CN110119682A - A kind of infrared remote sensing Image Fire point recognition methods - Google Patents
A kind of infrared remote sensing Image Fire point recognition methods Download PDFInfo
- Publication number
- CN110119682A CN110119682A CN201910270901.5A CN201910270901A CN110119682A CN 110119682 A CN110119682 A CN 110119682A CN 201910270901 A CN201910270901 A CN 201910270901A CN 110119682 A CN110119682 A CN 110119682A
- Authority
- CN
- China
- Prior art keywords
- response diagram
- remote sensing
- obtains
- infrared remote
- sensing image
- 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
- G06F18/21—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
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/188—Vegetation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of infrared remote sensing Image Fire point recognition methods, for the problem of spatial information deficiency in network, use original infrared remote sensing image as input, the space characteristics between each fiery point target are expressed by the Multilevel Response figure of acquisition, using convolution results at different levels as the textural characteristics of fiery point target, and the acquisition of rear class response diagram is used as input using prime response diagram, to merge spatial information and texture information, so that the contextual information for scheming to provide according to response when multistage network model training provides detection clue for the fiery point target for being not easy to detect, multistage network model is effectively improved to the discrimination of wood land infrared remote sensing image moderate heat point target.
Description
Technical field
The invention belongs to images steganalysis field more particularly to a kind of infrared remote sensing Image Fire point recognition methods.
Background technique
Satellite remote sensing technology has the characteristics that wide observation scope, acquisition abundant information, repeated measures ability are strong, utilizes remote sensing
Image is monitored forest fire, fire point position and variation, accurate evaluation fire damage and influence can be found in time, for dividing
Cloth range is wide and sends out the forest fire that calamity area should not approach, and remote sensing images forest fire more has unique and important excellent
Gesture.
It is mainly at present the conventional high-temperature target identification based on shallow-layer feature for remote sensing images fire point target detection method
Method and machine learning method based on convolutional neural networks.The former is based primarily upon shallow-layer feature and carries out fiery point target detection, often
Infrared heat target identification method has fiery vertex degree (NDFI) method of normalization, the polynary truncation method of mahalanobis distance, mahalanobis distance
Multicategory discriminant method, factor analysis etc..Such method often has preferable discrimination for simple scenario target, however it is constructed
Shallow-layer feature it is not fine enough to scene description, be difficult effectively adapt to all scenes (especially complex scene), and
Conventional method such as normalizes interference of fiery vertex degree (NDFI) method vulnerable to water body, color steel sheet roofing, it is difficult to effectively identify fiery point
Target.
Based on the machine learning method of convolutional neural networks due to its powerful feature extraction and scene objects capability of fitting,
More popular method is used as remote sensing images field in recent years.This method does not need artificial pointedly design feature, and
That advanced complicated semantic feature is gradually constructed by shallow-layer feature by layer-by-layer convolution, the alternate mode of pondization, to image object or
Scene has stronger feature descriptive power, therefore is widely used in image classification identification field.However, for more similar
Difficulty point different type target or scene, simple depth convolutional network still suffers from higher misrecognition for small data set
Rate.
Summary of the invention
To solve the above problems, the present invention provides a kind of infrared remote sensing Image Fire point recognition methods, can effectively improve more
Discrimination of the grade network model to wood land infrared remote sensing image moderate heat point target.
A kind of infrared remote sensing Image Fire point recognition methods, using multistage network model to the fire point mesh in infrared remote sensing image
Mark is identified, wherein multistage network model is at least three-level, acquisition methods specifically:
S1: using the infrared remote sensing image of wood land as training sample, and the position of training sample moderate heat point target is obtained
It sets；
S2: VGG19 network is trained using the infrared remote sensing image obtained at random on ImageNet data set, is obtained
The core size of the convolution kernel of VGG19 network；
S3: using the core size of the convolution kernel of the obtained VGG19 network of step S2 as the initial ginseng of first order convolutional network
Then number is trained first order convolutional network using part training sample according to setting ratio, obtains first order response diagram；
Wherein, response diagram includes several subgraphs；
S4: loss function is executed to first order response diagram and obtains operation, obtains first order loss function, the loss function
Obtain operation are as follows:
Obtain the same level response diagram each width subgraph maximum eigenvalue point, then using the position of the maximum eigenvalue point as
The position of fiery point target in each width subgraph；
By the quadratic sum of the location error of the fiery point target obtained in the position of obtained each fiery point target and step S1
As loss function；
S5: building second level convolutional network, wherein second level convolutional network includes at least two-stage convolutional layer；According to setting
Ratio, selected part training sample is trained second level convolutional network again, obtains the second level by all convolutional layers
Characteristic pattern and the intermediate features figure exported from any level-one convolutional layer；Then by second level characteristic pattern and first order response diagram into
Row series connection obtains series connection characteristic pattern；To connect characteristic pattern and default convolution kernel progress convolution, obtain second level response diagram；To
Secondary response figure executes loss function and obtains operation, obtains second level loss function；
S6: being divided into M sub- characteristic patterns for intermediate features figure, then obtains M corresponding M grades of sub- characteristic patterns respectively and responds
Figure, wherein M determines by the multiple proportion between the size of intermediate features figure and the size of response diagram, and response diagram at different levels obtains
Take method specifically:
Subcharacter figure is connected with prime response diagram respectively, then will connect result and default convolution kernel progress convolution,
Obtain the corresponding response diagram of subcharacter figure；Wherein, the prime response diagram of first sub- characteristic pattern is second level response diagram, remaining
The prime response diagram of each subcharacter figure is the corresponding response diagram of its previous subcharacter figure；
S7: response diagram at different levels corresponding to each subcharacter figure executes loss function and obtains operation respectively, obtains each subcharacter
Scheme corresponding M grades of loss function；Meanwhile by the position of the corresponding fiery point target of afterbody loss function, as finally obtaining
Fiery point target position, and the convolution kernel that uses collectively forms the multistage network model when obtaining all grades of response diagrams；
S8: all grades of loss function is added, the total losses of multistage network model is obtained；Then judge the total losses
Whether given threshold is less than, if being less than, the network obtained at this time is final multistage network model, if being not less than, into step
Rapid S9；
S9: total losses is used for backpropagation, the convolution that adjustment multistage network model is used when obtaining response diagram at different levels
The core size of core；Then the infrared remote sensing image of wood land is reacquired as new training sample, using volume adjusted
The core size for the convolution kernel that the core size of product core uses when obtaining response diagram at different levels before replacing respectively, repeats step S3~S8,
Total losses is obtained again, until total losses is less than given threshold.
Further, before in step S1 using the infrared remote sensing image of wood land as training sample, first by affine
It converts and data augmentation is carried out to the infrared remote sensing image of wood land, increase the quantity of the infrared remote sensing image of wood land, then
Using the infrared remote sensing image of all wood lands obtained after augmentation as training sample, wherein the affine transformation includes rotation
Turn, translation and small distortion.
Optionally, the M=4, and the acquisition methods of the corresponding 4 grades of response diagrams of 4 sub- characteristic patterns specifically:
Choose one of subcharacter figure and connect with the second level response diagram that step S5 is obtained, then will series connection result with
Default convolution kernel carries out convolution, obtains third level response diagram；
Choose second sub- characteristic pattern, connect with third level response diagram, then will series connection result and default convolution kernel into
Row convolution obtains fourth stage response diagram；
Choose third sub- characteristic pattern, connect with fourth stage response diagram, then will series connection result and default convolution kernel into
Row convolution obtains level V response diagram；
Choose the 4th sub- characteristic pattern, connect with level V response diagram, then will series connection result and default convolution kernel into
Row convolution obtains the 6th grade of response diagram.
Further, a kind of infrared remote sensing Image Fire point recognition methods, further comprising the steps of:
S10: the infrared remote sensing image of wood land is reacquired as test sample, while obtaining the fire in test sample
Point target position；
S11: detection identification is carried out to test sample using the multistage network model that step S8 is obtained, then will test identification
The position of obtained fiery point target is compared with the obtained fiery point target position step S10, obtains the identification of multistage network model
Rate；
S12: discrimination is divided into " 0.85 or more ", " 0.75~0.85 " and " 0.75 or less " three classifications, and is obtained
The corresponding sample of each classification and sample size；
S13: by affine transformation by the sample size augmentation of " 0.75~0.85 " and " 0.75 or less " two classifications extremely
The 2/3 of the sample size of " 0.85 or more " classification；
S14: by the sample of the sample of " 0.85 or more " classification and " 0.75~0.85 " and " 0.75 or less " classification after augmentation
Originally it merges, obtains total sample, then using total sample as training sample；
S15: the training sample that step S14 is obtained replaces the sample in step S1, repeats step S1~S9, obtains
The higher multistage network model of discrimination.
The utility model has the advantages that
The present invention provides a kind of infrared remote sensing Image Fire point recognition methods, for the problem of spatial information deficiency in network,
Use original infrared remote sensing image as input, the space expressed between each fiery point target by the Multilevel Response figure of acquisition is special
Sign, using convolution results at different levels as the textural characteristics of fiery point target, and the acquisition of rear class response diagram uses the conduct of prime response diagram
Input, so that spatial information and texture information are merged, so that scheming the context provided according to response when multistage network model training
Information provides detection clue for the fiery point target for being not easy to detect, and effectively improves multistage network model to wood land infrared remote sensing figure
As the discrimination of moderate heat point target；
Meanwhile the present invention cannot be guaranteed discrimination for the problem that shallow-layer network, and deep layer network is difficult to optimize, by net
Network is divided into multistage, and obtains the loss function of networks at different levels, and the ginseng of multistage network model is reversely adjusted finally by total losses
Number, belongs to Training, effectively gradient can be avoided to disappear, so that the recognition capability of finally obtained multistage network model is more
By force.
Detailed description of the invention
Fig. 1 is a kind of flow chart of infrared remote sensing Image Fire point recognition methods provided by the invention；
Fig. 2 is the structural schematic diagram of multistage network model provided by the invention；
Fig. 3 is the corresponding relationship of response diagrams at different levels and loss functions at different levels provided by the invention；
Fig. 4 is the infrared remote sensing image schematic diagram of wood land provided by the invention；
Fig. 5 is the infrared remote sensing image schematic diagram of another wood land provided by the invention.
Specific embodiment
In order to make those skilled in the art more fully understand application scheme, below in conjunction in the embodiment of the present application
Attached drawing, the technical scheme in the embodiment of the application is clearly and completely described.
Embodiment one
Referring to Fig. 1, which is a kind of flow chart of infrared remote sensing Image Fire point recognition methods provided in this embodiment.It is a kind of
Infrared remote sensing Image Fire point recognition methods identifies the fiery point target in infrared remote sensing image using multistage network model,
Wherein, multistage network model is at least three-level, acquisition methods specifically:
S1: using the infrared remote sensing image of wood land as training sample, and the position of training sample moderate heat point target is obtained
It sets.
Optionally, before using the infrared remote sensing image of wood land as training sample, first by affine transformation to forest
The infrared remote sensing image in region carries out data augmentation, increases the quantity of the infrared remote sensing image of wood land, then will obtain after augmentation
The infrared remote sensing image for all wood lands arrived is as training sample, wherein the affine transformation include rotation, translation and
Small distortion.
S2: VGG19 network is trained using the infrared remote sensing image obtained at random on ImageNet data set, is obtained
The core size of the convolution kernel of VGG19 network.
S3: using the core size of the convolution kernel of the obtained VGG19 network of step S2 as the initial ginseng of first order convolutional network
Then number is trained first order convolutional network using part training sample according to setting ratio, obtains first order response diagram；
Wherein, response diagram includes several subgraphs.
It should be noted that response diagram is 3-D image, wherein bidimensional is Size dimensional, is left one-dimensional for description fire point mesh
Target characteristic response dimension is sliced response diagram from each sampled point in characteristic response dimension, and each slice is to respond
A width subgraph in figure, and each subgraph describes the characteristic response of fiery point target.
S4: loss function is executed to first order response diagram and obtains operation, obtains first order loss function, the loss function
Obtain operation are as follows:
Obtain the same level response diagram each width subgraph maximum eigenvalue point, then using the position of the maximum eigenvalue point as
The position of fiery point target in each width subgraph；
By the quadratic sum of the location error of the fiery point target obtained in the position of obtained each fiery point target and step S1
As loss function.
S5: building second level convolutional network, wherein second level convolutional network includes at least two-stage convolutional layer；According to setting
Ratio, selected part training sample is trained second level convolutional network again, obtains the second level by all convolutional layers
Characteristic pattern and the intermediate features figure exported from any level-one convolutional layer；Then by second level characteristic pattern and first order response diagram into
Row series connection obtains series connection characteristic pattern；To connect characteristic pattern and default convolution kernel progress convolution, obtain second level response diagram；To
Secondary response figure executes loss function and obtains operation, obtains second level loss function.
S6: being divided into M sub- characteristic patterns for intermediate features figure, then obtains M corresponding M grades of sub- characteristic patterns respectively and responds
Figure, wherein M determines by the multiple proportion between the size of intermediate features figure and the size of response diagram, and response diagram at different levels obtains
Take method specifically:
Subcharacter figure is connected with prime response diagram respectively, then will connect result and default convolution kernel progress convolution,
Obtain the corresponding response diagram of subcharacter figure；Wherein, the prime response diagram of first sub- characteristic pattern is second level response diagram, remaining
The prime response diagram of each subcharacter figure is the corresponding response diagram of its previous subcharacter figure.
For example, in the case that intermediate features figure is divided into 4 sub- characteristic patterns, and 4 corresponding 4 grades of sub- characteristic patterns respond
The acquisition methods of figure specifically:
Choose one of subcharacter figure and connect with the second level response diagram that step S5 is obtained, then will series connection result with
Default convolution kernel carries out convolution, obtains third level response diagram；
Choose second sub- characteristic pattern, connect with third level response diagram, then will series connection result and default convolution kernel into
Row convolution obtains fourth stage response diagram；
Choose third sub- characteristic pattern, connect with fourth stage response diagram, then will series connection result and default convolution kernel into
Row convolution obtains level V response diagram；
Choose the 4th sub- characteristic pattern, connect with level V response diagram, then will series connection result and default convolution kernel into
Row convolution obtains the 6th grade of response diagram.
S7: response diagram at different levels corresponding to each subcharacter figure executes loss function and obtains operation respectively, obtains each subcharacter
Scheme corresponding M grades of loss function；Meanwhile by the position of the corresponding fiery point target of afterbody loss function, as finally obtaining
Fiery point target position, and the convolution kernel that uses collectively forms the multistage network model when obtaining all grades of response diagrams.
S8: all grades of loss function is added, the total losses of multistage network model is obtained；Then judge the total losses
Whether given threshold is less than, if being less than, the network obtained at this time is final multistage network model, if being not less than, into step
Rapid S9.
S9: total losses is used for backpropagation, the convolution that adjustment multistage network model is used when obtaining response diagram at different levels
The core size of core；Then the infrared remote sensing image of wood land is reacquired as new training sample, using volume adjusted
The core size for the convolution kernel that the core size of product core uses when obtaining response diagram at different levels before replacing respectively, repeats step S3~S8,
Total losses is obtained again, until total losses is less than given threshold.
It should be noted that backpropagation is a kind of common method for adjusting parameter model.In forward-propagating process
In, that is, step S1~S8, information is inputted by input layer through hidden layer, is successively handled and is transmitted to output layer.If defeated
Layer cannot get desired output valve out, then takes the quadratic sum of output and desired error as objective function, this is also loss letter
Several extensive definition, is transferred to backpropagation, successively finds out objective function to the partial derivative of each neuron weight, constitutes objective function
The ladder amount of weight vector, that is, the existing gradient descent method as modification weight foundation, the study of network are repaired in weight
It is completed during changing, when error reaches desired value, e-learning terminates.
In order to further obtain the higher multistage network model of discrimination, after executing the step S9, can also continue to execute
Following steps:
S10: the infrared remote sensing image of wood land is reacquired as test sample, while obtaining the fire in test sample
Point target position；
S11: detection identification is carried out to test sample using the multistage network model that step S8 is obtained, then will test identification
The position of obtained fiery point target is compared with the obtained fiery point target position step S10, obtains the identification of multistage network model
Rate；
S12: discrimination is divided into " 0.85 or more ", " 0.75~0.85 " and " 0.75 or less " three classifications, and is obtained
The corresponding sample of each classification and sample size；
S13: by affine transformation by the sample size augmentation of " 0.75~0.85 " and " 0.75 or less " two classifications extremely
The 2/3 of the sample size of " 0.85 or more " classification；
S14: by the sample of the sample of " 0.85 or more " classification and " 0.75~0.85 " and " 0.75 or less " classification after augmentation
Originally it merges, obtains total sample, then using total sample as training sample；
S15: the training sample that step S14 is obtained replaces the sample in step S1, repeats step S1~S9, obtains
The higher multistage network model of discrimination.
Embodiment two
It is tested using 10 meters of spatial resolution infrared remote sensing image data sets as experimental subjects, specific implementation step is such as
Under:
S1: sample number is used as using the fiery point target slice of 10 meters of 368 pixel * of resolution ratio, 368 pixel sizes of infrared image
According to, and data augmentation is carried out, sample is then divided into training sample and verifying sample set, specifically:
S11: data set sample has 200 width images, carries out augmentation to each image by the operation such as rotation, translation, scaling.
S12: the data set sample after augmentation is divided by training sample set and verifying sample set with the ratio of 7:3, and by two
Person's scramble is packaged into the data format of network needs.
S2: the VGG19 network trained on ImageNet large data sets is trained as base net network.
S3: the network hyper parameter that initialization S2 is obtained, hyper parameter includes the core size of convolution kernel, and every batch of amount of training data is
20.As shown in Fig. 2, the figure is the structural schematic diagram of multistage network model provided in this embodiment.The first order is a base volume
Product network (convs), each target response is directly predicted from original image.Convolutional layer structure includes 7 layers of convolution sum, 3 layers of pond,
Being originally inputted picture is 368*368, obtains 46*46 size original image by convolutional layer, the present embodiment network by 3 ponds
As long as training recognition result be set as identifying in infrared remote sensing image most apparent nine fiery point targets of feature, in addition
Comprising a background response, totally 10 layers of response diagram, therefore obtaining first order response diagram size is 46*46*10.Then to the first order
Response diagram executes loss function and obtains operation, obtains first order loss function.
That is, response diagram is the 3-D image that size is 46*46*10,46*46 is Size dimensional, is left one-dimensional to be to retouch
The characteristic response dimension of fiery point target is stated, i.e., has 10 sampled points, corresponding 10 width subgraphs, wherein there are 9 width subgraphs in characteristic response dimension
A fiery point target is respectively corresponded, a remaining width subgraph corresponds to background.
S4: second stage predicts each target response also according to original image, is added to a string in convolutional layer middle section
Join layer (contact).Tandem data includes three aspect contents, the stage convolution results (46* obtained based on VGG network structure
46*32, textural characteristics), (46*46*1, Gauss are rung for first order response diagram (46*46*10, space characteristics) and center constraint
Answer), result size constancy after series connection, depth becomes 32+10+1=43, i.e. size is 46*46*43.Then tandem data is based on
Convolutional network exports the second level response diagram (46*46*10) in the stage.Loss function is executed to second level response diagram and obtains behaviour
Make, obtains second level loss function.
S5: the characteristic pattern that depth in second stage calculating process is 128 by the phase III is divided into 4 parts and takes as input
Convolution results of a portion as the phase III, then using the comprehensive three parts of series connection layer: this stage convolution results, second
Phase targets response and center constraint, depth becomes 43 after series connection, i.e. size is 46*46*43.Then tandem data is based on convolution
Network exports third level response diagram (46*46*10).Loss function is executed to third level response diagram and obtains operation, obtains the third level
Loss function.
S6: the four to the 6th stage etch repeats the phase III, is subject to the response diagram of the last stage.Also
It is to say, by the position of the corresponding fiery point target of the 6th grade of loss function, as finally obtained fiery point target position, and obtains institute
The convolution kernel used when having grade response diagram collectively forms the multistage network model；Damage is executed to the four~six grade of response diagram respectively
It loses function and obtains operation, obtain corresponding four~six grade of loss function.As shown in figure 3, the figure is provided in this embodiment each
The corresponding relationship of grade response diagram and loss functions at different levels.
S7: all grades of loss function is added, the total losses of multistage network model is obtained；Then judge the total losses
Whether given threshold is less than, if being less than, the network obtained at this time is final multistage network model, if being not less than, into step
Rapid S8.
S8: total losses is used for backpropagation, the convolution that adjustment multistage network model is used when obtaining response diagram at different levels
The core size of core；Then the infrared remote sensing image of wood land is reacquired as new training sample, using volume adjusted
The core size for the convolution kernel that the core size of product core uses when obtaining response diagram at different levels before replacing respectively, repeats step S3~S8,
Total losses is obtained again, until total losses is less than given threshold.
S9: from the foregoing it will be appreciated that the present embodiment selects suitable n-th generation (here according to training precision and loss function
Selected for the 15th generation) training network parameter composition multistage network model, then reacquire the infrared remote sensing image work of wood land
For test sample, detection identification is carried out to all test samples using multistage network model, records discrimination.
S10: discrimination is divided into " 0.85 or more ", " 0.75~0.85 " and " 0.75 or less " three classifications, and is obtained
The corresponding sample of each classification and sample size.
S11: it is converted using rotation, translation, small distortion equiaffine, according to group data volume situation every in S10 to " 0.75
~0.85 " carries out augmentation with " 0.75 or less " two classifications, obtains new sample set.Specifically:
For the classification of " 0.85 or more ", sample data is more, without augmentation；Remaining two group augmentation to " 0.85 with
On " categorical data amount 2/3.
S12: by the sample of the sample of " 0.85 or more " classification and " 0.75~0.85 " and " 0.75 or less " classification after augmentation
Originally it merges, obtains total sample, then using total sample as training sample.
S13: the training sample that step S12 is obtained replaces the sample in step S1, repeats step S1~S8, obtains
The higher multistage network model of discrimination.
S14: it reacquires test sample and identification verifying is carried out to the multistage network model that step S13 is obtained, output is final
Recognition result.Referring to fig. 4 and Fig. 5, the infrared remote sensing of the wood land under two different scenes respectively provided in this embodiment
Image schematic diagram.Wherein, the box of Fig. 4 and Fig. 5 indicates identification of the multistage network model to infrared remote sensing image moderate heat point target
As a result, it follows that multistage network model provided in this embodiment being capable of infrared remote sensing image moderate heat point mesh to wood land
Mark is effectively identified.
Certainly, the invention may also have other embodiments, without deviating from the spirit and substance of the present invention, ripe
Various corresponding changes and modifications can be made according to the present invention certainly by knowing those skilled in the art, but these it is corresponding change and
Deformation all should fall within the scope of protection of the appended claims of the present invention.
Claims (4)
1. a kind of infrared remote sensing Image Fire point recognition methods, which is characterized in that using multistage network model to infrared remote sensing image
In fiery point target identified, wherein multistage network model is at least three-level, acquisition methods specifically:
S1: using the infrared remote sensing image of wood land as training sample, and the position of training sample moderate heat point target is obtained；
S2: VGG19 network is trained using the infrared remote sensing image obtained at random on ImageNet data set, is obtained
The core size of the convolution kernel of VGG19 network；
S3: using the core size of the convolution kernel of the obtained VGG19 network of step S2 as the initial parameter of first order convolutional network, so
First order convolutional network is trained using part training sample according to setting ratio afterwards, obtains first order response diagram；Wherein,
Response diagram includes several subgraphs；
S4: executing loss function to first order response diagram and obtain operation, obtains first order loss function, and the loss function obtains
Operation are as follows:
The maximum eigenvalue point for obtaining each width subgraph of the same level response diagram, then using the position of the maximum eigenvalue point as each width
The position of fiery point target in subgraph；
Using the quadratic sum of the location error of the fiery point target obtained in the position of obtained each fiery point target and step S1 as
Loss function；
S5: building second level convolutional network, wherein second level convolutional network includes at least two-stage convolutional layer；According to setting ratio,
Again selected part training sample is trained second level convolutional network, obtains the second level characteristic pattern by all convolutional layers
And the intermediate features figure exported from any level-one convolutional layer；Then second level characteristic pattern and first order response diagram are gone here and there
Connection obtains series connection characteristic pattern；To connect characteristic pattern and default convolution kernel progress convolution, obtain second level response diagram；To the second level
Response diagram executes loss function and obtains operation, obtains second level loss function；
S6: being divided into M sub- characteristic patterns for intermediate features figure, then obtains the corresponding M grades of response diagram of M sub- characteristic patterns respectively,
Wherein, M is determined by the multiple proportion between the size of intermediate features figure and the size of response diagram, and the acquisition side of response diagram at different levels
Method specifically:
Subcharacter figure is connected respectively, then will connect result and default convolution kernel progress convolution with prime response diagram, is obtained
The corresponding response diagram of subcharacter figure；Wherein, the prime response diagram of first sub- characteristic pattern is second level response diagram, remaining each son
The prime response diagram of characteristic pattern is the corresponding response diagram of its previous subcharacter figure；
S7: response diagram at different levels corresponding to each subcharacter figure executes loss function and obtains operation respectively, obtains each subcharacter figure pair
The M grade loss function answered；Meanwhile by the position of the corresponding fiery point target of afterbody loss function, as finally obtained fire
Point target position, and the convolution kernel used when all grades of response diagrams of acquisition collectively forms the multistage network model；
S8: all grades of loss function is added, the total losses of multistage network model is obtained；Then whether judge the total losses
Less than given threshold, if being less than, the network obtained at this time enters step S9 if being not less than for final multistage network model；
S9: being used for backpropagation for total losses, the convolution kernel that adjustment multistage network model is used when obtaining response diagram at different levels
Core size；Then the infrared remote sensing image of wood land is reacquired as new training sample, using convolution kernel adjusted
Core size replace respectively before the core size of convolution kernel that uses when obtaining response diagram at different levels, repeat step S3~S8, again
Total losses is obtained, until total losses is less than given threshold.
2. a kind of infrared remote sensing Image Fire point recognition methods as described in claim 1, which is characterized in that by forest in step S1
Before the infrared remote sensing image in region is as training sample, carried out first by infrared remote sensing image of the affine transformation to wood land
Data augmentation, increases the quantity of the infrared remote sensing image of wood land, then by the infrared of all wood lands obtained after augmentation
Remote sensing images are as training sample, wherein the affine transformation includes rotation, translation and small distortion.
3. a kind of infrared remote sensing Image Fire point recognition methods as described in claim 1, which is characterized in that the M=4, and 4
The acquisition methods of the corresponding 4 grades of response diagrams of subcharacter figure specifically:
It chooses one of subcharacter figure and connects with the second level response diagram that step S5 is obtained, then by series connection result and preset
Convolution kernel carries out convolution, obtains third level response diagram；
Second sub- characteristic pattern is chosen, is connected with third level response diagram, then series connection result is rolled up with default convolution kernel
Product, obtains fourth stage response diagram；
The sub- characteristic pattern of third is chosen, is connected with fourth stage response diagram, then series connection result is rolled up with default convolution kernel
Product, obtains level V response diagram；
The 4th sub- characteristic pattern is chosen, is connected with level V response diagram, then series connection result is rolled up with default convolution kernel
Product, obtains the 6th grade of response diagram.
4. a kind of infrared remote sensing Image Fire point recognition methods as described in claim 1, which is characterized in that further include following step
It is rapid:
S10: reacquiring the infrared remote sensing image of wood land as test sample, while obtaining the point mesh of the fire in test sample
Cursor position；
S11: detection identification is carried out to test sample using the multistage network model that step S8 is obtained, identification is then will test and obtains
The position of fiery point target compared with the obtained fiery point target position step S10, obtain the discrimination of multistage network model；
S12: discrimination is divided into " 0.85 or more ", " 0.75~0.85 " and " 0.75 or less " three classifications, and is obtained each
The corresponding sample of classification and sample size；
S13: by affine transformation by the sample size augmentation of " 0.75~0.85 " and " 0.75 or less " two classifications to " 0.85 with
On " sample size of classification 2/3；
S14: by the sample of " 0.85 or more " classification and the sample of " 0.75~0.85 " and " 0.75 or less " classification after augmentation into
Row merges, and obtains total sample, then using total sample as training sample；
S15: the training sample that step S14 is obtained replaces the sample in step S1, repeats step S1~S9, is identified
The higher multistage network model of rate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910270901.5A CN110119682A (en) | 2019-04-04 | 2019-04-04 | A kind of infrared remote sensing Image Fire point recognition methods |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910270901.5A CN110119682A (en) | 2019-04-04 | 2019-04-04 | A kind of infrared remote sensing Image Fire point recognition methods |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110119682A true CN110119682A (en) | 2019-08-13 |
Family
ID=67520826
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910270901.5A Pending CN110119682A (en) | 2019-04-04 | 2019-04-04 | A kind of infrared remote sensing Image Fire point recognition methods |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110119682A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113505627A (en) * | 2021-03-31 | 2021-10-15 | 北京苍灵科技有限公司 | Remote sensing data processing method and device, electronic equipment and storage medium |
| CN114078218A (en) * | 2021-11-24 | 2022-02-22 | 南京林业大学 | Self-adaptive fusion forest smoke and fire identification data augmentation method |
Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104680508A (en) * | 2013-11-29 | 2015-06-03 | 华为技术有限公司 | Convolutional neural network and target object detection method based on convolutional neural network |
| CN105913604A (en) * | 2016-05-18 | 2016-08-31 | 中国计量大学 | Fire occurrence determining method and device based on unmanned aerial vehicle |
| CN107928675A (en) * | 2017-11-22 | 2018-04-20 | 王华锋 | A kind of trunk measuring method being combined based on deep learning and red dot laser |
| CN108205654A (en) * | 2017-09-30 | 2018-06-26 | 北京市商汤科技开发有限公司 | A kind of motion detection method and device based on video |
| CN108256431A (en) * | 2017-12-20 | 2018-07-06 | 中车工业研究院有限公司 | A kind of hand position identification method and device |
| CN108334902A (en) * | 2018-02-02 | 2018-07-27 | 北京华纵科技有限公司 | A kind of track train equipment room smog fireproof monitoring method based on deep learning |
| CN108399378A (en) * | 2018-02-08 | 2018-08-14 | 北京理工雷科电子信息技术有限公司 | A kind of natural scene image recognition methods based on VGG depth convolutional networks |
| CN108537215A (en) * | 2018-03-23 | 2018-09-14 | 清华大学 | A kind of flame detecting method based on image object detection |
| CN108628306A (en) * | 2018-04-10 | 2018-10-09 | 平安科技（深圳）有限公司 | Robot ambulation disorder detection method, device, computer equipment and storage medium |
| CN108717582A (en) * | 2018-04-17 | 2018-10-30 | 平安科技（深圳）有限公司 | Forest fire prediction technique, device, computer equipment and storage medium |
| CN108765335A (en) * | 2018-05-25 | 2018-11-06 | 电子科技大学 | A kind of forest fire detection method based on remote sensing images |
| CN109376681A (en) * | 2018-11-06 | 2019-02-22 | 广东工业大学 | A kind of more people's Attitude estimation method and system |
-
2019
- 2019-04-04 CN CN201910270901.5A patent/CN110119682A/en active Pending
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104680508A (en) * | 2013-11-29 | 2015-06-03 | 华为技术有限公司 | Convolutional neural network and target object detection method based on convolutional neural network |
| CN105913604A (en) * | 2016-05-18 | 2016-08-31 | 中国计量大学 | Fire occurrence determining method and device based on unmanned aerial vehicle |
| CN108205654A (en) * | 2017-09-30 | 2018-06-26 | 北京市商汤科技开发有限公司 | A kind of motion detection method and device based on video |
| CN107928675A (en) * | 2017-11-22 | 2018-04-20 | 王华锋 | A kind of trunk measuring method being combined based on deep learning and red dot laser |
| CN108256431A (en) * | 2017-12-20 | 2018-07-06 | 中车工业研究院有限公司 | A kind of hand position identification method and device |
| CN108334902A (en) * | 2018-02-02 | 2018-07-27 | 北京华纵科技有限公司 | A kind of track train equipment room smog fireproof monitoring method based on deep learning |
| CN108399378A (en) * | 2018-02-08 | 2018-08-14 | 北京理工雷科电子信息技术有限公司 | A kind of natural scene image recognition methods based on VGG depth convolutional networks |
| CN108537215A (en) * | 2018-03-23 | 2018-09-14 | 清华大学 | A kind of flame detecting method based on image object detection |
| CN108628306A (en) * | 2018-04-10 | 2018-10-09 | 平安科技（深圳）有限公司 | Robot ambulation disorder detection method, device, computer equipment and storage medium |
| CN108717582A (en) * | 2018-04-17 | 2018-10-30 | 平安科技（深圳）有限公司 | Forest fire prediction technique, device, computer equipment and storage medium |
| CN108765335A (en) * | 2018-05-25 | 2018-11-06 | 电子科技大学 | A kind of forest fire detection method based on remote sensing images |
| CN109376681A (en) * | 2018-11-06 | 2019-02-22 | 广东工业大学 | A kind of more people's Attitude estimation method and system |
Non-Patent Citations (3)
| Title |
|---|
| YANMIN LUO等: "Fire smoke detection algorithm based on motion characteristic and convolutional neural networks", 《MULTIMED TOOLS APPL》 * |
| 傅天驹等: "复杂背景下基于深度卷积神经网络的森林火灾识别", 《网络出版地址：HTTPS://KNS.CNKI.NET/KCMS/DETAIL/36.1137.TP.20160323.1556.022.HTML》 * |
| 林作永等: "基于深度卷积神经网络的火灾预警算法研究", 《信息通信》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113505627A (en) * | 2021-03-31 | 2021-10-15 | 北京苍灵科技有限公司 | Remote sensing data processing method and device, electronic equipment and storage medium |
| CN114078218A (en) * | 2021-11-24 | 2022-02-22 | 南京林业大学 | Self-adaptive fusion forest smoke and fire identification data augmentation method |
| CN114078218B (en) * | 2021-11-24 | 2024-03-29 | 南京林业大学 | Adaptive fusion forest smoke and fire identification data augmentation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109670528B (en) | Data expansion method facing pedestrian re-identification task and based on paired sample random occlusion strategy | |
| CN111062973B (en) | Vehicle tracking method based on target feature sensitivity and deep learning | |
| CN106203354B (en) | Scene recognition method based on interacting depth structure | |
| CN105184309B (en) | Classification of Polarimetric SAR Image based on CNN and SVM | |
| CN105528638B (en) | The method that gray relative analysis method determines convolutional neural networks hidden layer characteristic pattern number | |
| CN110348399B (en) | Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network | |
| CN109902715B (en) | Infrared dim target detection method based on context aggregation network | |
| CN105975931A (en) | Convolutional neural network face recognition method based on multi-scale pooling | |
| CN107729819A (en) | A kind of face mask method based on sparse full convolutional neural networks | |
| CN107341518A (en) | A kind of image classification method based on convolutional neural networks | |
| CN103226820B (en) | The two-dimensional maximum entropy division night vision image fusion target detection algorithm improved | |
| CN107844795A (en) | Convolutional neural networks feature extracting method based on principal component analysis | |
| CN110472667A (en) | Small object classification method based on deconvolution neural network | |
| CN106204779A (en) | The check class attendance method learnt based on plurality of human faces data collection strategy and the degree of depth | |
| CN108665005A (en) | A method of it is improved based on CNN image recognition performances using DCGAN | |
| CN109785344A (en) | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration | |
| CN113850242B (en) | Storage abnormal target detection method and system based on deep learning algorithm | |
| CN108596818A (en) | A kind of image latent writing analysis method based on multi-task learning convolutional neural networks | |
| CN106874952A (en) | Feature fusion based on stack self-encoding encoder | |
| CN109886267A (en) | A kind of soft image conspicuousness detection method based on optimal feature selection | |
| CN107784308A (en) | Conspicuousness object detection method based on the multiple dimensioned full convolutional network of chain type | |
| CN109753864A (en) | A kind of face identification method based on caffe deep learning frame | |
| CN110119682A (en) | A kind of infrared remote sensing Image Fire point recognition methods | |
| CN110751195A (en) | Fine-grained image classification method based on improved YOLOv3 | |
| CN109598711A (en) | A kind of thermal image defect extracting method based on feature mining and neural 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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190813 |
|
| WD01 | Invention patent application deemed withdrawn after publication |