CN108664994A - A kind of remote sensing image processing model construction system and method - Google Patents
A kind of remote sensing image processing model construction system and method Download PDFInfo
- Publication number
- CN108664994A CN108664994A CN201810344784.8A CN201810344784A CN108664994A CN 108664994 A CN108664994 A CN 108664994A CN 201810344784 A CN201810344784 A CN 201810344784A CN 108664994 A CN108664994 A CN 108664994A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- convolutional layer
- image processing
- sensing image
- detection window
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
- 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
- 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
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of remote sensing image processing model construction system and method, this method includes:A, based on average pond by the semantic feature of the shallow convolutional layer, the semantic feature normalization of deep convolutional layer;B, in conjunction with the shallow convolutional layer, deep convolutional layer in such a way that feature channel expands;C, the RPN units of detection window for generating target are set.The system is for executing method.The present invention can extract enough abstract characteristics by integrating the semantic feature of the convolutional network of multilayer, while reduce the loss of the marginal information of image, provide the precision of remote sensing image processing.
Description
Technical field
The present invention relates to remote sensing images technical field more particularly to a kind of remote sensing image processing model construction system and sides
Method.
Background technology
The target detection for being directed to remote sensing images at this stage has been achieved for many achievements, although current detection means is
Wherein by depth learning technology application, when but current deep learning detection framework detects the target in remote sensing images, effect
Fruit is simultaneously bad, especially when the target for needing to detect is smaller and more intensive, will produce a large amount of target missing inspection and flase drop is existing
As these limit the practical application of Remote Sensing Target detection.
Due to remote sensing images be using the image of satellite shooting, have contain much information, target is more, shooting height is higher leads
The features such as target is smaller in the remote sensing images of cause, these features increase the difficulty of Remote Sensing Target detection, this also leads to mesh
Effect of the preceding detection method in the practical application that Remote Sensing Target detects is poor.
Therefore, in order to improve the accuracy of Remote Sensing Target detection, academia and relevant enterprise have all done and have largely ground
Study carefully and technological development, but most of work is to be improved existing conventional method, to improve Remote Sensing Target detection
Accuracy rate.Although achieving certain progress, produce little effect.
Invention content
To solve the above-mentioned problems, the present invention provides a kind of remote sensing image processing model construction system and method.
On the one hand the technical solution adopted by the present invention is a kind of remote sensing image processing model construction system, including:Convolution net
Network module, including shallow convolutional layer and deep convolutional layer;Preprocessing module, for based on average pond by the semanteme of the shallow convolutional layer
The semantic feature normalized of feature, deep convolutional layer;Processing module, for by way of being expanded feature channel in conjunction with described
Shallow convolutional layer, deep convolutional layer;RPN units, the detection window for generating target.
Preferably, the convolutional network module is based on formulaProcessing input x and output y, wherein
The β coefficients that upper layer inputs in order to control,For the output valve based on neural network convolution, W is the weights of convolutional layer.
Preferably, further include correcting module,
Activation primitive is corrected for being arrangedWherein, xiFor positive value to be repaired, η is to correct activation letter
Several negative gradient coefficients.
Preferably, the RPN units are for executing step:The target signature in the training dataset of remote sensing images is obtained,
Target signature is clustered to k class using K-means clustering algorithms, finds out the central value of each class, it is special according to the information of acquisition
The central value of the K that seeks peace clusters constructs corresponding detection window length-width ratio.
Preferably, the RPN units are additionally operable to execute step:The target area of training dataset is modeled, according to
The distribution situation of area data, be distributed it is discrete turn to several casees, the central value of each case is calculated, according to each case and its
Central value increases the size of detection window.
Preferably, further include correction module, the Europe at the center at center and training data for calculating detection window position
Family name's distance is detected the offset correction of window based on the Euclidean distance, and convolution, sampling are carried out to revised detection window
To update the detection feature of detection window, based on softmax functions to be detected the precise classification of feature.
On the one hand the technical solution adopted by the present invention is a kind of remote sensing image processing model building method, including step:A、
Based on average pond by the semantic feature of the shallow convolutional layer, the semantic feature normalized of deep convolutional layer;B, pass through feature
The mode of channel amplification is in conjunction with the shallow convolutional layer, deep convolutional layer;C, the RPN that the detection window for generating target is arranged is mono-
Member.
Preferably, the convolutional network module is based on formulaProcessing input x and output y, wherein
The β coefficients that upper layer inputs in order to control,For the output valve based on neural network convolution, W is the weights of convolutional layer.
Preferably, further include step B, step D between C:
Activation primitive is corrected in settingWherein, xiFor positive value to be repaired, η is to correct activation primitive
Negative gradient coefficient.
Preferably, the RPN units are for executing step:The target signature in the training dataset of remote sensing images is obtained,
Target signature is clustered to k class using K-means clustering algorithms, finds out the central value of each class, it is special according to the information of acquisition
The central value of the K that seeks peace clusters constructs corresponding detection window length-width ratio.
Preferably, the RPN units are additionally operable to execute step:The target area of training dataset is modeled, according to
The distribution situation of area data, be distributed it is discrete turn to several casees, the central value of each case is calculated, according to each case and its
Central value increases the size of detection window.
Preferably, further include step E:The Euclidean distance at the center of detection window position and the center of training data is calculated,
It is detected the offset correction of window based on the Euclidean distance, convolution, sampling are carried out to update to revised detection window
The detection feature of detection window, based on softmax functions to be detected the precise classification of feature.
Beneficial effects of the present invention, which are the semantic feature of the convolutional network by integrating multilayer, can extract enough be abstracted
Feature, while the loss of the marginal information of image is reduced, the precision of remote sensing image processing is provided.
Description of the drawings
Fig. 1 show a kind of schematic diagram of remote sensing image processing model building method based on the embodiment of the present invention;
Fig. 2 show the fusion flow diagram based on the embodiment of the present invention;
Fig. 3 show the fusion frame diagram based on the embodiment of the present invention;
Fig. 4 show the RPN based on the embodiment of the present invention and constitutes flow diagram;
Fig. 5 show the multiwindow processing flow schematic diagram based on the embodiment of the present invention.
Specific implementation mode
The present invention will be described with reference to embodiments.
Embodiment 1 based on invention, a kind of remote sensing image processing model building method as shown in Figure 1, including step:A、
Based on average pond by the semantic feature of the shallow convolutional layer, the semantic feature normalization of deep convolutional layer;B, logical by feature
The mode of road amplification is in conjunction with the shallow convolutional layer, deep convolutional layer;C, the RPN units of detection window for generating target are set.
The convolutional network module is based on formulaProcessing input x and output y, wherein β is in order to control
The coefficient of upper layer input,For the output valve based on neural network convolution, W is the weights of convolutional layer.
Step B, between C further include step D:
Activation primitive is corrected in settingWherein, xiFor positive value to be repaired, η is to correct activation primitive
Negative gradient coefficient.
The RPN units are for executing step:The target signature in the training dataset of remote sensing images is obtained, using K-
Means clustering algorithms cluster target signature to k class, find out the central value of each class, poly- according to the information characteristics of acquisition and K
The central value of class constructs corresponding detection window length-width ratio.
The RPN units are additionally operable to execute step:The target area of training dataset is modeled, according to area data
Distribution situation, be distributed it is discrete turn to several casees, calculate the central value of each case, increased according to each case and its central value
Add the size of detection window.
Method further includes step E:The Euclidean distance for calculating the center of detection window position and the center of training data, is based on
The Euclidean distance is detected the offset correction of window, and convolution, sampling are carried out to update detection to revised detection window
The detection feature of window, based on softmax functions to be detected the precise classification of feature.
It include mainly two stages for being further improved for embodiment:
One, the training stage:The acquisition of remote sensing image data collection and making, the feature extraction of remote sensing images, remote sensing images mesh
Target positions and classification, model training;
Two, detection-phase:The feature extraction of remote sensing images, the positioning of target and classification;As based on the distant of deep learning
The part for feeling image object detection algorithm, by marking the target of the remote sensing images obtained from satellite to make for training
The training set (i.e. training dataset) of Remote Sensing Target detection model;Training stage includes step:
1, the remote sensing images mesh of the acquisition and making of remote sensing images training dataset, adaptive targets strategy and deep learning
Target detection algorithm belongs to supervised learning task, therefore, is marked first against remote sensing image data collection (i.e. training dataset);
These remote sensing image datas have diversity from the satellites such as high score two, wind and cloud No.1, data source, to ensure that instruction
The model practised has stronger robustness for the remote sensing images in a variety of sources;It is emerging that the content of mark mainly contains some senses
Interesting target, such as a plurality of types of aircrafts, ship;Mark mainly uses the mode for the minimum enclosed rectangle for outlining target, mark
The interesting target that each image contains is noted, training set is finally obtained.
2, target detection model is built:
2.1, remote sensing images include a large amount of information, and shallower network is unable to fully extraction feature, and current depth is refreshing
Through network because gradient diffusing phenomenon causes actual effect bad, therefore the present invention is based on the short directly mapping convolution network modules that is connected
The convolutional neural networks for constructing ultra-deep (including multiple convolutional layers), solve what deep neural network was constantly deepened in depth
In the case of backpropagation gradient disappear or explosion the phenomenon that, wherein it is short be connected directly mapping convolution network module (the i.e. described volume
Product network module) the door machine system that applies Recognition with Recurrent Neural Network, input x primarily directed to convolutional network module and by nerve
Output after network convolutionOutput y after weighting with the results added after volume as this network module, wherein W are
The specific formula of weights of convolutional layer is such as:The β coefficients that upper layer (convolutional layer) inputs in order to control, together
When β can learn and update in training;The module will make the feature that ultra-deep convolutional network is extracted in remote sensing images not
It is multiplexed under the premise of increasing network burden, therefore produces more abstract, more expression in the highest semantic layer of feature and anticipate
The feature of justice, the final generalization ability for improving model.
Because the depth of neural network will have a direct impact on the effect of extraction feature, therefore, the present invention is based on basic short phases
The deep neural network that convolution network module has built totally 121 layers is even directly mapped, the spy to remote sensing images can be fully learnt
Sign;Meanwhile remote sensing images, with the intensification of neural network depth, the feature gone out from image zooming-out is more abstract, is conducive to target
Classification, but the marginal information loss of target can be than more serious in remote sensing images, the especially ship of Small object and close
Connected ship, which results in current remote sensing images detection model is ineffective to small target deteection;To ensure to extract
Feature there is sufficient location information, fusion flow diagram as shown in Figure 2, this programme proposes a kind of polynary fusion machine
System both ensure that the ultra-deep convolutional network based on door machine system (threshold structure (gate)) was abstracted enough in the feature extracted
Meanwhile in turn ensuring that the marginal information loss of target in image is less;Mainly using a kind of, polynary based on door machine melts mechanism
Close the convolutional layer feature of ultra-deep convolutional neural networks diversification;The stage is extracted in diverse characteristics, by the shallow-layer DC_2c of diversification
The semantic feature of layer, the semantic feature of DC_3d layers (i.e. shallow-layer convolutional layer) and DC_4e layers of ultra deep (i.e. deep layer convolutional layer), (
Polynary semantic feature fusing stage) it first passes through polynary semantic feature and normalizes to same size in a manner of average pond, then will
This three-layer network is combined in such a way that feature channel expands and (merges frame diagram as shown in Figure 3), is activated finally by correcting
Polynary fusion feature is output to RPN candidate frames and generates network (i.e. RPN modules) by function, wherein correcting activation primitive formula such as
Under:
Wherein η is the negative gradient coefficient for correcting activation primitive, and the polynary semantic feature that activation primitive will increase merges non-thread
Sexual intercourse makes the Characteristics of The Remote Sensing Images that network extracts have more expression and significance.
Fusion frame diagram as shown in Figure 3:By based on door machine ultra-deep convolutional neural networks and diverse characteristics fusion
Mechanism is combined, and ensure that the demand that the target signature of remote sensing images is classified to network depth, while also ensuring the side of target
Edge characteristic information is fully conducive to the positioning of target enough;I.e. by convolutional network, (convolutional network includes several convolution to input picture
Layer, Conv1, DC_2c, DC_3d and DC_4e etc.) processing;By Max Pooling by shallow-layer convolutional layer (i.e. DC_2c, DC_3d)
It is normalized to and normalizes to same size with deep layer convolutional layer (i.e. DC_4e) characteristic pattern, be combined by Concat functions, passed through
RPN modules are identified, and finally carry out the classification and positioning of target.
2.2, this programme devises the target detection frame extraction network (i.e. RPN modules) of remote sensing images, constructs and is suitable for
The RPN algorithms of the adaptive targets of remote sensing images;The top layer in neural network that the RPN algorithms mainly use sliding window is special
Sign figure sliding generates ROI (Region Of Interest), and generate window (anchor) shape of ROI mainly by length and width when
Its area determines.
RPN as shown in Figure 4 constitutes flow, this algorithm is first for the feature that Remote Sensing Target is small, target is closely coupled
The target aspect ratio information (target signature i.e. described in claim) in remote sensing images training dataset is first excavated, is then used
The cluster of aspect ratio information feature to k class is found out the central value of each class by K-means clustering algorithms, according to excavating
Information characteristics and the central values of K clusters construct corresponding detection window length-width ratio;Secondly, smaller for target in remote sensing images
The case where, adaptive targets RPN networks model the target area of training dataset, using the histogram branch mailbox side of area
Formula, the primary operational of branch mailbox are the distribution situation according to area data, be distributed it is discrete turn to several casees (bin), finally
The central value for calculating each bin increases the size of detection window, the RPN of the adaptive targets by each bin and its central value
Network can generate candidate frame with higher recall rate for target not of uniform size and (be used for scanning the sliding window of remote sensing images
Mouth, i.e. detection window), to solve current remote sensing images detection framework because of the IOU in candidate frame detection window
(Intersection-Over-Union) remote sensing images Small object detection leakage phenomenon caused by too small.
2.3, adaptive RPN networks extract candidate frame in the highest characteristic layer of ultra-deep convolutional neural networks and (are used for
Detect the detection window of target) after, algorithm constructs self-correction and sorts out network with the detection window to ROI into line position with accurate
Self-correction is set, while also being classified to the target in ROI.
First, algorithm is modified the position of detection window using self-correction strategy;The strategy calculates detection window first
The Euclidean distance at the center and the center of target GroundTruth of mouth position, the formula for calculating Euclidean distance are as follows:
Wherein, XpIndicate the center of detection window, XgThe center for indicating GroundTruth, (in machine learning
Middle ground truth indicate the classification accuracy of the training set of supervised learning, for proving or overthrowing some it is assumed that having
The machine learning of supervision can to training data marking, just imagine if training marked erroneous, will be to test data
Detection have an impact, here by the numerical nomenclature of those correct labelings be GroundTruth) and d represents the two offset
Amount, is detected window in the enterprising line displacement amendment of characteristic pattern, revised distance is according to Euclidean distance is obtained:
x′p,x=xp,x+ d, x 'p,y=xp,y+ d (4),
Accurate bezel locations can be obtained by the position offset based on Euclidean distance, realize test position
Self-correction.Convolution down-sampling is carried out to the detection window after offset correction and obtains a more abstract target signature (i.e. with fixed
The benchmark of target of the justice in the remote sensing images that detection window streaks, such as including how much each sides, which the angle that side is formed is
A range, for example, steamer then include two straight flanges, two camber lines), finally use the newer target signature of softmax function pairs into
Row precise classification;The target loss function of detection framework training is improved simultaneously, and the Euclidean distance penalty term having to offset is set
ReSmooth L1Loss, wherein α is penalty coefficient, and objectives function is as follows:
2.4, the remotely-sensed data marked and Remote Sensing Target detection model are instructed with deep learning frame caffe
Practice, ultra-deep convolutional neural networks are carried out tune ginseng by the method that transfer learning is used in training on imagenet data sets
It trains, in the M-RPN models for then moving to trained model parameter, and is repeatedly instructed based on online difficult example parser
Practice remote sensing image data set, the parameter setting of best model training, preservation model result are found by validation data set.
3, detection-phase:
Step 1 loads the model of training stage:
Trained Remote Sensing Target detection model is loaded using the python interfaces of caffe deep learning frames,
Reading model relevant parameter.
The integrated detection of different transform of the step 2 based on multiwindow:
Multiwindow processing flow schematic diagram as shown in Figure 5, since the resolution ratio of remote sensing images is higher, it is meant that processing mould
Type needs are largely calculated, and slowing for detection is caused, and even result in model collapse;General processing scheme is by image
It zooms in and out, such scheme can make image impairment bulk information, be positioned so as to cause target flase drop, target missing inspection and target
The problems such as inaccurate, seriously affects detection result.Therefore this programme uses multiwindow detection mode in detection-phase, passes through reading
The resolution ratio of image by high-resolution remote sensing images Detection task using the window of different scale to be decomposed into several subgraphs
Picture, such decomposition avoids image scaling, and then ensure that localized target after the multiple down-sampling of ultra-deep convolutional network
Still there is sufficient location information;Meanwhile more change process are done to each subgraph, including to image carry out flip horizontal,
Left and right overturning, rotation, amplification, enhancing detection framework is to the generalization ability of the targets direction such as ship, then in detection sub task
Subtask testing result used is carried out integrated fusion by image finally by NMS (non-maxima suppression).
Such as the image resolution ratio of input is 3000*5000, then it is 600*1000 as subtask window to use window, together
When in order to ensure that window edge target is not cut, it is the intersection region of 300 pixels to have distance between each window.According to subtask
Window size, it is 600*1000 subgraphs that original image, which is broken down into 5*5 resolution ratio, and the calculation formula of subtask number is:
Wherein Ih, IwThe respectively length and width of input picture, Ph, PwThe respectively length and width of subtask window, d are safe spacing;
Flip horizontal, left and right overturning, rotation, amplification are done into 25 subtasks after decomposition respectively, 25* (4+1) a subgraph is obtained,
All subgraphs are detected, are then integrated all subtask testing results by NMS (non-maxima suppression),
The prediction result for rejecting high superposed, finally obtains final testing result.
Embodiment 2 based on invention, a kind of remote sensing image processing model construction system, including:Convolutional network module, including
Shallow convolutional layer and deep convolutional layer;Preprocessing module, for based on average pond by the semantic feature of the shallow convolutional layer, depth convolution
The semantic feature normalization of layer;Processing module, for being rolled up in conjunction with the shallow convolutional layer, deeply by way of being expanded feature channel
Lamination;RPN units, the detection window for generating target.
The convolutional network module is based on formulaProcessing input x and output y, wherein β is in order to control
The coefficient of upper layer input,For the output valve based on neural network convolution, W is the weights of convolutional layer.
Further include correcting module,
Activation primitive is corrected for being arrangedWherein, xiFor positive value to be repaired, η is to correct to activate
The negative gradient coefficient of function.
The RPN units are for executing step:The target signature in the training dataset of remote sensing images is obtained, using K-
Means clustering algorithms cluster target signature to k class, find out the central value of each class, poly- according to the information characteristics of acquisition and K
The central value of class constructs corresponding detection window length-width ratio.
The RPN units are additionally operable to execute step:The target area of training dataset is modeled, according to area data
Distribution situation, be distributed it is discrete turn to several casees, calculate the central value of each case, increased according to each case and its central value
Add the size of detection window.
System further includes correction module, the Euclidean at the center at center and training data for calculating detection window position away from
From, the offset correction of window is detected based on the Euclidean distance, to revised detection window carry out convolution, sampling with more
The detection feature of new detection window, based on softmax functions to be detected the precise classification of feature.
The above, only presently preferred embodiments of the present invention, the invention is not limited in the above embodiments, as long as
It reaches the technique effect of the present invention with identical means, should all belong to the scope of protection of the present invention.In the protection model of the present invention
Its technical solution embodiment can have a variety of different modifications and variations in enclosing.
Claims (12)
1. a kind of remote sensing image processing model construction system, which is characterized in that including:
Convolutional network module, including shallow convolutional layer and deep convolutional layer;
Preprocessing module, for being returned the semantic feature of the semantic feature of the shallow convolutional layer, deep convolutional layer based on average pond
One processing;
Processing module, for by way of being expanded feature channel in conjunction with the shallow convolutional layer, deep convolutional layer;
RPN units, the detection window for generating target.
2. a kind of remote sensing image processing model construction system according to claim 1, which is characterized in that the convolutional network
Module is based on formulaProcessing input x and output y, wherein
The β coefficients that upper layer inputs in order to control,For the output valve based on neural network convolution, W is the weights of convolutional layer.
3. a kind of remote sensing image processing model construction system according to claim 1, which is characterized in that further include correcting mould
Block,
Activation primitive is corrected for being arrangedWherein, xiFor positive value to be repaired, η is to correct activation primitive
Negative gradient coefficient.
4. a kind of remote sensing image processing model construction system according to claim 1, which is characterized in that the RPN units
For executing step:
The target signature in the training dataset of remote sensing images is obtained, is clustered target signature to k using K-means clustering algorithms
A class finds out the central value of each class, and it is long to construct corresponding detection window according to the information characteristics of acquisition and the K central value clustered
Wide ratio.
5. a kind of remote sensing image processing model construction system according to claim 4, which is characterized in that the RPN units
It is additionally operable to execute step:
The target area of training dataset is modeled, according to the distribution situation of area data, if being distributed discrete turn to
Dry case, calculates the central value of each case, increases the size of detection window according to each case and its central value.
6. according to a kind of remote sensing image processing model construction system of Claims 1 to 5 any one of them, which is characterized in that also
Including correction module, the Euclidean distance at the center at center and training data for calculating detection window position is based on the Europe
Family name's distance is detected the offset correction of window, and convolution, sampling are carried out to update detection window to revised detection window
Feature is detected, based on softmax functions to be detected the precise classification of feature.
7. a kind of remote sensing image processing model building method is suitable for system described in claim 1, which is characterized in that including step
Suddenly:
A, based on average pond by the semantic feature of the shallow convolutional layer, the semantic feature normalization of deep convolutional layer;
B, in conjunction with the shallow convolutional layer, deep convolutional layer in such a way that feature channel expands;
C, the RPN units of detection window for generating target are set.
8. remote sensing image processing model building method according to claim 7, which is characterized in that the convolutional network module
Based on formulaProcessing input x and output y, wherein
The β coefficients that upper layer inputs in order to control,For the output valve based on neural network convolution, W is the weights of convolutional layer.
9. remote sensing image processing model building method according to claim 7, which is characterized in that step B, also wrapped between C
Include step D:
Activation primitive is corrected in settingWherein, xiFor positive value to be repaired, η is the negative ladder for correcting activation primitive
Spend coefficient.
10. remote sensing image processing model building method according to claim 7, which is characterized in that the RPN units are used for
Execute step:
The target signature in the training dataset of remote sensing images is obtained, is clustered target signature to k using K-means clustering algorithms
A class finds out the central value of each class, and it is long to construct corresponding detection window according to the information characteristics of acquisition and the K central value clustered
Wide ratio.
11. remote sensing image processing model building method according to claim 10, which is characterized in that the RPN units are also
For executing step:
The target area of training dataset is modeled, according to the distribution situation of area data, if being distributed discrete turn to
Dry case, calculates the central value of each case, increases the size of detection window according to each case and its central value.
12. according to claim 7~11 any one of them remote sensing image processing model building method, which is characterized in that also wrap
Include step E:
The Euclidean distance for calculating the center of detection window position and the center of training data, is detected based on the Euclidean distance
The offset correction of window carries out convolution, sampling to update the detection feature of detection window to revised detection window, is based on
Softmax functions are to be detected the precise classification of feature.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810344784.8A CN108664994A (en) | 2018-04-17 | 2018-04-17 | A kind of remote sensing image processing model construction system and method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810344784.8A CN108664994A (en) | 2018-04-17 | 2018-04-17 | A kind of remote sensing image processing model construction system and method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108664994A true CN108664994A (en) | 2018-10-16 |
Family
ID=63782226
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810344784.8A Pending CN108664994A (en) | 2018-04-17 | 2018-04-17 | A kind of remote sensing image processing model construction system and method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108664994A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109583517A (en) * | 2018-12-26 | 2019-04-05 | 华东交通大学 | A kind of full convolution example semantic partitioning algorithm of the enhancing suitable for small target deteection |
CN109784375A (en) * | 2018-12-22 | 2019-05-21 | 国网内蒙古东部电力有限公司 | Adaptive transformer part detection recognition method based on Faster RCNN |
CN110009648A (en) * | 2019-03-04 | 2019-07-12 | 东南大学 | Trackside image Method of Vehicle Segmentation based on depth Fusion Features convolutional neural networks |
CN111209813A (en) * | 2019-12-27 | 2020-05-29 | 南京航空航天大学 | Remote sensing image semantic segmentation method based on transfer learning |
CN111536991A (en) * | 2020-04-26 | 2020-08-14 | 哈尔滨工业大学 | Remote sensing satellite multi-mode imaging load simulation device |
CN111967734A (en) * | 2020-07-30 | 2020-11-20 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | Method for planning ground observation requirements of micro-nano remote sensing network |
CN113558644A (en) * | 2021-07-20 | 2021-10-29 | 陕西科技大学 | Emotion classification method, medium and equipment for 3D matrix and multidimensional convolution network |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104102929A (en) * | 2014-07-25 | 2014-10-15 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
CN104751162A (en) * | 2015-04-03 | 2015-07-01 | 哈尔滨工业大学 | Hyperspectral remote sensing data feature extraction method based on convolution neural network |
CN105512661A (en) * | 2015-11-25 | 2016-04-20 | 中国人民解放军信息工程大学 | Multi-mode-characteristic-fusion-based remote-sensing image classification method |
CN106295714A (en) * | 2016-08-22 | 2017-01-04 | 中国科学院电子学研究所 | A kind of multi-source Remote-sensing Image Fusion based on degree of depth study |
WO2017087334A1 (en) * | 2015-11-16 | 2017-05-26 | Orbital Insight, Inc. | Moving vehicle detection and analysis using low resolution remote sensing imagery |
CN107122796A (en) * | 2017-04-01 | 2017-09-01 | 中国科学院空间应用工程与技术中心 | A kind of remote sensing image sorting technique based on multiple-limb network integration model |
-
2018
- 2018-04-17 CN CN201810344784.8A patent/CN108664994A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104102929A (en) * | 2014-07-25 | 2014-10-15 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
CN104751162A (en) * | 2015-04-03 | 2015-07-01 | 哈尔滨工业大学 | Hyperspectral remote sensing data feature extraction method based on convolution neural network |
WO2017087334A1 (en) * | 2015-11-16 | 2017-05-26 | Orbital Insight, Inc. | Moving vehicle detection and analysis using low resolution remote sensing imagery |
CN105512661A (en) * | 2015-11-25 | 2016-04-20 | 中国人民解放军信息工程大学 | Multi-mode-characteristic-fusion-based remote-sensing image classification method |
CN106295714A (en) * | 2016-08-22 | 2017-01-04 | 中国科学院电子学研究所 | A kind of multi-source Remote-sensing Image Fusion based on degree of depth study |
CN107122796A (en) * | 2017-04-01 | 2017-09-01 | 中国科学院空间应用工程与技术中心 | A kind of remote sensing image sorting technique based on multiple-limb network integration model |
Non-Patent Citations (4)
Title |
---|
MIAO KANG 等: "Contextual Region-Based Convolutional Neural Network with Multilayer Fusion for SAR", 《REMOTE SENSING》 * |
李宇等: "基于卷积神经网络的光学遥感图像检索", 《光学精密工程》 * |
李铁 等: "高光谱遥感图像空谱联合分类方法研究", 《仪器仪表学报》 * |
陈雯柏: "《人工神经网络原理与实践》", 1 June 2016 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109784375A (en) * | 2018-12-22 | 2019-05-21 | 国网内蒙古东部电力有限公司 | Adaptive transformer part detection recognition method based on Faster RCNN |
CN109583517A (en) * | 2018-12-26 | 2019-04-05 | 华东交通大学 | A kind of full convolution example semantic partitioning algorithm of the enhancing suitable for small target deteection |
CN110009648A (en) * | 2019-03-04 | 2019-07-12 | 东南大学 | Trackside image Method of Vehicle Segmentation based on depth Fusion Features convolutional neural networks |
CN110009648B (en) * | 2019-03-04 | 2023-02-24 | 东南大学 | Roadside image vehicle segmentation method based on depth feature fusion convolutional neural network |
CN111209813A (en) * | 2019-12-27 | 2020-05-29 | 南京航空航天大学 | Remote sensing image semantic segmentation method based on transfer learning |
CN111536991A (en) * | 2020-04-26 | 2020-08-14 | 哈尔滨工业大学 | Remote sensing satellite multi-mode imaging load simulation device |
CN111967734A (en) * | 2020-07-30 | 2020-11-20 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | Method for planning ground observation requirements of micro-nano remote sensing network |
CN111967734B (en) * | 2020-07-30 | 2023-08-25 | 西南电子技术研究所(中国电子科技集团公司第十研究所) | Method for planning earth observation requirement of micro-nano remote sensing network |
CN113558644A (en) * | 2021-07-20 | 2021-10-29 | 陕西科技大学 | Emotion classification method, medium and equipment for 3D matrix and multidimensional convolution network |
CN113558644B (en) * | 2021-07-20 | 2024-03-22 | 陕西科技大学 | Emotion classification method, medium and equipment for 3D matrix and multidimensional convolution network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108664994A (en) | A kind of remote sensing image processing model construction system and method | |
CN109919108B (en) | Remote sensing image rapid target detection method based on deep hash auxiliary network | |
CN104463209B (en) | Method for recognizing digital code on PCB based on BP neural network | |
CN105975931B (en) | A kind of convolutional neural networks face identification method based on multiple dimensioned pond | |
CN104978580B (en) | A kind of insulator recognition methods for unmanned plane inspection transmission line of electricity | |
CN107610087B (en) | Tongue coating automatic segmentation method based on deep learning | |
CN104573731B (en) | Fast target detection method based on convolutional neural networks | |
CN103942577B (en) | Based on the personal identification method for establishing sample database and composite character certainly in video monitoring | |
CN109800628A (en) | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance | |
CN108388896A (en) | A kind of licence plate recognition method based on dynamic time sequence convolutional neural networks | |
CN110097568A (en) | A kind of the video object detection and dividing method based on the double branching networks of space-time | |
CN105825502B (en) | A kind of Weakly supervised method for analyzing image of the dictionary study based on conspicuousness guidance | |
CN108009509A (en) | Vehicle target detection method | |
CN105447473A (en) | PCANet-CNN-based arbitrary attitude facial expression recognition method | |
CN106778835A (en) | The airport target by using remote sensing image recognition methods of fusion scene information and depth characteristic | |
CN106682569A (en) | Fast traffic signboard recognition method based on convolution neural network | |
CN107909015A (en) | Hyperspectral image classification method based on convolutional neural networks and empty spectrum information fusion | |
CN107403200A (en) | Improve the multiple imperfect picture sorting technique of image segmentation algorithm combination deep learning | |
CN112395987B (en) | SAR image target detection method based on unsupervised domain adaptive CNN | |
CN107392224A (en) | A kind of crop disease recognizer based on triple channel convolutional neural networks | |
CN113569667B (en) | Inland ship target identification method and system based on lightweight neural network model | |
CN110543906B (en) | Automatic skin recognition method based on Mask R-CNN model | |
CN106651915A (en) | Target tracking method of multi-scale expression based on convolutional neural network | |
CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
CN108447057A (en) | SAR image change detection based on conspicuousness and depth convolutional network |
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 | ||
CB03 | Change of inventor or designer information |
Inventor after: Xu Yong Inventor after: He Jianting Inventor after: Yu Shiqi Inventor before: Xu Yong |
|
CB03 | Change of inventor or designer information | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181016 |
|
RJ01 | Rejection of invention patent application after publication |