CN110334765A - Remote Image Classification based on the multiple dimensioned deep learning of attention mechanism - Google Patents

Remote Image Classification based on the multiple dimensioned deep learning of attention mechanism Download PDF

Info

Publication number
CN110334765A
CN110334765A CN201910603799.6A CN201910603799A CN110334765A CN 110334765 A CN110334765 A CN 110334765A CN 201910603799 A CN201910603799 A CN 201910603799A CN 110334765 A CN110334765 A CN 110334765A
Authority
CN
China
Prior art keywords
neural networks
convolutional neural
remote sensing
training
loss
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910603799.6A
Other languages
Chinese (zh)
Other versions
CN110334765B (en
Inventor
唐旭
马秋硕
马晶晶
焦李成
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian University of Electronic Science and Technology
Original Assignee
Xian University of Electronic Science and Technology
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 Xian University of Electronic Science and Technology filed Critical Xian University of Electronic Science and Technology
Priority to CN201910603799.6A priority Critical patent/CN110334765B/en
Publication of CN110334765A publication Critical patent/CN110334765A/en
Application granted granted Critical
Publication of CN110334765B publication Critical patent/CN110334765B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • 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
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • 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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

The invention discloses a kind of Remote Image Classification based on the multiple dimensioned deep learning of attention mechanism, mainly solve the problems, such as that prior art classification accuracy is low.Its scheme is: establishing remote sensing images library and the corresponding classification of image library, and selects 80% remote sensing images sample building training image library at random from every class remote sensing images after normalized；Building one includes convolutional network module, the convolutional neural networks for paying attention to power module, SCDA module and full articulamentum；Training sample in training image library is input to the classification results that convolutional neural networks obtain training sample, and determines the loss function of convolutional neural networks；Loss function iteration is updated until penalty values stabilization by gradient descent method, obtains trained convolutional neural networks；By remote sensing image to be sorted after normalizing, it is input to trained convolutional neural networks and obtains classification results；Nicety of grading of the present invention is high, and strong robustness can be applied to the analysis and management of remote sensing image data.

Description

Remote Image Classification based on the multiple dimensioned deep learning of attention mechanism
Technical field
The invention belongs to technical field of image processing, in particular to a kind of Remote Sensing Images classification method can be applied to The analysis and management of remote sensing image data.
Background technique
As satellite remote sensing images and aerial remote sensing images resolution ratio are continuously improved, can be obtained from remote sensing images more Useful data and information.And it is directed to the application of different occasions, also there is different requirements to the processing of remote sensing images, so being Effectively these remote sensing image datas are analyzed and managed, need to stick semantic label to image according to picture material. And scene classification is exactly a kind of important channel for solving the problems, such as such.What scene classification referred to is exactly to distinguish to provide from multiple image There is the image of similar scene feature, and correctly classifies to these images.Compared to natural image, remote sensing images itself have The characteristics of there is itself, it is limitation and jljl different spectrum of the classification results due to the spatial resolution of remote sensing images itself, different Object usually will cause the phenomenon that mistake is divided, this is as caused by the complexity of remote sensing images itself with the presence of spectrum phenomenon.Therefore, How classification more accurately to be carried out to remote sensing images and also becomes a challenge.
Classification based on convolutional neural networks, some pictures of training will be needed by referring to, input convolutional Neural in batches In network, by the repetition training of high-volume data, so that objective optimization loss function reduces.To realize the mesh of classification 's.Nowadays many more mature, famous convolutional Neurals are suggested.Such as 2012, AlexKrizhevsky was just proposed A kind of depth convolutional network model " AlexNet ".
Although existing convolutional neural networks can be realized the task of picture scene classification, but in study picture semantic letter It is insufficient of both being still had when breath, first is that the classification information position inaccurate as caused by remote sensing images complexity, two It is that volume neural network can usually fall into local marking area in training, as shown in Figure 1.The two deficiencies will lead in actual field There are problems that poor robustness during the classification of scape and is easy to generate mistake and divides.
Summary of the invention
Present invention aims at above-mentioned prior art there are aiming at the problem that, propose a kind of multiple dimensioned based on attention mechanism Classifying Method in Remote Sensing Image expand convolutional network to reduce the probability that remote sensing image classification target falls into regional area and pay attention to The classification accuracy of remote sensing images is improved in power region.
Technical thought of the invention is: the convolution feature of picture is obtained using convolutional neural networks, according to attention mechanism Principle is obtained the useful information for being conducive to classification using attention mechanism, multiple dimensioned convolutional layer feature is extracted from useful information, And pass through full articulamentum network implementations image classification.
According to above-mentioned thinking, realization step of the invention includes the following:
(1) remote sensing images library { I is established1,I2,…In…,IN, the corresponding classification of image library is { Y1,Y2,…Yn…,YN}, And the remote sensing images library of foundation is normalized, wherein n-th of sample number in n representative image library, n ∈ [0, N], N Represent the number of pictures in remote sensing images library；
(2) it selects 80% sample at random from every class image after normalized, constructs training image library { T1, T2,…Tj…,TM, wherein M < N, wherein TjJ-th of picture in training image library is represented, j ∈ [0, M], M are the total of training sample Number；
(3) constructing one includes convolutional network module, the convolutional Neural net for paying attention to power module, SCDA module and full articulamentum Network；
(4) loss function of convolutional neural networks is determined:
(4a) is by training image library { T1,T2,…Tj…,TMIt is input to the convolutional layer neural network with pre-training weight, Export the last layer feature F of convolutional layer；
The last layer feature F is input to the attention power module of convolutional neural networks by (4b), exports convolution feature A, then will Convolutional layer feature A is input to multiple SCDA modules that convolutional neural networks have different average thresholds, and output T group mask convolution is special Sign: { M1,M2,…MT, wherein T is the number of SCDA module；
T group mask convolution feature is input to the full articulamentum of convolutional neural networks by (4c) by global average Chi Huahou, The classification results for exporting training data, obtain the loss function of convolutional neural networks:
Wherein, loss1For the cross entropy of output category result and actual result, loss2For T group mask convolution feature process After full articulamentum the absolute value of output category result and actual result cross entropy and,For convolutional neural networks weight vectors L2 norm, λr、λs, η be respectively loss1, loss2,Hyper parameter；
(5) setting the number of iterations is P, training is iterated to convolutional neural networks by gradient decline optimization, until damage Lose functionDo not decline or exercise wheel number reaches the number of iterations, obtains trained convolutional neural networks；
(6) user is input in trained convolutional neural networks after normalizing remote sensing figure I' picture to be sorted, obtains To classification results, picture classification is completed.
The present invention has the advantage that compared with prior art
1, the present invention can be quickly found more apparent special in remote sensing images due to being based on attention mechanism principle The feature for being used to classify more is concentrated on a certain region with obvious semantic information, enhances remote sensing images scene point by sign Class accuracy；
2, the present invention uses SCDA module, expands the impression visual field of convolutional neural networks, reduces remote sensing image classification Target falls into the probability of regional area, enhances the accuracy and robustness of remote sensing image classification；
3, the present invention devises loss function, further clarifies classification task, improves the accuracy of remote sensing image classification.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention；
Fig. 2 is the convolutional neural networks structure chart constructed in the present invention；
Fig. 3 is the remote sensing images master drawing that present invention emulation uses.
Specific embodiment
The embodiment of the present invention effect is described in further detail below in conjunction with attached drawing.
Referring to Fig.1, steps are as follows for realization of the invention:
Step 1, remote sensing images library is established, training sample and test sample are obtained.
UC Merced image 1a) is downloaded from the official website weegee, establishes remote sensing images library { I1,I2,…In…,IN, image The corresponding classification in library is { Y1,Y2,…Yn…,YN, wherein InN-th image in representative image library, YnN-th in representative image library The corresponding classification of image, n-th of sample number in n representative image library, n ∈ [0, N]；
1b) the remote sensing images library of foundation is normalized according to following formula:
Wherein VmaxFor the point maximum value of all pixels in remote sensing images library, VminFor the point of all pixels in remote sensing images library Minimum value, { I'1,I'2,…I'n…,I'NBe normalized after remote sensing images library, I'nFor image set after normalized N-th of sample, n ∈ [0, N], N represent normalization after remote sensing images library number of pictures；
The remote sensing images for 1c) selecting 80% at random from every class image in remote sensing images library after normalized are used as training Sample set { T1,T2,…Tj…,TM, using remaining 20% remote sensing images as test sample collection { T1,T2,…Td…,Tm, Middle TjIndicate j-th of sample in training sample, j ∈ [0, M], tdIndicate d-th of sample in test sample, d ∈ [0, m], M are The total number of training sample, m are the total number of test sample, m < N, M < N.
Step 2, convolutional neural networks are constructed.
Referring to Fig. 2, this step is accomplished by
Convolutional network module 2a) is set, and the module is by five convolutional layers sequentially connected in pre-training AlexNet network { conv1, conv2, conv3, conv4, conv5 } is constituted；
2b) setting pay attention to power module, the module by an overall situation be averaged pond layer, the first full articulamentum, Relu active coating, Second full articulamentum and Sigmoid function composition, structure are as shown in Figure 3；
The global average pond layer: its convolution feature sizes inputted is W × H × C, for the convolution to C W × H It is averaging, exports the convolution feature of 1 × 1 × C；
The first full articulamentum: its convolution kernel is dimensioned to 1 × 1 × C'/16, wherein C' is that input first connects entirely The characteristic dimension of layer；
The second full articulamentum: its convolution kernel is dimensioned to 1 × 1 × C ", wherein C " is that input second connects layer entirely Characteristic dimension；
The Relu activation primitive and Sigmoid activation primitive are respectively as follows:
Wherein x is the input function of Relu activation primitive, and x' is the input function of Sigmoid activation primitive；
SCDA module 2c) is set, for exporting convolution mask feature
Referring to Fig. 2, the working principle of the SCDA module is as follows:
Notice that the Three dimensional convolution feature of the second full articulamentum output of power module inputs to SCDA module and carries out third dimension Degree summation, obtains two-dimensional convolution feature, averages to obtained two-dimensional convolution feature；
Convolution mask is carried out according to average value, i.e., is compared the value at two-dimensional convolution characteristic strong point with average value, if The value at two-dimensional convolution characteristic strong point is greater than average value, then is encoded to 1, if the value of the data point of two-dimensional convolution feature is less than averagely Value, then be encoded to 0, obtain convolution mask；
Convolution mask feature is extracted, i.e., convolution mask is multiplied with the average value threshold value E of setting, and to the convolution after multiplication Mask data value adds 1, then is multiplied with the Three dimensional convolution feature of input SCDA module, obtains mask feature；Before mask feature Two dimensions are averaged, and are obtained convolution mask feature and are exported；
Full articulamentum 2d) is set, the full articulamentum successively by convolution kernel size be respectively 512 × 1024,102 4 × 1024,1024 × 21 three convolution kernels composition；
2e) the convolutional network module of above-mentioned setting, attention power module, SCDA module and full articulamentum are sequentially connected, obtained To convolutional neural networks.
Step 3 determines the loss function of convolutional neural networks:
3a) by training sample set { T1,T2,…Tj…,TMIt is input to the convolutional network module of convolutional layer neural network, it is defeated The last layer feature F of convolutional layer out；
The last layer feature F 3b) is input to the attention power module of convolutional neural networks, exports convolution feature A, then will volume Lamination feature A is input to multiple SCDA modules that convolutional neural networks have different average thresholds, and output T group mask convolution is special Sign: { M1,M2,…MT, wherein T is the number of SCDA module；
T group mask convolution feature 3c) is input to the full articulamentum of convolutional neural networks, exports the classification knot of training data Fruit obtains the loss function loss of convolutional neural networksop:
Wherein:
For the L2 norm of convolutional neural networks weight vectors, λr、λs, η be respectively loss1, loss2,It is super Parameter；
Indicate the cross entropy of output category result and actual result, yjFor in training image library TjPrediction category probability, ojFor T in training image libraryjPractical category；
Indicate T group mask convolution feature output category result after full articulamentum With the absolute value of actual result cross entropy and, T is that SCDA module says wood, lossmFor T in training image libraryjIn m convolution mask Under feature, lossnFor T in training image libraryjLoss under the n-th convolution mask feature1。
Step 4, training is iterated to convolutional neural networks.
Being iterated trained existing method to convolutional neural networks has gradient optimization algorithm, Nesterov gradient to add Fast method, Adagrad method, the present invention use but are not limited to gradient descent algorithm, and implementation step is as follows:
4a) setting the number of iterations is P, and it is L, attenuation rate β that trained initial learning rate, which is arranged, by training image library { T1, T2,…Tj…,TMBe divided into the convolutional neural networks for being input to step 2 building for G times, the number of pictures Q inputted every time are as follows:
Wherein M is the total number of training image library sample；
4b) set the corresponding learning rate l of input picture every time are as follows:
L=L* βG；
The update of G subparameter 4c) is carried out to convolutional neural networks by following formula, obtains updated weight vectors Wnew；
Wherein, W is the weight vectors of convolutional neural networks parameter；
By updated weight vectors WnewBring 3c into) in loss function, obtain the updated loss function of weight vectors lossop；
4d) picture will be trained to be input to convolutional neural networks next time, loss function loss updated to weight vectorsop It is updated, so that loss function lossopValue constantly decline；
4e) repeat 4d), until loss function lossopNo longer decline, and current exercise wheel number is less than the iteration time of setting Number P then stops the training to the network, obtains trained convolutional neural networks；Otherwise, when training round reaches changing for setting When generation number P, stops the training to the network, obtain trained convolutional neural networks；
Step 5 classifies to the remote sensing scene picture that user inputs.
5a) remote sensing images to be sorted are normalized in user, i.e., first obtain remote sensing images pixel to be sorted Maximum value V'maxWith the minimum value V' of pixelmin, then to the value of all pixels points of remote sensing images to be sorted divided by V'maxWith V'minDifference, the remote sensing images to be sorted after obtaining normalized；
5b) remote sensing images after normalized are input in trained convolutional network model, obtain classification results.
Effect of the invention can be further illustrated by following emulation:
1. simulated conditions
This example in HP-Z840-Workstation with Xeon (R) CPU E5-2630, GeForce TITAN XP, Under 64G RAM, Ubuntu system, on TensorFlow operation platform, the present invention and existing remote sensing images scene classification are completed Emulation.
Simulation parameter setting is as follows, and iteration round P is 100 times, learning rate 0.00001, λr=0.7, λs=0.3, η= 0.0001, inputting picture number G every time is 6 times, and attenuation rate β is that 0.9, SCDA module takes 3 groups altogether, and three cell mean threshold values are respectively p1=1.0, p2=0.8, p3=0.6, by training data Random-Rotation, enhancing is four times of original data number.Learning sequence is In repetitive exercise each time, to category arbiter, difference optimizer of classifying, common training.
2. emulation content
UC Merced remote sensing images collection is downloaded, shown in Fig. 3, and it is normalized, i.e., first obtains UC Merced image set pixel maximum value V "maxWith the minimum value V " of pixelmin, then to all pictures of UC Merced image set The value of vegetarian refreshments is divided by V "maxWith V "minDifference, the UC Merced image set after obtaining normalized；
80% remote sensing images are selected at random from the UC Merced image after normalized as training sample set DT, using remaining 20% remote sensing images as test sample collection Dt；
Under above-mentioned simulated conditions, using training sample set DTRespectively with the present invention and existing representative three kinds of images point Class model is trained, using test sample collection DtIt is tested, compares the accuracy rate of its classification, as a result such as table 1.
The image that the training sample set and test sample are concentrated has 21 types, respectively agricultural, airplane、baseball diamond、beach、buildings、chaparral、dense residential、forest、 freeway、golf course、harbor、intersection、medium residential、mobilehomepark、 overpass、parking lot、river、runway、sparse residential、storage tanks、tennis Court,
1 present invention of table is evaluated with existing remote sensing image classification model performance
Test sample accuracy rate
The present invention 0.9849
MSCP 0.9782
SHHTFM 0.9789
DCA 0.9690
MSCP is the existing Classifying Method in Remote Sensing Image based on multiple pileup covariance pond in table 1, and SHHTFM is existing base In the Classifying Method in Remote Sensing Image that isomorphism isomery is sparse, DCA is the existing Classifying Method in Remote Sensing Image based on depth characteristic fusion.
As it can be seen from table 1 in training sample set DTWhen the percentage for accounting for UC Merced image set is 80%, sent out with this Bright trained convolutional neural networks to 20% test sample collection DtClassify, accuracy rate is representative more distant than existing It is high to feel image classification model accurate rate.
In conclusion the present invention is substantially better than other remote sensing image classification models for the classifying quality of remote sensing images.

Claims (6)

1. a kind of Classifying Method in Remote Sensing Image based on the multiple dimensioned deep learning of attention mechanism, which is characterized in that include the following:
(1) remote sensing images library { I is established1,I2,…In…,IN, the corresponding classification of image library is { Y1,Y2,…Yn…,YN, and it is right The remote sensing images library of foundation is normalized, wherein InN-th image in representative image library, YnN-th in representative image library The corresponding classification of image, n-th of sample number in n representative image library, n ∈ [0, N], N represent the number of pictures in remote sensing images library；
(2) it selects 80% remote sensing images sample at random from every class remote sensing images after normalized, constructs training image Library n ∈ [0, N], using remaining 20% remote sensing images as test sample collection { T1,T2,…Td…,Tm, wherein TjIndicate training J-th of sample in sample, j ∈ [0, M], tdIndicate that d-th of sample in test sample, d ∈ [0, m], M are the total of training sample Number, m are the total number of test sample, m < N, M < N.Wherein, wherein TjRepresent j-th of picture in training image library, j ∈ [0, M], M is the total number of training sample；
(3) constructing one includes convolutional network module, the convolutional neural networks for paying attention to power module, SCDA module and full articulamentum；
(4) loss function of convolutional neural networks is determined:
(4a) is by training image library { T1,T2,…Tj…,TMIt is input to the convolutional network module of convolutional layer neural network, output volume The last layer feature F of lamination；
The last layer feature F is input to the attention power module of convolutional neural networks by (4b), exports convolution feature A, then by convolution Layer feature A is input to multiple SCDA modules that convolutional neural networks have different average thresholds, exports T group mask convolution feature: {M1,M2,…,MT, wherein T is the number of SCDA module；
T group mask convolution feature is input to the full articulamentum of convolutional neural networks, output by (4c) by global average Chi Huahou The classification results of training data obtain the loss function of convolutional neural networks:
Wherein, loss1For the cross entropy of output category result and actual result, loss2It is T group mask convolution feature by connecting entirely Connect after layer the absolute value of output category result and actual result cross entropy and,For the L2 of convolutional neural networks weight vectors Norm, λr、λs, η be respectively loss1, loss2,Hyper parameter；
(5) setting the number of iterations is P, training is iterated to convolutional neural networks by gradient decline optimization, until losing letter Number lossopDo not decline or exercise wheel number reaches the number of iterations, obtains trained convolutional neural networks；
(6) user is input in trained convolutional neural networks after normalizing remote sensing images I' to be sorted, is divided Class is as a result, complete picture classification.
2. passing through the method according to claim 1, wherein remote sensing images library is normalized in (1) Following formula carries out:
Wherein VmaxFor the point maximum value of all pixels in remote sensing images library, VminIt is minimum for the point of all pixels in remote sensing images library Value, { I'1,I'2,…I'n…,I'NBe normalized after remote sensing images library, I'nFor remote sensing images after normalized N-th of sample, n ∈ [0, N].
3. the method according to claim 1, wherein (3) in composition convolutional neural networks convolutional network module, Notice that power module, SCDA module and full articulamentum, parameter setting are as follows:
The convolutional network module, by five convolutional layers sequentially connected in pre-training AlexNet network conv1, conv2, Conv3, conv4, conv5 } it constitutes；
The attention mechanism module is connected entirely by the average pond layer of the overall situation, the first full articulamentum, Relu activation primitive, second It connects layer and Sigmoid function is constituted；
The SCDA module, successively it is made of convolutional channel summation layer and mask layer.
4. the method according to claim 1, wherein output category result in (4c) and actual result are intersected Entropy loss1, formula is as follows:
Wherein, yjFor T in training image libraryjPrediction category probability, ojFor T in training image libraryjPractical category.
5. the method according to claim 1, wherein the T group mask convolution feature in (4c) passes through full articulamentum Afterwards the absolute value of output category result and actual result cross entropy and, formula is as follows:
Wherein T represents the number of SCDA module, lossmRepresent T in training image libraryjLoss under m convolution mask feature1, lossnRepresent T in training image libraryjLoss under the n-th convolution mask feature1。
6. the method according to claim 1, wherein by gradient decline optimization to convolutional neural networks in (5) It is iterated training, is accomplished by
The initial learning rate of (5a) setting training is L, attenuation rate β, by training image library { T1,T2,…Tj…,TMBe divided into G times It inputs in the convolutional neural networks of building, the number of pictures Q inputted every time are as follows:
Wherein M is the total number of training image library sample；
(5b) sets the corresponding learning rate l of input picture every time are as follows:
L=L* βG；
(5c) carries out the update of G subparameter to convolutional neural networks by following formula, obtains updated weight vectors Wnew；
Wherein, W is the weight vectors of convolutional neural networks parameter；
(5d) will train picture to input convolutional neural networks, loss function loss updated to weight vectors next timeopIt carries out It updates, so that loss function lossopValue constantly decline；
(5e) repeats (5d), until loss function lossopNo longer decline, and current exercise wheel number is less than the number of iterations of setting P then stops the training to the network, obtains trained convolutional neural networks；Otherwise, when training round reaches the iteration of setting When number P, stops the training to the network, obtain trained convolutional neural networks.
CN201910603799.6A 2019-07-05 2019-07-05 Remote sensing image classification method based on attention mechanism multi-scale deep learning Active CN110334765B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910603799.6A CN110334765B (en) 2019-07-05 2019-07-05 Remote sensing image classification method based on attention mechanism multi-scale deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910603799.6A CN110334765B (en) 2019-07-05 2019-07-05 Remote sensing image classification method based on attention mechanism multi-scale deep learning

Publications (2)

Publication Number Publication Date
CN110334765A true CN110334765A (en) 2019-10-15
CN110334765B CN110334765B (en) 2023-03-24

Family

ID=68144267

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910603799.6A Active CN110334765B (en) 2019-07-05 2019-07-05 Remote sensing image classification method based on attention mechanism multi-scale deep learning

Country Status (1)

Country Link
CN (1) CN110334765B (en)

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866494A (en) * 2019-11-14 2020-03-06 三亚中科遥感研究所 Optical remote sensing image-based town group extraction method and system
CN111046962A (en) * 2019-12-16 2020-04-21 中国人民解放军战略支援部队信息工程大学 Sparse attention-based feature visualization method and system for convolutional neural network model
CN111178304A (en) * 2019-12-31 2020-05-19 江苏省测绘研究所 High-resolution remote sensing image pixel level interpretation method based on full convolution neural network
CN111275192A (en) * 2020-02-28 2020-06-12 交叉信息核心技术研究院(西安)有限公司 Auxiliary training method for simultaneously improving accuracy and robustness of neural network
CN111339862A (en) * 2020-02-17 2020-06-26 中国地质大学（武汉） Remote sensing scene classification method and device based on channel attention mechanism
CN111723674A (en) * 2020-05-26 2020-09-29 河海大学 Remote sensing image scene classification method based on Markov chain Monte Carlo and variation deduction and semi-Bayesian deep learning
CN111738124A (en) * 2020-06-15 2020-10-02 西安电子科技大学 Remote sensing image cloud detection method based on Gabor transformation and attention
CN111797941A (en) * 2020-07-20 2020-10-20 中国科学院长春光学精密机械与物理研究所 Image classification method and system carrying spectral information and spatial information
CN111861880A (en) * 2020-06-05 2020-10-30 昆明理工大学 Image super-fusion method based on regional information enhancement and block self-attention
CN112101190A (en) * 2020-09-11 2020-12-18 西安电子科技大学 Remote sensing image classification method, storage medium and computing device
CN112580557A (en) * 2020-12-25 2021-03-30 深圳市优必选科技股份有限公司 Behavior recognition method and device, terminal equipment and readable storage medium
CN112926380A (en) * 2021-01-08 2021-06-08 浙江大学 Novel underwater laser target intelligent recognition system
CN113177523A (en) * 2021-05-27 2021-07-27 青岛杰瑞工控技术有限公司 Fish behavior image identification method based on improved AlexNet
CN113191285A (en) * 2021-05-08 2021-07-30 山东大学 River and lake remote sensing image segmentation method and system based on convolutional neural network and Transformer
CN113435531A (en) * 2021-07-07 2021-09-24 中国人民解放军国防科技大学 Zero sample image classification method and system, electronic equipment and storage medium
CN113449712A (en) * 2021-09-01 2021-09-28 武汉方芯科技有限公司 Goat face identification method based on improved Alexnet network
CN113505651A (en) * 2021-06-15 2021-10-15 杭州电子科技大学 Mosquito identification method based on convolutional neural network
CN114092832A (en) * 2022-01-20 2022-02-25 武汉大学 High-resolution remote sensing image classification method based on parallel hybrid convolutional network
CN114286113A (en) * 2021-12-24 2022-04-05 国网陕西省电力有限公司西咸新区供电公司 Image compression recovery method and system based on multi-head heterogeneous convolution self-encoder

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106909924A (en) * 2017-02-18 2017-06-30 北京工业大学 A kind of remote sensing image method for quickly retrieving based on depth conspicuousness
CN108830296A (en) * 2018-05-18 2018-11-16 河海大学 A kind of improved high score Remote Image Classification based on deep learning
WO2018214195A1 (en) * 2017-05-25 2018-11-29 中国矿业大学 Remote sensing imaging bridge detection method based on convolutional neural network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106909924A (en) * 2017-02-18 2017-06-30 北京工业大学 A kind of remote sensing image method for quickly retrieving based on depth conspicuousness
WO2018214195A1 (en) * 2017-05-25 2018-11-29 中国矿业大学 Remote sensing imaging bridge detection method based on convolutional neural network
CN108830296A (en) * 2018-05-18 2018-11-16 河海大学 A kind of improved high score Remote Image Classification based on deep learning

Cited By (30)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866494B (en) * 2019-11-14 2022-09-06 三亚中科遥感研究所 Urban group extraction method and system based on optical remote sensing image
CN110866494A (en) * 2019-11-14 2020-03-06 三亚中科遥感研究所 Optical remote sensing image-based town group extraction method and system
CN111046962A (en) * 2019-12-16 2020-04-21 中国人民解放军战略支援部队信息工程大学 Sparse attention-based feature visualization method and system for convolutional neural network model
CN111178304A (en) * 2019-12-31 2020-05-19 江苏省测绘研究所 High-resolution remote sensing image pixel level interpretation method based on full convolution neural network
CN111178304B (en) * 2019-12-31 2021-11-05 江苏省测绘研究所 High-resolution remote sensing image pixel level interpretation method based on full convolution neural network
CN111339862A (en) * 2020-02-17 2020-06-26 中国地质大学（武汉） Remote sensing scene classification method and device based on channel attention mechanism
CN111275192A (en) * 2020-02-28 2020-06-12 交叉信息核心技术研究院(西安)有限公司 Auxiliary training method for simultaneously improving accuracy and robustness of neural network
CN111275192B (en) * 2020-02-28 2023-05-02 交叉信息核心技术研究院(西安)有限公司 Auxiliary training method for improving accuracy and robustness of neural network simultaneously
CN111723674A (en) * 2020-05-26 2020-09-29 河海大学 Remote sensing image scene classification method based on Markov chain Monte Carlo and variation deduction and semi-Bayesian deep learning
CN111723674B (en) * 2020-05-26 2022-08-05 河海大学 Remote sensing image scene classification method based on Markov chain Monte Carlo and variation deduction and semi-Bayesian deep learning
CN111861880B (en) * 2020-06-05 2022-08-30 昆明理工大学 Image super-fusion method based on regional information enhancement and block self-attention
CN111861880A (en) * 2020-06-05 2020-10-30 昆明理工大学 Image super-fusion method based on regional information enhancement and block self-attention
CN111738124B (en) * 2020-06-15 2023-08-22 西安电子科技大学 Remote sensing image cloud detection method based on Gabor transformation and attention
CN111738124A (en) * 2020-06-15 2020-10-02 西安电子科技大学 Remote sensing image cloud detection method based on Gabor transformation and attention
CN111797941A (en) * 2020-07-20 2020-10-20 中国科学院长春光学精密机械与物理研究所 Image classification method and system carrying spectral information and spatial information
CN112101190B (en) * 2020-09-11 2023-11-03 西安电子科技大学 Remote sensing image classification method, storage medium and computing device
CN112101190A (en) * 2020-09-11 2020-12-18 西安电子科技大学 Remote sensing image classification method, storage medium and computing device
CN112580557A (en) * 2020-12-25 2021-03-30 深圳市优必选科技股份有限公司 Behavior recognition method and device, terminal equipment and readable storage medium
CN112926380A (en) * 2021-01-08 2021-06-08 浙江大学 Novel underwater laser target intelligent recognition system
CN112926380B (en) * 2021-01-08 2022-06-24 浙江大学 Novel underwater laser target intelligent recognition system
CN113191285A (en) * 2021-05-08 2021-07-30 山东大学 River and lake remote sensing image segmentation method and system based on convolutional neural network and Transformer
CN113177523A (en) * 2021-05-27 2021-07-27 青岛杰瑞工控技术有限公司 Fish behavior image identification method based on improved AlexNet
CN113505651A (en) * 2021-06-15 2021-10-15 杭州电子科技大学 Mosquito identification method based on convolutional neural network
CN113435531A (en) * 2021-07-07 2021-09-24 中国人民解放军国防科技大学 Zero sample image classification method and system, electronic equipment and storage medium
CN113435531B (en) * 2021-07-07 2022-06-21 中国人民解放军国防科技大学 Zero sample image classification method and system, electronic equipment and storage medium
CN113449712A (en) * 2021-09-01 2021-09-28 武汉方芯科技有限公司 Goat face identification method based on improved Alexnet network
CN114286113A (en) * 2021-12-24 2022-04-05 国网陕西省电力有限公司西咸新区供电公司 Image compression recovery method and system based on multi-head heterogeneous convolution self-encoder
CN114286113B (en) * 2021-12-24 2023-05-30 国网陕西省电力有限公司西咸新区供电公司 Image compression recovery method and system based on multi-head heterogeneous convolution self-encoder
CN114092832B (en) * 2022-01-20 2022-04-15 武汉大学 High-resolution remote sensing image classification method based on parallel hybrid convolutional network
CN114092832A (en) * 2022-01-20 2022-02-25 武汉大学 High-resolution remote sensing image classification method based on parallel hybrid convolutional network

Also Published As

Publication number Publication date
CN110334765B (en) 2023-03-24

Similar Documents

Publication Publication Date Title
CN110334765A (en) Remote Image Classification based on the multiple dimensioned deep learning of attention mechanism
CN110516596B (en) Octave convolution-based spatial spectrum attention hyperspectral image classification method
CN110516085B (en) Image text mutual retrieval method based on bidirectional attention
CN107784320B (en) Method for identifying radar one-dimensional range profile target based on convolution support vector machine
CN110097103A (en) Based on the semi-supervision image classification method for generating confrontation network
CN105678284B (en) A kind of fixed bit human body behavior analysis method
CN110136154A (en) Remote sensing images semantic segmentation method based on full convolutional network and Morphological scale-space
CN110852227A (en) Hyperspectral image deep learning classification method, device, equipment and storage medium
CN112052754B (en) Polarization SAR image ground object classification method based on self-supervision characterization learning
CN108549893A (en) A kind of end-to-end recognition methods of the scene text of arbitrary shape
CN108596101A (en) A kind of remote sensing images multi-target detection method based on convolutional neural networks
CN107229904A (en) A kind of object detection and recognition method based on deep learning
CN108961245A (en) Picture quality classification method based on binary channels depth parallel-convolution network
CN107563428A (en) Classification of Polarimetric SAR Image method based on generation confrontation network
CN107832797B (en) Multispectral image classification method based on depth fusion residual error network
CN107944483B (en) Multispectral image classification method based on dual-channel DCGAN and feature fusion
CN108846426A (en) Polarization SAR classification method based on the twin network of the two-way LSTM of depth
CN110399856A (en) Feature extraction network training method, image processing method, device and its equipment
CN106022273A (en) Handwritten form identification system of BP neural network based on dynamic sample selection strategy
CN109726748B (en) GL-CNN remote sensing image scene classification method based on frequency band feature fusion
CN107886123A (en) A kind of synthetic aperture radar target identification method based on auxiliary judgement renewal learning
CN106600595A (en) Human body characteristic dimension automatic measuring method based on artificial intelligence algorithm
CN106203625A (en) A kind of deep-neural-network training method based on multiple pre-training
CN108460391A (en) Based on the unsupervised feature extracting method of high spectrum image for generating confrontation network
CN110163213A (en) Remote sensing image segmentation method based on disparity map and multiple dimensioned depth network model

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