍的数据增强；对旋转后的SAR图像加入随机整数作为随机噪声再进一步增强数据，随机整数为[-8,8]中的任意整数。
进一步，在下采样模块4中，对数据增强的SAR图像下采样处理，下采样处理为对图像进行缩小像素处理以得到多尺度的SAR图像。
进一步，在LBP特征提取模块5中，将多尺度的SAR图像经过LBP特征算子提取LBP特征。LBP特征为数据矩阵，级联融合即是将LBP特征对应的数据矩阵横向并联，进而得到LBP特征向量X LBP。其中，LBP特征算子为
Figure PCTCN2020102822-appb-000006
(x c,y c)为质心坐标，g c为质心灰度值，g p为等距离分布于以质心为圆心、r为半径的圆周上的像素点，p为g p的数量。
进一步，在LPQ特征提取模块6中，对多尺度的SAR图像通过LPQ算法提取LPQ特征，再将LPQ特征级联融合，进而得到LPQ特征向量X LPQ。
进一步，在融合特征模块7中，将LBP特征提取模块5中得到的LBP特征向量X LBP以及LPQ特征提取模块6中得到的LPQ特征向量X LPQ级联后通过主成分分析方法实现降维得到融合特征数据X m。融合LBP特征和LPQ特征的融合特征数据具有更强的数据特征，有利于提高 SAR图像识别的准确性。
进一步，宽度学习网络模块8包含了宽度学习网络的模型。先输入从融合特征模块7得到的融合特征数据X m，再将融合特征数据X m分为训练集X 1和测试集X 2，并对训练集X 1和测试集X 2进行z分数标准化。
输入训练集X 1，对宽度学习网络进行训练；
生成特征节点：
输入训练集X 1实现投影并产生i组特征节点，则第i组特征节点为Z i＝φ(X 1W ei+β ei)，其中W ei为随机矩阵，β ei为偏置，φ(·)为非线性激活函数。
生成增强节点：
输入训练集X 1产生m组增强节点，则第m组特征节点为H m≡ξ(Z nW hm+β hm)，其中W hm为与W ei不同的随机矩阵，β hm为与β ei不同的偏置，ξ(·)为与φ(·)不同的非线性激活函数，Z n表示所有特征节点。
优化连接权重：
将i组特征节点和m组增强节点相互连接得到合并矩阵，作为系统的实际输入，宽度学习网络的输出矩阵为：Y＝＝[Z i|H m]W m；那么宽度学习网络的连接权重为：W m＝[Z i|H m] +Y；通过训练集X 1对宽度学习网络不断训练以优化连接权重。
输入测试集X 2到训练后的宽度学习网络中得到识别结果；利用宽度学习网络的资源需求量少、快速训练和增量学习特性，实现了准确、 快速、自适应的SAR图像识别技术。
此外，本发明另一个实施例提供了SAR图像识别装置，包括处理器和用于与所述处理器连接的存储器；所述存储器存储有可被所述处理器执行的指令，所述指令被所述处理器执行，以使所述处理器能够执行如上所述的基于多尺度特征与宽度学习的SAR图像识别方法。
本发明另一个实施例提供了一种存储介质，所述存储介质存储有计算机可执行指令，所述计算机可执行指令用于使计算机执行如上所述的基于多尺度特征与宽度学习的SAR图像识别方法。
上述应用了基于多尺度特征与宽度学习的SAR图像识别方法的SAR图像识别装置及存储介质，通过利用融合LBP特征和LPQ特征的融合特征数据作为宽度学习网络的输入，在充分利用互补信息的同时降低了冗余信息并提高特征提取性能；利用宽度学习网络对SAR图像识别，训练速度快，时间成本大大降低，同时减小SAR图像中的噪声影响；使得识别效果更加稳定、鲁棒、可靠。
以上所述，只是本发明的较佳实施例而已，本发明并不局限于上述实施方式，只要其以相同的手段达到本发明的技术效果，都应属于本发明的保护范围。

Claims (6)

  1. 基于多尺度特征与宽度学习的SAR图像识别方法，其特征在于，包括以下步骤：
    输入原始SAR图像；
    通过质心定位法对原始SAR图像提取感兴趣区域，得到包含感兴趣区域的SAR图像；
    对包含感兴趣区域的SAR图像旋转并添加噪声以增强数据量，得到数据增强的SAR图像；
    对数据增强的SAR图像下采样处理，得到多尺度的SAR图像；
    对多尺度的SAR图像提取LBP特征，并得到LBP特征向量X LBP；
    对多尺度的SAR图像提取LPQ特征，并得到LPQ特征向量X LPQ；
    将LBP特征向量X LBP和LPQ特征向量X LPQ级联后通过主成分分析方法实现降维得到融合特征数据X m；
    输入融合特征数据X m到宽度学习网络进行图像识别并输出识别结果。
  2. 根据权利要求1所述的基于多尺度特征与宽度学习的SAR图像识别方法，其特征在于，所述通过质心定位法对原始SAR图像提取感兴趣区域，得到包含感兴趣区域的SAR图像具体包括以下步骤：通过质心公式确定质心坐标；以质心为中心得到L*L的感兴趣区域，进而得到包含感兴趣区域的SAR图像；其中，L为感兴趣区域的边长；质心公式为
    Figure PCTCN2020102822-appb-100001
    (x c,y c)为质心坐标，m 10和 m 01为原始SAR图像的一阶原始矩阵，m 00为原始SAR图像的零阶原始矩阵。
  3. 根据权利要求2所述的基于多尺度特征与宽度学习的SAR图像识别方法，其特征在于，所述对多尺度的SAR图像提取LBP特征，并得到LBP特征向量X LBP具体为：将多尺度的SAR图像经过LBP特征算子提取LBP特征，将LBP特征级联融合得到LBP特征向量X LBP；其中，LBP特征算子为
    Figure PCTCN2020102822-appb-100002
    Figure PCTCN2020102822-appb-100003
    (x c,y c)为质心坐标，g c为质心灰度值，g p为等距离分布于以质心为圆心、r为半径的圆周上的像素点，p为g p的数量。
  4. 根据权利要求1所述的基于多尺度特征与宽度学习的SAR图像识别方法，其特征在于，所述输入融合特征数据X m到宽度学习网络进行图像识别并输出识别结果具体包括：
    将融合特征数据X m分为训练集X 1和测试集X 2；
    输入训练集X 1至宽度学习网络，训练宽度学习网络优化连接权重；
    输入测试集X 2至训练后的宽度学习网络得到识别结果。
  5. 根据权利要求4所述的基于多尺度特征与宽度学习的SAR图像识别方法，其特征在于，所述输入训练集X 1至宽度学习网络，训练宽度学习网络优化连接权重具体包括：
    生成特征节点：输入训练集X 1实现投影并产生i组特征节点，其中第i组特征节点为Z i＝φ(X 1W ei+β ei)，式中W ei为随机矩阵，β ei为偏置，φ(·)为非线性激活函数；
    生成增强节点：输入训练集X 1产生m组增强节点，其中第m组特征节点为H m≡ξ(Z nW hm+β hm)，式中W hm为与W ei不同的随机矩阵，β hm为与β ei不同的偏置，ξ(·)为与φ(·)不同的非线性激活函数，Z n表示所有特征节点；
    优化连接权重：将i组特征节点和m组增强节点相互连接得到合并矩阵作为宽度学习网络的实际输入，根据输出矩阵Y计算宽度学习网络的连接权重为：W m＝[Z i|H m] +Y；通过训练集X 1对宽度学习网络不断训练以优化连接权重。
  6. 应用权利要求1-5任一项所述的基于多尺度特征与宽度学习的SAR图像识别方法的基于多尺度特征与宽度学习的SAR图像识别装置，其特征在于，包括：
    图像输入模块，用于输入原始SAR图像；
    感兴趣区域提取模块，用于通过质心定位法对原始SAR图像提取
    感兴趣区域，得到包含感兴趣区域的SAR图像；
    数据增强模块，用于对包含感兴趣区域的SAR图像旋转并添加噪声以增强数据量，得到数据增强的SAR图像；
    下采样模块，用于对数据增强的SAR图像下采样处理，得到多尺度的SAR图像；
    LBP特征提取模块，用于对多尺度的SAR图像提取LBP特征，并得到LBP特征向量X LBP；
    LPQ特征提取模块，用于对多尺度的SAR图像提取LPQ特征，并得到LPQ特征向量X LPQ；
    融合特征模块，用于将LBP特征向量X LBP和LPQ特征向量X LPQ级联后通过主成分分析方法实现降维得到融合特征数据X m；
    宽度学习网络模块，用于输入融合特征数据X m进行图像识别并输出识别结果。
PCT/CN2020/102822 2019-05-30 2020-07-17 基于多尺度特征与宽度学习的sar图像识别方法及装置 WO2020239140A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910460194.6A CN110222700A (zh) 2019-05-30 2019-05-30 基于多尺度特征与宽度学习的sar图像识别方法及装置
CN201910460194.6 2019-05-30

Publications (1)

Publication Number Publication Date
WO2020239140A1 true WO2020239140A1 (zh) 2020-12-03

Family

ID=67818896

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/102822 WO2020239140A1 (zh) 2019-05-30 2020-07-17 基于多尺度特征与宽度学习的sar图像识别方法及装置

Country Status (3)

Country Link
US (1) US10977526B2 (zh)
CN (1) CN110222700A (zh)
WO (1) WO2020239140A1 (zh)

Families Citing this family (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110222700A (zh) * 2019-05-30 2019-09-10 五邑大学 基于多尺度特征与宽度学习的sar图像识别方法及装置
CN110929622B (zh) * 2019-11-15 2024-01-05 腾讯科技（深圳）有限公司 视频分类方法、模型训练方法、装置、设备及存储介质
CN111046936A (zh) * 2019-12-05 2020-04-21 湖北民族大学 基于级联宽度神经网络桥梁裂缝分类识别方法
CN111982302B (zh) * 2020-08-24 2023-12-29 广东工业大学 一种带噪声滤波和环境温度补偿的测温方法
CN112215095A (zh) * 2020-09-24 2021-01-12 西北工业大学 违禁品检测方法、装置、处理器和安检系统
CN112819007B (zh) * 2021-01-07 2023-08-01 北京百度网讯科技有限公司 图像识别方法、装置、电子设备及存储介质
CN112784856A (zh) * 2021-01-29 2021-05-11 长沙理工大学 胸部x射线图像的通道注意力特征提取方法和识别方法
DE102021102633A1 (de) * 2021-02-04 2022-08-04 Jungheinrich Aktiengesellschaft Verfahren zur Bestimmung einer Orientierung eines Flurförderzeugs
CN112926662B (zh) * 2021-02-25 2022-05-03 电子科技大学 一种基于多尺度语言嵌入rec的目标检测方法
CN113673534A (zh) * 2021-04-22 2021-11-19 江苏大学 一种基于Faster RCNN的RGB-D图像果实检测方法
CN113378660B (zh) * 2021-05-25 2023-11-07 广州紫为云科技有限公司 一种低数据成本的人脸识别的方法及装置
CN113390806B (zh) * 2021-06-15 2023-11-10 崔薇 基于多尺度图像的生菜重金属含量检测方法
CN113870194B (zh) * 2021-09-07 2024-04-09 燕山大学 深层特征和浅层lbp特征融合的乳腺肿瘤超声图像处理装置
WO2023139773A1 (ja) * 2022-01-24 2023-07-27 三菱電機株式会社 データ拡張方法、及びデータ拡張装置
CN114744946B (zh) * 2022-05-18 2023-06-06 江南大学 基于宽度学习干扰观测器的永磁同步电机控制方法及系统
CN114926753B (zh) * 2022-06-16 2023-10-13 无锡慧眼人工智能科技有限公司 一种海量图像条件下的快速目标场景信息提取方法
CN114821133B (zh) * 2022-06-27 2022-09-09 江苏瑞立环保工程股份有限公司 一种基于人工智能的环保设备故障检测方法及系统
CN115393301B (zh) * 2022-08-16 2024-03-12 中山大学附属第一医院 一种肝脏二维剪切波弹性图像的影像组学分析方法和装置
CN117036893B (zh) * 2023-10-08 2023-12-15 南京航空航天大学 一种基于局部跨阶段和快速下采样的图像融合方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103020654A (zh) * 2012-12-12 2013-04-03 北京航空航天大学 基于样本产生与核局部特征融合的sar图像仿生识别方法
KR101314293B1 (ko) * 2012-08-27 2013-10-02 재단법인대구경북과학기술원 조명변화에 강인한 얼굴인식 시스템
CN108805915A (zh) * 2018-04-19 2018-11-13 南京市测绘勘察研究院股份有限公司 一种抗视角变化的近景影像区域特征匹配方法
CN108921877A (zh) * 2018-07-25 2018-11-30 大连海事大学 一种基于宽度学习的长期目标跟踪算法
CN109585021A (zh) * 2018-11-28 2019-04-05 华南理工大学 基于全息投影技术的精神状态评估系统及其工作方法
CN110222700A (zh) * 2019-05-30 2019-09-10 五邑大学 基于多尺度特征与宽度学习的sar图像识别方法及装置

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9928406B2 (en) * 2012-10-01 2018-03-27 The Regents Of The University Of California Unified face representation for individual recognition in surveillance videos and vehicle logo super-resolution system
CN103605993B (zh) * 2013-12-04 2017-01-25 康江科技(北京)有限责任公司 一种基于面向场景判别性分析的图像到视频人脸识别方法
SG10201405182WA (en) * 2014-08-25 2016-03-30 Univ Singapore Technology & Design Method and system
CN104732601B (zh) * 2014-11-19 2017-05-10 东北大学 一种基于人脸识别技术的高识别率自动考勤装置及方法
US10872230B2 (en) * 2015-03-27 2020-12-22 Intel Corporation Low-cost face recognition using Gaussian receptive field features
CN106940782B (zh) * 2016-01-05 2020-08-07 中国科学院空天信息创新研究院 基于变差函数的高分sar新增建设用地提取软件
CN106650568B (zh) * 2016-08-31 2020-05-26 浙江大华技术股份有限公司 一种人脸识别方法及装置
CN106682621B (zh) * 2016-12-28 2019-01-29 西安电子科技大学 基于贝叶斯网络的sar图像识别方法
US20190108447A1 (en) * 2017-11-30 2019-04-11 Intel Corporation Multifunction perceptrons in machine learning environments
CN108416378B (zh) * 2018-02-28 2020-04-14 电子科技大学 一种基于深度神经网络的大场景sar目标识别方法
CN108960422B (zh) * 2018-06-19 2022-04-15 河南工业大学 一种基于主成分分析的宽度学习方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101314293B1 (ko) * 2012-08-27 2013-10-02 재단법인대구경북과학기술원 조명변화에 강인한 얼굴인식 시스템
CN103020654A (zh) * 2012-12-12 2013-04-03 北京航空航天大学 基于样本产生与核局部特征融合的sar图像仿生识别方法
CN108805915A (zh) * 2018-04-19 2018-11-13 南京市测绘勘察研究院股份有限公司 一种抗视角变化的近景影像区域特征匹配方法
CN108921877A (zh) * 2018-07-25 2018-11-30 大连海事大学 一种基于宽度学习的长期目标跟踪算法
CN109585021A (zh) * 2018-11-28 2019-04-05 华南理工大学 基于全息投影技术的精神状态评估系统及其工作方法
CN110222700A (zh) * 2019-05-30 2019-09-10 五邑大学 基于多尺度特征与宽度学习的sar图像识别方法及装置

Also Published As

Publication number Publication date
CN110222700A (zh) 2019-09-10
US20200380294A1 (en) 2020-12-03
US10977526B2 (en) 2021-04-13

Similar Documents

Publication Publication Date Title
WO2020239140A1 (zh) 基于多尺度特征与宽度学习的sar图像识别方法及装置
CN111862126B (zh) 深度学习与几何算法结合的非合作目标相对位姿估计方法
WO2019020075A1 (zh) 图像处理方法、装置、存储介质、计算机程序和电子设备
WO2022206020A1 (zh) 图像场景深度的估计方法、装置、终端设备和存储介质
CN107516322A (zh) 一种基于对数极空间的图像物体大小和旋转估计计算方法
CN112508991B (zh) 一种前后景分离的熊猫照片卡通化方法
CN111709307B (zh) 一种基于分辨率增强的遥感图像小目标检测方法
CN116097307A (zh) 图像的处理方法及相关设备
CN110580715B (zh) 一种基于照度约束和格网变形的图像对齐方法
CN114358150A (zh) 一种sar-可见光遥感图像匹配方法
Zhao et al. Unsupervised color retention network and new quantization metric for blind motion deblurring
WO2024082602A1 (zh) 一种端到端视觉里程计方法及装置
CN114693755B (zh) 多模图像最大矩与空间一致性的非刚性配准方法及系统
CN114399547B (zh) 一种基于多帧的单目slam鲁棒初始化方法
CN114529455A (zh) 一种基于任务解耦的有参图像超分辨率方法和系统
Zhang et al. Bilateral upsampling network for single image super-resolution with arbitrary scaling factors
CN110991354A (zh) 基于逆空间变换网络的isar图像识别方法
CN113409365B (zh) 图像处理方法及相关终端、设备和存储介质
WO2018120043A1 (zh) 图像重建方法及装置
CN116310657B (zh) 特征点检测模型训练方法、图像特征匹配方法及装置
Pal et al. MAML-SR: Self-adaptive super-resolution networks via multi-scale optimized attention-aware meta-learning
CN117237858B (zh) 一种回环检测方法
US20230177722A1 (en) Apparatus and method with object posture estimating
CN117788296B (zh) 基于异构组合深度网络的红外遥感图像超分辨率重建方法
Wei et al. Remote Sensing Image Registration Based on Grid Prediction of Pixel Deformation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20813574

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20813574

Country of ref document: EP

Kind code of ref document: A1