CN105243154B - Remote sensing image retrieval method based on notable point feature and sparse own coding and system - Google Patents
Remote sensing image retrieval method based on notable point feature and sparse own coding and system Download PDFInfo
- Publication number
- CN105243154B CN105243154B CN201510708598.4A CN201510708598A CN105243154B CN 105243154 B CN105243154 B CN 105243154B CN 201510708598 A CN201510708598 A CN 201510708598A CN 105243154 B CN105243154 B CN 105243154B
- Authority
- CN
- China
- Prior art keywords
- image
- notable
- feature
- matrix
- sparse
- 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; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
-
- 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]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Library & Information Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The characteristic point of each image obtains characteristic point matrix in a kind of remote sensing image retrieval method and system based on notable point feature and sparse own coding, including extraction image library, and the notable figure of each image is calculated using visual attention model；Using Adaptive Thresholding by notable figure binaryzation, and filtered remarkable characteristic is obtained into line mask operation with characteristic point matrix；Several remarkable characteristics are chosen respectively from each training image and construct training sample, and sparse autoencoder network is trained according to the training sample set after albefaction, obtains feature extractor；Feature extraction is carried out using feature extractor, threshold function table is used in combination to carry out LS-SVM sparseness to the characteristics of image of extraction, obtains the final feature vector for retrieval；Feature vector based on extraction carries out image retrieval according to preset similarity measurement criterion.The present invention realizes automatically extracting for characteristics of image by trained sparse autoencoder network, and the feature extracted has good discernment, ensure that retrieval precision ratio.
Description
Technical field
The invention belongs to technical field of image processing, are related to a kind of remote sensing figure based on notable point feature and sparse own coding
As search method and system.
Background technology
With the raising of remote sensing earth observation ability, retrievable remotely-sensed data shows the spy of diversification and magnanimity
Point.However, while mass remote sensing data provides abundant data source for all kinds of major application demands, at current ground data
Reason and analysis ability are insufficient, and " data magnanimity, information are flooded " problem of remote sensing big data becomes increasingly conspicuous.How emerging section is utilized
Computing technique and means are learned, quickly positioning and the interesting target in intelligent retrieval remote sensing images or region, are remote sensing big datas
Processing and analysis facing challenges and field of remote sensing image processing problem in science urgently to be resolved hurrily.Remote Sensing Image Retrieval technology
It is effective ways the bottleneck that solution, studies efficient image retrieval technologies and have great importance.
Current remote Sensing Image Retrieval technology is mainly to carry out similarity measurement by the low-level feature to image, and then return
Return similar image.Compared to traditional search method based on keyword, content-based retrieval method efficiency and accuracy higher,
But a kind of character description method that can effectively describe various complicated remote sensing images scenes of design is very difficult.In recent years, deep
Degree learns the research hotspot due to being increasingly becoming field of image recognition with good feature learning ability.Compared to engineer's
Feature, the method based on deep learning can obtain a feature extractor to realize that the automatic of characteristics of image carries by sample training
It takes, is suitable for inclusion in the remote Sensing Image Retrieval of complex scene.Since network design and training are relatively easy, sparse own coding has become
For a kind of common deep learning method, and it is widely used in image procossing.
Sparse autoencoder network is trained, in terms of constructing training sample, existing method is typically from training image
Certain amount and the image block construction training sample of size are randomly selected, this sample architecture method has following defect.First, from
For the angle of human eye vision theory, of people's attention is the specific objective on remote sensing images, and the image block randomly selected can
The specific objective of concern can and not be included.Second, since the size of training image is fixed, randomly select image block construction instruction
The method for practicing sample is likely to result in lack of training samples.Third utilizes trained network since training sample is image block
Obtained when feature extraction be image block rather than entire image feature, therefore cannot be directly used to image retrieval.It is terrible
To the feature of entire image, it usually needs the method for using convolution, but not only computational efficiency is low but also can introduce it for this process
His parameter.In terms of activation primitive selection, existing method generally use sigmoid functions swash as network hidden layer neuron
Function living, and the problems such as sigmoid functions disappear in network backpropagation there are serious gradient, it is unfavorable for network training.
For sparse autoencoder network feature extraction, existing method be typically directly using the activation value of hidden layer as the feature extracted and
Do not pass through LS-SVM sparseness, and tests and show that sparse features performance is more preferable.
Invention content
In view of the deficiencies of the prior art, the present invention provides a kind of distant based on notable point feature and sparse own coding
Feel image retrieval technologies scheme.The present invention will extract the notable point features of remote sensing images as the input of sparse autoencoder network into
And it is trained, it is final to realize remote Sensing Image Retrieval using the feature extractor extraction characteristics of image of training.
The technical solution adopted in the present invention is a kind of remote Sensing Image Retrieval based on notable point feature and sparse own coding
Method includes the following steps：
Step 1, the characteristic point for extracting each image in image library obtains characteristic point matrix, and is calculated using visual attention model
The notable figure of each image；
Step 2, for the notable figure of each image in image library, Adaptive Thresholding is respectively adopted by notable figure binaryzation,
And characteristic point matrix corresponding with image obtains filtered remarkable characteristic into line mask operation；Realization method is as follows,
When using Adaptive Thresholding by notable figure binaryzation, according to the conspicuousness size of notable figure pixel, determine notable
The binary-state threshold T of figure is as follows,
Wherein, w and h indicates that the width and height of notable figure, I (x, y) indicate the saliency value of notable figure pixel (x, y) respectively；
If obtaining binaryzation notable figure to notable figure binaryzation according to binary-state threshold T, should mutually there be matrix IbinaryIf P
Indicate the characteristic point matrix of image, PIIndicating filtered notable feature dot matrix, calculating notable feature dot matrix is as follows,
Step 3, it takes several images as training image from image library, several notable spies is chosen respectively from each training image
Sign point construction training sample, obtains training sample set X, and sparse autoencoder network is trained according to the training sample set X ' after albefaction,
Obtain feature extractor；
The sparse autoencoder network input layer, hidden layer and output layer, wherein hidden layer neuron use ReLU letters
Number is used as activation primitive, and output layer neuron is using softplus functions as activation primitive, the cost of sparse autoencoder network
Function is defined as follows,
Wherein, first item is mean square error item, and Section 2 is regular terms, HW,bIndicate the network output of training sample set X '
Value, W=[W1,W2] and b=[b1,b2] respectively indicate network input layer and implicit interlayer weight W1With biasing b1And hidden layer
With the weight W of output interlayer2With biasing b2The weight matrix of composition, λ indicate regularization coefficient；
Step 4, to all images in image library, feature extraction is carried out using the feature extractor of step 3 training gained,
It is used in combination threshold function table to carry out LS-SVM sparseness to the characteristics of image of extraction, obtains the final feature vector for retrieval；It realizes
Mode is as follows,
The characteristics of image Y of extraction indicates as follows,
Y=f1(W1PI′+b1)
Wherein, notable feature dot matrix PI' it is according to filtered notable feature dot matrix P obtained by step 2IAfter albefaction
As a result；
For the characteristics of image Y of extraction, carries out following LS-SVM sparseness and obtains sparse features matrix Z,
Z=[Z+,Z-Max]=[(0, Y- α), max (0, α-Y)]
Wherein, α indicates the threshold value of threshold function table, matrix Z+=max (0, Y- α), Z-=max (0, α-Y)；
If the SIFT point numbers detected from piece image are n, sparse features matrix Z is further processed, spy is obtained
It is as follows to levy vector F,
Wherein,WithRepresenting matrix Z respectively+And Z-I-th of column vector.
Step 5, the feature vector extracted based on step 4 carries out image retrieval according to preset similarity measurement criterion.
Moreover, in step 1, the characteristic point for extracting each image in image library obtains characteristic point matrix, is carried using SIFT operators
Take realization.
Moreover, in step 5, preset similarity measurement criterion uses city distance.
The present invention correspondingly provides a kind of Content-based Remote Sensing Image Retrieval System based on notable point feature and sparse own coding, including
With lower module,
Feature point extraction module, the characteristic point for extracting each image in image library obtains characteristic point matrix, and utilizes and regard
Feel that attention model calculates the notable figure of each image；
Adaptive threshold is respectively adopted for the notable figure for each image in image library in remarkable characteristic extraction module
Method is by notable figure binaryzation, and characteristic point matrix corresponding with image obtains filtered remarkable characteristic into line mask operation；
Realization method is as follows,
When using Adaptive Thresholding by notable figure binaryzation, according to the conspicuousness size of notable figure pixel, determine notable
The binary-state threshold T of figure is as follows,
Wherein, w and h indicates that the width and height of notable figure, I (x, y) indicate the saliency value of notable figure pixel (x, y) respectively；
If obtaining binaryzation notable figure to notable figure binaryzation according to binary-state threshold T, should mutually there be matrix IbinaryIf P
Indicate the characteristic point matrix of image, PIIndicating filtered notable feature dot matrix, calculating notable feature dot matrix is as follows,
Training module, for taking several images as training image from image library, if being chosen respectively from each training image
Dry remarkable characteristic constructs training sample, obtains training sample set X, sparse self-editing according to the training sample set X ' training after albefaction
Code network, obtains feature extractor；
The sparse autoencoder network input layer, hidden layer and output layer, wherein hidden layer neuron use ReLU letters
Number is used as activation primitive, and output layer neuron is using softplus functions as activation primitive, the cost of sparse autoencoder network
Function is defined as follows,
Wherein, first item is mean square error item, and Section 2 is regular terms, HW,bIndicate the network output of training sample set X '
Value, W=[W1,W2] and b=[b1,b2] respectively indicate network input layer and implicit interlayer weight W1With biasing b1And hidden layer
With the weight W of output interlayer2With biasing b2The weight matrix of composition, λ indicate regularization coefficient；
Characteristic extracting module, for all images in image library, using step 3 training gained feature extractor into
Row feature extraction is used in combination threshold function table to carry out LS-SVM sparseness to the characteristics of image of extraction, obtains the final spy for retrieval
Sign vector；Realization method is as follows,
The characteristics of image Y of extraction indicates as follows,
Y=f1(W1PI′+b1)
Wherein, notable feature dot matrix PI' it is according to filtered notable feature dot matrix P obtained by step 2IAfter albefaction
As a result；
For the characteristics of image Y of extraction, carries out following LS-SVM sparseness and obtains sparse features matrix Z,
Z=[Z+,Z-Max]=[(0, Y- α), max (0, α-Y)]
Wherein, α indicates the threshold value of threshold function table, matrix Z+=max (0, Y- α), Z-=max (0, α-Y)；
If the SIFT point numbers detected from piece image are n, sparse features matrix Z is further processed, spy is obtained
It is as follows to levy vector F,
Wherein,WithRepresenting matrix Z respectively+And Z-I-th of column vector.
Module is retrieved, for the feature vector of feature based extraction module extraction, according to preset similarity measurement criterion
Carry out image retrieval.
Moreover, in feature point extraction module, the characteristic point for extracting each image in image library obtains characteristic point matrix, utilizes
SIFT operator extractions are realized.
Moreover, in retrieval module, preset similarity measurement criterion uses city distance.
Compared with prior art, the present invention has following features and advantageous effect,
1, the notable figure of image, and the characteristic point that notable figure binaryzation extracts SIFT are calculated using visual attention model
It is filtered to obtain the remarkable characteristic of image, not only conforms with the vision attention feature of human eye and can preferably reflect people
Search Requirement.
2, the remarkable characteristic for choosing image constructs training sample, compensates for traditional grab sample structure on training image
Make the defect of training sample.
3, the feature extractor trained using sparse autoencoder network realizes automatically extracting for characteristics of image, exempts
For the characteristic Design process of complicated remote sensing images.
4, favorable expandability, training sample include but are not limited to remarkable characteristic.
Description of the drawings
Fig. 1 is the flow chart of the embodiment of the present invention.
Specific implementation mode
Remote sensing image retrieval method proposed by the present invention based on notable point feature and sparse own coding extracts image first
Characteristic point obtain characteristic point matrix, and calculate the notable figure of image, then use adaptive threshold by notable figure binaryzation with
Characteristic point matrix carries out " mask " operation and obtains remarkable characteristic, then chooses the remarkable characteristic construction training sample of certain amount
The sparse autoencoder network of this training, and automatically extract characteristics of image using trained feature extractor and obtain the feature for retrieval
Vector finally carries out image retrieval according to preset method for measuring similarity and returns to similar image.
For the present invention will be described in detail, technical solution provides embodiment flow and is described as follows referring to Fig. 1：
Step 1, the characteristic point for extracting each image in image library obtains characteristic point matrix, and is calculated using visual attention model
The notable figure of each image.
When it is implemented, existing image library or the image library voluntarily built by those skilled in the art may be used.Example
The high-resolution remote sensing image that a width includes multiple atural object classifications is such as chosen, cutting structure packet is carried out using Tiles partitioned modes
Retrieval image library containing multiple classifications.For every piece image in image library, embodiment uses SIFT (Scale first
Invariant Feature Transform) characteristic point (key point) of operator extraction image obtains characteristic point matrix, then adopts
The notable figure of image, Tile method of partition, SIFT operators are calculated with GBVS (Graph-Based Visual Saliency) model
And GBVS models are the prior arts, it will not go into details by the present invention.
Step 2, for the notable figure of each image in image library, Adaptive Thresholding is respectively adopted by notable figure binaryzation,
And characteristic point matrix corresponding with image carries out " mask " operation and obtains filtered remarkable characteristic.
The binary-state threshold of notable figure, binaryzation notable figure and feature are determined in embodiment according to the conspicuousness size of pixel
Dot matrix obtains remarkable characteristic after carrying out " mask " operation, realizes as follows：
According to the conspicuousness size of notable figure pixel, the binary-state threshold T of notable figure is determined by formula (1).
Wherein, w and h indicates that the width and height of notable figure, I (x, y) indicate the saliency value of pixel at notable figure (x, y) respectively.
According to binary-state threshold T to notable figure binaryzation, binaryzation notable figure is obtained, should mutually there is matrix Ibinary.Utilize two
Value notable figure is filtered to obtain remarkable characteristic to the characteristic point matrix of image.Let p represent the characteristic point matrix of image, PI
Indicate filtered notable feature dot matrix, then notable feature dot matrix can be calculated by formula (2).
Wherein,
Matrix
The corresponding feature vector of one SIFT key point of each element representation of matrix P, and SIFT key points are corresponding
Feature vector is usually 128 dimensions, and the embodiment of the present invention accordingly uses 128 dimensions；
Matrix
Wherein, P128(x, y) indicates the corresponding feature vector of characteristic point, if pixel does not have characteristic point P at (x, y)128(x,
Y)=0.IbinaryMiddle each element is 0 or 1, Ibinary(x, y) indicates value of the binaryzation notable figure at (x, y).Symbol
For scale multiplication symbol.
Step 3, several images are chosen from image library as training image, are chosen respectively from each training image several notable
Characteristic point constructs training sample, and the sparse autoencoder network of training obtains feature extractor.
In embodiment, in step 3 choose certain amount training image remarkable characteristic and unconventional image block structure
Make training sample, when training selects ReLU (Rectified Linear Units) functions and unconventional sigmoid functions are made
For the activation primitive of sparse autoencoder network hidden layer neuron.For example, in step 3 each remarkable characteristic be one 4 × 4 ×
The feature vector of 8=128 dimensions, a characteristic point constitute a training sample.When it is implemented, the number of training image, a width
The number of remarkable characteristic can voluntarily be specified by those skilled in the art in training image.
It is implemented as follows：
First, the remarkable characteristic of image is chosen, training sample set is constructed.
Embodiment randomly selects the image of certain amount as training image from image library first, then randomly selects one
The remarkable characteristic of fixed number purpose training image constructs training sample set.Training sample set can use formula (3) to indicate：
Wherein, m indicates that the number of training sample, each row of X indicate a remarkable characteristic, i.e. a training sample.
For example, [x1,1,x2,1,…,x128,1] it is the 1st training sample, [x1,2,x2,2,…,x128,2] it is the 2nd training sample.
Then, the sparse autoencoder network of training obtains feature extractor.
There are certain correlations for the remarkable characteristic extracted due to same width training image, cannot be directly training
Sample set X inputs sparse autoencoder network and is trained.ZCA (Zero Component Analysis) albefaction is used before training
It is handled to obtain the training sample set X ' after albefaction to training sample, and preserves relevant parameter when ZCA albefactions, ZCA albefactions
It is embodied as the prior art, it will not go into details by the present invention.
Embodiment defines a sparse autoencoder network comprising input layer, 3 layers of hidden layer and output layer, wherein hidden
Neuron containing layer uses ReLU functions f1=max (0, x) is used as activation primitive, output layer neuron to use softplus functions f2
=ln (1+ex) it is used as activation primitive.Compared to traditional sigmoid functions, ReLU functions can alleviate gradient disappearance to a certain extent
Problem is more conducive to network training.Training sample set X ' is given, then the cost function of sparse autoencoder network may be defined as formula (4).
First item is mean square error item in formula, and Section 2 is regular terms, HW,bIndicate the network output of training sample set X '
Value, W=[W1,W2] and b=[b1,b2] respectively indicate network input layer and implicit interlayer weight W1With biasing b1And hidden layer
With the weight W of output interlayer2With biasing b2The weight matrix of composition, λ indicate regularization coefficient.When it is implemented, can be adopted when training
Weight and bias matrix parameter W and b are obtained with the cost function in the methods of gradient decline optimized-type (4).
Step 4, to all images in image library, feature extraction is carried out using the feature extractor of step 3 training gained,
It is used in combination threshold function table to carry out LS-SVM sparseness to the feature of extraction, obtains the final feature vector for retrieval.
It is mapped the remarkable characteristic input feature vector extractor of image to obtain corresponding image in the step 4 of embodiment
Feature, recycle threshold function table to the feature of extraction carry out LS-SVM sparseness can be obtained the final feature for retrieval to
Amount.
The characteristics of image Y of extraction can use formula (5) to indicate as follows,
Y=f1(W1PI′+b1) (5)
Wherein, by W1PI+b1ReLU functions f is substituted into as variable x1=max (0, x), remarkable characteristic square used herein
Battle array PI' it is according to filtered notable feature dot matrix obtained by step 2, using identical as when carrying out albefaction to training sample set X
ZCA whitening parameters carry out pretreated result.For the characteristics of image Y of extraction, with formula (6) carry out LS-SVM sparseness obtain it is dilute
Dredge eigenmatrix Z.
Z=[Z+,Z-Max]=[(0, Y- α), max (0, α-Y)] (6)
Wherein, α indicates the threshold value of threshold function table f=max (0, x- α) and f=max (0, α-Y), matrix Z+=max (0, Y-
α), Z-=max (0, α-Y).
The final feature vector F for retrieval in order to obtain, if the SIFT point numbers detected from piece image are n
It is a, sparse features matrix Z is further processed with formula (7).
Wherein,WithRepresenting matrix Z respectively+And Z-I-th of column vector.
Step 5, the feature vector extracted based on step 4 carries out image retrieval according to preset similarity measurement criterion：Tool
When body is implemented, those skilled in the art can voluntarily preset similarity measurement criterion.Embodiment is counted using city distance (L1 norms)
The similitude of query image and other images is calculated, and associated picture is returned by similitude size.When it is implemented, can be with image library
Middle any image is query image, obtains the associated picture returned by similitude size, to other images other than image library,
Same mode may be used and extract feature vector, and retrieved from image library.
When it is implemented, computer software mode, which can be used, in the above flow realizes automatic running flow, mould can also be used
Block mode provides corresponding system.The present invention correspondingly provides a kind of remote sensing images based on notable point feature and sparse own coding
Searching system comprises the following modules,
Feature point extraction module, the characteristic point for extracting each image in image library obtains characteristic point matrix, and utilizes and regard
Feel that attention model calculates the notable figure of each image；
Adaptive threshold is respectively adopted for the notable figure for each image in image library in remarkable characteristic extraction module
Method is by notable figure binaryzation, and characteristic point matrix corresponding with image obtains filtered remarkable characteristic into line mask operation；
Realization method is as follows,
When using Adaptive Thresholding by notable figure binaryzation, according to the conspicuousness size of notable figure pixel, determine notable
The binary-state threshold T of figure is as follows,
Wherein, w and h indicates that the width and height of notable figure, I (x, y) indicate the saliency value of notable figure pixel (x, y) respectively；
If obtaining binaryzation notable figure to notable figure binaryzation according to binary-state threshold T, should mutually there be matrix IbinaryIf P
Indicate the characteristic point matrix of image, PIIndicating filtered notable feature dot matrix, calculating notable feature dot matrix is as follows,
Training module, for taking several images as training image from image library, if being chosen respectively from each training image
Dry remarkable characteristic constructs training sample, obtains training sample set X, sparse self-editing according to the training sample set X ' training after albefaction
Code network, obtains feature extractor；
The sparse autoencoder network input layer, hidden layer and output layer, wherein hidden layer neuron use ReLU letters
Number is used as activation primitive, and output layer neuron is using softplus functions as activation primitive, the cost of sparse autoencoder network
Function is defined as follows,
Wherein, first item is mean square error item, and Section 2 is regular terms, HW,bIndicate the network output of training sample set X '
Value, W=[W1,W2] and b=[b1,b2] respectively indicate network input layer and implicit interlayer weight W1With biasing b1And hidden layer
With the weight W of output interlayer2With biasing b2The weight matrix of composition, λ indicate regularization coefficient；
Query characteristics extraction module, for image to be checked, being carried out using the feature extractor of step 3 training gained
Feature extraction is used in combination threshold function table to carry out LS-SVM sparseness to the characteristics of image of extraction, obtains the final feature for retrieval
Vector；Realization method is as follows,
The characteristics of image Y of extraction indicates as follows,
Y=f1(W1PI′+b1)
Wherein, notable feature dot matrix PI' it is according to filtered notable feature dot matrix P obtained by step 2IAfter albefaction
As a result；
For the characteristics of image Y of extraction, carries out following LS-SVM sparseness and obtains sparse features matrix Z,
Z=[Z+,Z-Max]=[(0, Y- α), max (0, α-Y)]
Wherein, α indicates the threshold value of threshold function table, matrix Z+=max (0, Y- α), Z-=max (0, α-Y)；
If the SIFT point numbers detected from piece image are n, sparse features matrix Z is further processed, spy is obtained
It is as follows to levy vector F,
Wherein,WithRepresenting matrix Z respectively+And Z-I-th of column vector.
Module is retrieved, the feature vector for being extracted based on query characteristics extraction module, according to preset similarity measurement
Criterion carries out image retrieval.
Specific embodiment described herein is given an example to the present invention.The skill of the technical field of the invention
Art personnel can do various modifications or supplement to described specific embodiment or substitute by a similar method, but can't be inclined
Spirit or beyond the scope defined by the appended claims from the present invention.
Claims (6)
1. a kind of remote sensing image retrieval method based on notable point feature and sparse own coding, it is characterised in that：Including following step
Suddenly,
Step 1, the characteristic point for extracting each image in image library obtains characteristic point matrix, and calculates each figure using visual attention model
The notable figure of picture；
Step 2, for the notable figure of each image in image library, Adaptive Thresholding is respectively adopted by notable figure binaryzation, and with
The corresponding characteristic point matrix of image obtains filtered remarkable characteristic into line mask operation；Realization method is as follows,
When using Adaptive Thresholding by notable figure binaryzation, according to the conspicuousness size of notable figure pixel, notable figure is determined
Binary-state threshold T is as follows,
Wherein, w and h indicates that the width and height of notable figure, I (x, y) indicate the saliency value of notable figure pixel (x, y) respectively；
If obtaining binaryzation notable figure to notable figure binaryzation according to binary-state threshold T, should mutually there be matrix Ibinary, let p represent
The characteristic point matrix of image, PIIndicating filtered notable feature dot matrix, calculating notable feature dot matrix is as follows,
Step 3, it takes several images as training image from image library, several remarkable characteristics is chosen respectively from each training image
Training sample is constructed, training sample set X is obtained, according to the sparse autoencoder network of training sample set X ' training after albefaction, is obtained
Feature extractor；
The sparse autoencoder network input layer, hidden layer and output layer, wherein hidden layer neuron are made using ReLU functions
For activation primitive, output layer neuron is using softplus functions as activation primitive, the cost function of sparse autoencoder network
It is defined as follows,
Wherein, first item is mean square error item, and Section 2 is regular terms, HW,bIndicate the network output valve of training sample set X ', W
=[W1,W2] and b=[b1,b2] respectively indicate network input layer and implicit interlayer weight W1With biasing b1And hidden layer and defeated
Go out the weight W of interlayer2With biasing b2The weight matrix of composition, λ indicate regularization coefficient；
Step 4, to all images in image library, feature extraction is carried out using the feature extractor of step 3 training gained, is used in combination
Threshold function table carries out LS-SVM sparseness to the characteristics of image of extraction, obtains the final feature vector for retrieval；Realization method
It is as follows,
The characteristics of image Y of extraction indicates as follows,
Y=f1(W1PI′+b1)
Wherein, f1() is ReLU functions, notable feature dot matrix PI' it is according to filtered notable feature dot matrix obtained by step 2
PIResult after albefaction；
For the characteristics of image Y of extraction, carries out following LS-SVM sparseness and obtains sparse features matrix Z,
Z=[Z+,Z-Max]=[(0, Y- α), max (0, α-Y)]
Wherein, α indicates the threshold value of threshold function table, matrix Z+=max (0, Y- α), Z-=max (0, α-Y)；
If the SIFT point numbers detected from piece image are n, sparse features matrix Z is further processed, obtain feature to
It is as follows to measure F,
Wherein,WithRepresenting matrix Z respectively+And Z-I-th of column vector；
Step 5, the feature vector extracted based on step 4 carries out image retrieval according to preset similarity measurement criterion.
2. the remote sensing image retrieval method based on notable point feature and sparse own coding, feature exist according to claim 1
In：In step 1, the characteristic point for extracting each image in image library obtains characteristic point matrix, is realized using SIFT operator extractions.
3. the remote sensing image retrieval method according to claim 1 or claim 2 based on notable point feature and sparse own coding, feature
It is：In step 5, preset similarity measurement criterion uses city distance.
4. a kind of Content-based Remote Sensing Image Retrieval System based on notable point feature and sparse own coding, it is characterised in that：Including with lower die
Block,
Feature point extraction module, the characteristic point for extracting each image in image library obtains characteristic point matrix, and is noted using vision
Meaning model calculates the notable figure of each image；
Remarkable characteristic extraction module, for the notable figure for each image in image library, Adaptive Thresholding is respectively adopted will
Notable figure binaryzation, and characteristic point matrix corresponding with image obtains filtered remarkable characteristic into line mask operation；It realizes
Mode is as follows,
When using Adaptive Thresholding by notable figure binaryzation, according to the conspicuousness size of notable figure pixel, notable figure is determined
Binary-state threshold T is as follows,
Wherein, w and h indicates that the width and height of notable figure, I (x, y) indicate the saliency value of notable figure pixel (x, y) respectively；
If obtaining binaryzation notable figure to notable figure binaryzation according to binary-state threshold T, should mutually there be matrix Ibinary, let p represent
The characteristic point matrix of image, PIIndicating filtered notable feature dot matrix, calculating notable feature dot matrix is as follows,
Training module is chosen several aobvious respectively for taking several images as training image from image library from each training image
It writes characteristic point and constructs training sample, obtain training sample set X, according to the sparse own coding net of training sample set X ' training after albefaction
Network obtains feature extractor；
The sparse autoencoder network input layer, hidden layer and output layer, wherein hidden layer neuron are made using ReLU functions
For activation primitive, output layer neuron is using softplus functions as activation primitive, the cost function of sparse autoencoder network
It is defined as follows,
Wherein, first item is mean square error item, and Section 2 is regular terms, HW,bIndicate the network output valve of training sample set X ', W
=[W1,W2] and b=[b1,b2] respectively indicate network input layer and implicit interlayer weight W1With biasing b1And hidden layer and defeated
Go out the weight W of interlayer2With biasing b2The weight matrix of composition, λ indicate regularization coefficient；
Characteristic extracting module, for all images in image library, being carried out using the feature extractor of step 3 training gained special
Sign extraction, is used in combination threshold function table to carry out LS-SVM sparseness to the characteristics of image of extraction, obtain the final feature for retrieval to
Amount；Realization method is as follows,
The characteristics of image Y of extraction indicates as follows,
Y=f1(W1PI′+b1)
Wherein, f1() is ReLU functions, notable feature dot matrix PI' it is according to filtered notable feature dot matrix obtained by step 2
PIResult after albefaction；
For the characteristics of image Y of extraction, carries out following LS-SVM sparseness and obtains sparse features matrix Z,
Z=[Z+,Z-Max]=[(0, Y- α), max (0, α-Y)]
Wherein, α indicates the threshold value of threshold function table, matrix Z+=max (0, Y- α), Z-=max (0, α-Y)；
If the SIFT point numbers detected from piece image are n, sparse features matrix Z is further processed, obtain feature to
It is as follows to measure F,
Wherein,WithRepresenting matrix Z respectively+And Z-I-th of column vector；
Module is retrieved, the feature vector for being extracted based on query characteristics extraction module, according to preset similarity measurement criterion
Carry out image retrieval.
5. the Content-based Remote Sensing Image Retrieval System based on notable point feature and sparse own coding, feature exist according to claim 4
In：In feature point extraction module, the characteristic point for extracting each image in image library obtains characteristic point matrix, utilizes SIFT operator extractions
It realizes.
6. the Content-based Remote Sensing Image Retrieval System based on notable point feature and sparse own coding according to claim 4 or 5, feature
It is：It retrieves in module, preset similarity measurement criterion uses city distance.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510708598.4A CN105243154B (en) | 2015-10-27 | 2015-10-27 | Remote sensing image retrieval method based on notable point feature and sparse own coding and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510708598.4A CN105243154B (en) | 2015-10-27 | 2015-10-27 | Remote sensing image retrieval method based on notable point feature and sparse own coding and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105243154A CN105243154A (en) | 2016-01-13 |
| CN105243154B true CN105243154B (en) | 2018-08-21 |
Family
ID=55040802
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510708598.4A Active CN105243154B (en) | 2015-10-27 | 2015-10-27 | Remote sensing image retrieval method based on notable point feature and sparse own coding and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105243154B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718531B (en) * | 2016-01-14 | 2019-12-17 | 广州市万联信息科技有限公司 | Image database establishing method and image identification method |
| CN106228130B (en) * | 2016-07-19 | 2019-09-10 | 武汉大学 | Remote sensing image cloud detection method of optic based on fuzzy autoencoder network |
| CN106295613A (en) * | 2016-08-23 | 2017-01-04 | 哈尔滨理工大学 | A kind of unmanned plane target localization method and system |
| CN106909924B (en) * | 2017-02-18 | 2020-08-28 | 北京工业大学 | Remote sensing image rapid retrieval method based on depth significance |
| CN107122809B (en) * | 2017-04-24 | 2020-04-28 | 北京工业大学 | Neural network feature learning method based on image self-coding |
| CN107515895B (en) * | 2017-07-14 | 2020-06-05 | 中国科学院计算技术研究所 | Visual target retrieval method and system based on target detection |
| CN108830172A (en) * | 2018-05-24 | 2018-11-16 | 天津大学 | Aircraft remote sensing images detection method based on depth residual error network and SV coding |
| CN109259733A (en) * | 2018-10-25 | 2019-01-25 | 深圳和而泰智能控制股份有限公司 | Apnea detection method, apparatus and detection device in a kind of sleep |
| CN111144483B (en) * | 2019-12-26 | 2023-10-17 | 歌尔股份有限公司 | Image feature point filtering method and terminal |
| CN112731410B (en) * | 2020-12-25 | 2021-11-05 | 上海大学 | Underwater target sonar detection method based on CNN |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102073748A (en) * | 2011-03-08 | 2011-05-25 | 武汉大学 | Visual keyword based remote sensing image semantic searching method |
| CN102867196A (en) * | 2012-09-13 | 2013-01-09 | 武汉大学 | Method for detecting complex sea-surface remote sensing image ships based on Gist characteristic study |
| CN103309982A (en) * | 2013-06-17 | 2013-09-18 | 武汉大学 | Remote sensing image retrieval method based on vision saliency point characteristics |
| CN104462494A (en) * | 2014-12-22 | 2015-03-25 | 武汉大学 | Remote sensing image retrieval method and system based on non-supervision characteristic learning |
-
2015
- 2015-10-27 CN CN201510708598.4A patent/CN105243154B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102073748A (en) * | 2011-03-08 | 2011-05-25 | 武汉大学 | Visual keyword based remote sensing image semantic searching method |
| CN102867196A (en) * | 2012-09-13 | 2013-01-09 | 武汉大学 | Method for detecting complex sea-surface remote sensing image ships based on Gist characteristic study |
| CN103309982A (en) * | 2013-06-17 | 2013-09-18 | 武汉大学 | Remote sensing image retrieval method based on vision saliency point characteristics |
| CN104462494A (en) * | 2014-12-22 | 2015-03-25 | 武汉大学 | Remote sensing image retrieval method and system based on non-supervision characteristic learning |
Non-Patent Citations (2)
| Title |
|---|
| 利用视觉注意模型和局部特征的遥感影像检索方法;周维勋 等;《武汉大学学报.信息科学版》;20150105;第40卷(第1期);第46-52页 * |
| 基于视觉显著点特征的遥感影像检索方法;王星 等;《测绘科学》;20140420;第39卷(第4期);第34-38页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105243154A (en) | 2016-01-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105243154B (en) | Remote sensing image retrieval method based on notable point feature and sparse own coding and system | |
| CN108804530B (en) | Subtitling areas of an image | |
| CN110598029B (en) | Fine-grained image classification method based on attention transfer mechanism | |
| Yuan et al. | Exploring a fine-grained multiscale method for cross-modal remote sensing image retrieval | |
| Yi et al. | ASSD: Attentive single shot multibox detector | |
| CN112750140B (en) | Information mining-based disguised target image segmentation method | |
| CN107818314B (en) | Face image processing method, device and server | |
| CN110222140A (en) | A kind of cross-module state search method based on confrontation study and asymmetric Hash | |
| CN110309856A (en) | Image classification method, the training method of neural network and device | |
| CN109344821A (en) | Small target detecting method based on Fusion Features and deep learning | |
| CN104462494B (en) | A kind of remote sensing image retrieval method and system based on unsupervised feature learning | |
| CN106909924A (en) | A kind of remote sensing image method for quickly retrieving based on depth conspicuousness | |
| CN109559300A (en) | Image processing method, electronic equipment and computer readable storage medium | |
| CN107256246A (en) | PRINTED FABRIC image search method based on convolutional neural networks | |
| CN106227851A (en) | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end | |
| CN109508675B (en) | Pedestrian detection method for complex scene | |
| CN110929080B (en) | Optical remote sensing image retrieval method based on attention and generation countermeasure network | |
| CN105243139A (en) | Deep learning based three-dimensional model retrieval method and retrieval device thereof | |
| CN105868706A (en) | Method for identifying 3D model based on sparse coding | |
| CN114758362B (en) | Clothing changing pedestrian re-identification method based on semantic perception attention and visual shielding | |
| CN107958067A (en) | It is a kind of based on without mark Automatic Feature Extraction extensive electric business picture retrieval system | |
| CN111666919A (en) | Object identification method and device, computer equipment and storage medium | |
| CN112597324A (en) | Image hash index construction method, system and equipment based on correlation filtering | |
| CN114842208A (en) | Power grid harmful bird species target detection method based on deep learning | |
| Termritthikun et al. | NU-LiteNet: Mobile landmark recognition using convolutional neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |