CN109389593A - A kind of detection method, device, medium and the equipment of infrared image Small object - Google Patents
A kind of detection method, device, medium and the equipment of infrared image Small object Download PDFInfo
- Publication number
- CN109389593A CN109389593A CN201811166867.9A CN201811166867A CN109389593A CN 109389593 A CN109389593 A CN 109389593A CN 201811166867 A CN201811166867 A CN 201811166867A CN 109389593 A CN109389593 A CN 109389593A
- Authority
- CN
- China
- Prior art keywords
- residual error
- error network
- notable
- depth residual
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10048—Infrared image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides detection method, device, medium and the equipment of a kind of infrared image Small object.The method, comprising: extract the characteristics of image of source images；Residual error network operations are carried out to described image feature using regression depth residual error network, the notable figure after obtaining clutter reduction background；Extract the candidate target in the notable figure；Classified using classifying type depth residual error network to the candidate target, obtains target detection notable figure.By the way that regression depth residual error network and classifying type depth residual error network integration get up to handle source images, it can preferably inhibit background clutter and noise jamming, obtain the notable figure that object edge is more clear.
Description
Technical field
The present invention relates to technical field of image processing, and in particular to a kind of detection method of infrared image Small object, device,
Medium and equipment.
Background technique
Infrared imaging mechanism determines that the target in infrared image lacks complete texture and shape information, while Small object exists
Shared number of pixels is less in infrared image, is almost in dotted, and due to detecting at a distance and Atmospheric propagating effects
It influences, is easily flooded by noise and background clutter.In order to inhibit background clutter and noise jamming, the essence of detection algorithm is improved
Degree, researcher constantly propose all kinds of infrared small target detection methods.
Summary of the invention
For the defects in the prior art, the present invention provides a kind of detection method of infrared image Small object, device, medium
And equipment, it can preferably inhibit background clutter and noise jamming, obtain the notable figure that object edge is more clear.
In a first aspect, the present invention provides a kind of detection methods of infrared image Small object, comprising:
Extract the characteristics of image of source images；
Residual error network operations are carried out to described image feature using regression depth residual error network, obtain clutter reduction background
Notable figure afterwards；
Extract the candidate target in the notable figure；
Classified using classifying type depth residual error network to the candidate target, obtains target detection notable figure.
Optionally, the characteristics of image for extracting source images, comprising:
Using multi-channel Gabor filtering, the multidimensional image feature of source images is extracted.
Optionally, the regression depth residual error network is free of pond layer；The classifying type depth residual error network is free of pond
Change layer.
Optionally, residual error network operations are carried out to described image feature using regression depth residual error network described, obtained
Before the step of obtaining the notable figure of clutter reduction background, further includes:
Obtain sample infrared image；
Using sliding window mode, traversal cutting is carried out to the sample infrared image, obtains training image collection；
The scene information that each training image is concentrated according to the training image carries out contingency table to the training image collection
Note；
According to the training image collection after classification annotation, to regression depth residual error network and classifying type depth residual error network into
Row pre-training, the network parameter of optimized regression moldeed depth degree residual error network and classifying type depth residual error network.
Optionally, the candidate target extracted in the notable figure, comprising:
Using threshold processing method, the candidate target in the notable figure is extracted.
Optionally, described to be classified using classifying type depth residual error network to the candidate target, obtain target detection
Notable figure, comprising:
Using four neighborhood clustering methods, the neighborhood image data of each candidate target are extracted；
According to the neighborhood image data, is classified using classifying type depth residual error network to the candidate target, obtained
Obtain target detection notable figure.
Optionally, in the external mean square error loss function of output layer of the regression depth residual error network；
According to the loss function, the quality for predicting the regression depth residual error network is measured.
Second aspect, the present invention provide a kind of detection device of infrared image Small object, comprising:
Characteristic extracting module, for extracting the characteristics of image of source images；
Clutter recognition module is pressed down for being handled using regression depth residual error network described image feature
Notable figure after clutter background processed；
Object extraction module, for extracting the candidate target in the notable figure；
Categorization module obtains target inspection for classifying using classifying type depth residual error network to the candidate target
Survey notable figure.
The third aspect, the present invention provide a kind of computer readable storage medium, are stored thereon with computer program, the program
The detection method of one of first aspect infrared image Small object is realized when being executed by processor.
Fourth aspect, the present invention provide a kind of detection device of infrared image Small object, comprising: memory, processor and
The computer program that can be run on a memory and on a processor is stored, the processor realizes first when executing described program
The detection method of one of aspect infrared image Small object.
The present invention in the way of end-to-end by detecting candidate target by regression depth residual error network, to whole picture input
Source images execute disposable residual error network operations, the notable figure after obtaining clutter reduction background, then extract the candidate in notable figure
The input data amount of subsequent classification moldeed depth degree residual error network can be greatly lowered in object, using classifying type depth residual error network energy
Realize the high-precision identification of candidate target.By the way that regression depth residual error network and classifying type depth residual error network integration are got up
Source images are handled, can preferably inhibit background clutter and noise jamming, obtain the notable figure that object edge is more clear.
Detection device, a kind of computer readable storage medium and one kind of a kind of infrared image Small object provided by the invention
The detection device of infrared image Small object, with a kind of above-mentioned detection method of infrared image Small object for identical invention structure
Think, beneficial effect having the same.
Detailed description of the invention
It, below will be to specific in order to illustrate more clearly of the specific embodiment of the invention or technical solution in the prior art
Embodiment or attached drawing needed to be used in the description of the prior art are briefly described.In all the appended drawings, similar element
Or part is generally identified by similar appended drawing reference.In attached drawing, each element or part might not be drawn according to actual ratio.
Fig. 1 is a kind of flow chart of the detection method of infrared image Small object provided by the invention；
Fig. 2 is a kind of flow diagram of the detection method of infrared image Small object provided by the invention；
Fig. 3 is a kind of effect diagram of infrared image small target deteection provided by the invention；
Fig. 4 is a kind of schematic diagram of the detection device of infrared image Small object provided by the invention.
Specific embodiment
It is described in detail below in conjunction with embodiment of the attached drawing to technical solution of the present invention.Following embodiment is only used for
Clearly illustrate technical solution of the present invention, therefore be intended only as example, and cannot be used as a limitation and limit protection of the invention
Range.
It should be noted that unless otherwise indicated, technical term or scientific term used in this application should be this hair
The ordinary meaning that bright one of ordinary skill in the art are understood.
The present invention provides detection method, device, medium and the equipment of a kind of infrared image Small object.With reference to the accompanying drawing
The embodiment of the present invention is illustrated.
Referring to FIG. 1, Fig. 1 is a kind of stream of the detection method for infrared image Small object that the specific embodiment of the invention provides
Cheng Tu, a kind of detection method of infrared image Small object provided in this embodiment, comprising:
Step S101: the characteristics of image of source images is extracted.
Step S102: residual error network operations are carried out to described image feature using regression depth residual error network, are pressed down
Notable figure after clutter background processed.
Step S103: the candidate target in the notable figure is extracted.
Step S104: classifying to the candidate target using classifying type depth residual error network, and it is aobvious to obtain target detection
Write figure.
By detecting candidate target in the way of end-to-end by regression depth residual error network, to the source images of whole picture input
Disposable residual error network operations are executed, the notable figure after obtaining clutter reduction background, then extract the candidate target in notable figure, energy
The input data amount of subsequent classification moldeed depth degree residual error network is greatly lowered, candidate is able to achieve using classifying type depth residual error network
The high-precision of object identifies.By getting up to handle source figure for regression depth residual error network and classifying type depth residual error network integration
Picture can preferably inhibit background clutter and noise jamming, obtain the notable figure that object edge is more clear.
In a specific embodiment provided by the invention, the characteristics of image for extracting source images, further includes: utilize more
Channel Gabor filter extracts the multidimensional image feature of source images.
Wherein, characteristics of image may include multidimensional image feature.
For example, source images are by can be obtained the characteristics of image of four direction after Gabor filter, comprising: θ=0 °,
45 °, 90 °, 135 ° } four direction.
Two-dimensional Gabor function expression are as follows:
Wherein, x', y' are defined as follows:
X'=x cos θ+y sin θ
Y'=-x sin θ+y cos θ
Wherein, g is for the output valve for the Gaussian function modulated.γ is space length ratio (spatial aspect
Ratio), the ellipticity of Gabor function is determined.σ is the standard deviation of Gaussian function.λ be harmonic factor wavelength parameter, ratio σ/
λ is the band bandwidth for determining the space of Gabor filter.Value range beIndicate that harmonic factor is auxiliary
The phase offset at angle decides the symmetry of Gabor function.Wherein, θ (θ ∈ [0, π)) direction of filter is specified, it is fixed
Justice is the angle that Gaussian function deviates principal axes line.
Again using the characteristics of image of four direction as the input of regression depth residual error network, which is that four-way end is arrived
Network is held, clutter background is executed and inhibits retention point target blob picture, and the output picture size after input picture and clutter recognition
It is consistent.In output layer, Fusion Features are carried out to the image of four dimensions, finally export a notable figure.
In the present invention, regression depth residual error network is free of pond layer, and regression depth residual error network is with end-to-end
Form inhibit picture noise background, to highlight target signature.By taking single channel regression as an example, convolution kernel is having a size of 1 × 1+3
× 3+1 × 1, the activation primitive of each convolutional layer are ReLU (amendment linear unit), and output layer contains 1 width characteristic pattern.Due to the network
Without pond layer, input picture and each network layer characteristic pattern resolution ratio are remained unchanged, and small size training data can be used and carry out net
Network training, the network after training can carry out feedforward operation to arbitrary dimension image.Finally utilize improvement multichannel regression depth
Residual error network obtains notable figure.
In a specific embodiment provided by the invention, utilize regression depth residual error network to described image described
Before the step of feature is handled, and the notable figure of clutter reduction background is obtained, further includes: obtain sample infrared image；Using
Sliding window mode carries out traversal cutting to the sample infrared image, obtains training image collection；It is concentrated according to the training image every
The scene information of a training image carries out classification annotation to the training image collection；According to the training image collection after classification annotation,
Pre-training carried out to regression depth residual error network and classifying type depth residual error network, optimized regression moldeed depth degree residual error network and point
The network parameter of type depth residual error network.
Before being detected using regression depth residual error network and classifying type depth residual error network to source images, need
The two networks are trained, network parameter is optimized.
Since the two networks do not need pond layer, before training optimizes network, so that it may remove pond layer
Fall, to improve optimization efficiency.
Optimize detailed process are as follows: obtain sample infrared image, training is obtained by sliding window mode on sample infrared image
Image, sliding window size can be taken as 18 × 18, by adding varying strength oversampled points target in the picture, generate 10,000
Group training image.
Wherein, sample infrared image can choose sea and sky background infrared image.
Background, target, target context mixing three classes training are classified as further according to the scene information of training image reflection
Image simultaneously marks.In classification, can be classified using MATTLAB, in mark, can mark background is 0, target 1,
Target context is mixed into 2, distinguishes three different scene informations according to number.
Using sorted training image as the input of regression depth residual error network, to regression depth residual error network into
Row training, optimizes its network parameter.The notable figure of sample infrared image can be obtained by regression depth residual error network.To this
Notable figure carries out threshold processing, extracts candidate target, and extract candidate target neighborhood image data.
Using candidate target neighborhood image data as the input of classifying type depth residual error network, to classifying type depth residual error net
Network is trained, and optimizes its network parameter.
Wherein, in the external mean square error loss function of output layer of regression depth residual error network, formula is as follows:
Y' in formulaiFor i-th of data true value in batch training data, yiFor neural network forecast value.N is the number of output data.Its
In, yi,y'i∈Rm, m is picture size in training data.L is the assessed value for predicting clutter reduction background.The value can predict mould
The quality of type measures the performance of model.
In the network parameter of optimized regression moldeed depth degree residual error network, network parameter can be optimized in conjunction with the assessed value,
Improve the performance of regression depth residual error network.
In a specific embodiment provided by the invention, the candidate target extracted in the notable figure be can wrap
It includes: using threshold processing method, extracting the candidate target in the notable figure.
By the way that threshold value is arranged, the pixel of notable figure is judged, the candidate target in notable figure is extracted.
Threshold processing formula is as follows:
Th=μimg+k·σimg
In formula, Th is threshold value, μimgFor the gray value mean value of notable figure after clutter reduction background image, σimgTo inhibit
The gray value mean square deviation of notable figure after clutter background image, k >=3, I (x, y), T (x, y) are respectively to inhibit miscellaneous at coordinate (x, y)
The sum of the grayscale values of notable figure crosses threshold image value after wave background image.
When the threshold value of specific image vegetarian refreshments is greater than threshold value, judge the pixel for the pixel of candidate target；
When the threshold value of specific image pixel is not more than threshold value, judge the pixel for background pixel point.
The candidate target in notable figure can be extracted by this method.When extracting candidate target, 9X9 size can be extracted
Image, facilitate input classifying type depth residual error network.When there is a candidate target in notable figure, the image of extraction is one
The image of candidate target.When there is multiple candidate targets in notable figure, the image of extraction is the image of multiple candidate targets, also
It is the image of multiple 9X9 sizes, wherein may include the candidate target image of false-alarm targets and real goal.
It is described right to the candidate using classifying type depth residual error network in a specific embodiment provided by the invention
As classifying, target detection notable figure is obtained, may include: that it is right to extract each candidate using four neighborhood clustering methods
The neighborhood image data of elephant；According to the neighborhood image data, using classifying type depth residual error network to the candidate target into
Row classification, obtains target detection notable figure.
When extracting the image of the corresponding 9X9 size of candidate target, need using four neighborhood clustering methods, to each candidate
Object images cluster calculates candidate target estimated location with gray value weighting scheme, and centered on the position, and it is candidate right to extract
The neighborhood image data of elephant, then using the neighborhood image data of extraction as the input of classifying type depth residual error network, to all
Candidate target is classified, and is rejected the false-alarm targets in candidate target, is finally obtained target detection notable figure.
Wherein, classifying type depth residual error network includes convolutional layer and 1 layer of full articulamentum, and convolution kernel is having a size of 1 × 1+3 × 3+
1 × 1, the activation primitive of each convolutional layer is ReLU.The network is free of pond layer, so that being avoided that reduces the significant of Small object feature
Property.Full articulamentum is followed by output layer.The neighborhood image data of candidate target after simple threshold detection positioning are input to point
Type depth residual error network, the network reject false-alarm targets from candidate target and obtain classification results, and finally obtain target inspection
Survey notable figure.
Classifying type depth residual error network needs the external Softmax of full articulamentum in a network when realizing target classification
Function, formula are as follows:
Wherein, z=[z1z2…zc]TFor the output of full articulamentum, c is the number of types that network differentiates.L is Softmax function
Value.
As shown in figure 3, for the comparison result of the present invention and two methods in the prior art, wherein (a-i) is actual photographed
Infrared image result figure；(b-i) for using the result figure of robustness principal component analytical method (RPCA)；It (c-i) is use
The result figure of intensive sparse reconstruction method (DSR)；(d-i) for using the result figure of the method for the present invention；Wherein, i=1,2,3,4.
(a-1) the aerial single goal infrared image of actual photographed is indicated；(a-2) it indicates to use robustness principal component analysis side
The aerial single goal infrared image that method (RPCA) obtains；(a-3) it indicates to obtain using intensive sparse reconstruction method (DSR) aerial
Single goal infrared image；(a-4) the aerial single goal infrared image obtained using the method for the present invention is indicated.
(b-1) 3D rendering of the aerial single goal infrared image of actual photographed is indicated；(b-2) indicate using robustness it is main at
The 3D rendering for the aerial single goal infrared image for dividing analysis method (RPCA) to obtain；(b-3) it indicates using intensive sparse reconstruction side
The 3D rendering for the aerial single goal infrared image that method (DSR) obtains；(b-4) the aerial monocular obtained using the method for the present invention is indicated
Mark the 3D rendering of infrared image.
(c-1) the marine single goal infrared image of actual photographed is indicated；(c-2) it indicates to use robustness principal component analysis side
The marine single goal infrared image that method (RPCA) obtains；(c-3) sea obtained using intensive sparse reconstruction method (DSR) is indicated
Single goal infrared image；(c-4) the marine single goal infrared image obtained using the method for the present invention is indicated.
(d-1) 3D rendering of the marine single goal infrared image of actual photographed is indicated；(d-2) indicate using robustness it is main at
The 3D rendering for the marine single goal infrared image for dividing analysis method (RPCA) to obtain；(d-3) it indicates using intensive sparse reconstruction side
The 3D rendering for the marine single goal infrared image that method (DSR) obtains；(d-4) the marine monocular obtained using the method for the present invention is indicated
Mark the 3D rendering of infrared image.
Being compared by two and three dimensions display figure can see, and testing result generates apparent filtering effect in target area
It answers, it is smaller in smooth background area effect.Therefore, the present invention can effectively enhance mesh while inhibiting background clutter
The contrast of signal strength and target background is marked, target detection precision is promoted.
As shown in table 1, LSCR indicates local signal to noise ratio in table 1, and LSCRG indicates local signal to noise ratio gain, LSCR and LSCRG
Index is objectively evaluated for algorithm performance.Biggish LSCR represents that target is more significant relative to clutter background, and detection effect is more preferable；
Biggish LSCRG indicates that object detection method becomes apparent LSCR promotion, and detection performance is more excellent.Wherein, PRCA is robust
Property principal component analytical method, DSR is intensive sparse reconstruction, and Proposed method indicates method of the invention.By number in table 1
According to can see, LSCR the and LSCRG value of the method for the present invention is maximum, and improves 1-2 quantity relative to other methods
Grade indicates that this method specific image clutter reduction background generated is obvious, and image texture is relatively abundant, and details is prominent, and detection is quasi-
Exactness is high.
Table 1
According to the Selective Attention Mechanism of human vision, the angle analysis differentiated from the vision system multichannel more, infrared mesh
Indicate the feature of different directions.The multi-channel feature for carrying out different frequency and different directions to the infrared image of input extracts, after
Continuous input regression depth residual error network clutter reduction background, inputs classifying type depth residual error network class candidate target, can
Infrared image small target deteection precision is improved, the notable figure that infrared target detection obtains can preferably inhibit background clutter and noise
Interference, object edge are more clear.
It is corresponding based on inventive concept identical with a kind of above-mentioned detection method of infrared image Small object, this
Inventive embodiments additionally provide a kind of detection device of infrared image Small object, as shown in Figure 4.Due to the basic phase of Installation practice
Like and embodiment of the method, so describing fairly simple, the relevent part can refer to the partial explaination of embodiments of method.
A kind of detection device of infrared image Small object provided by the invention, comprising:
Characteristic extracting module 101, for extracting the characteristics of image of source images；
Clutter recognition module 102 is obtained for being handled using regression depth residual error network described image feature
Notable figure after clutter reduction background；
Object extraction module 103, for extracting the candidate target in the notable figure；
Categorization module 104 obtains target for classifying using classifying type depth residual error network to the candidate target
Detect notable figure.
In a specific embodiment provided by the invention, the characteristic extracting module 101 is specifically used for:
Using multi-channel Gabor filtering, the multidimensional image feature of source images is extracted.
In a specific embodiment provided by the invention, the regression depth residual error network is free of pond layer；It is described
Classifying type depth residual error network is free of pond layer.
In a specific embodiment provided by the invention, described device, further includes:
Sample acquisition module, for obtaining sample infrared image；
Sliding window module carries out traversal cutting to the sample infrared image, obtains training image for using sliding window mode
Collection；
Classification annotation module, for concentrating the scene information of each training image according to the training image, to the instruction
Practice image set and carries out classification annotation；
Training module, for according to the training image collection after classification annotation, to regression depth residual error network and classifying type
Depth residual error network carries out pre-training, the network parameter of optimized regression moldeed depth degree residual error network and classifying type depth residual error network.
In a specific embodiment provided by the invention, the object extraction module 103 is specifically used for:
Using threshold processing method, the candidate target in the notable figure is extracted.
In a specific embodiment provided by the invention, the categorization module 104, comprising:
Data extracting unit extracts the neighborhood image number of each candidate target for using four neighborhood clustering methods
According to；
Target classification unit is used for according to the neighborhood image data, using classifying type depth residual error network to the time
It selects object to classify, obtains target detection notable figure.
In a specific embodiment provided by the invention, described device, further includes:
Costing bio disturbance module loses letter for the external mean square error of output layer in the regression depth residual error network
Number；According to the loss function, the quality for predicting the regression depth residual error network is measured.
More than, it is a kind of detection device of infrared image Small object provided by the invention.
It is corresponding based on inventive concept identical with a kind of above-mentioned detection method of infrared image Small object, this
Inventive embodiments additionally provide a kind of computer readable storage medium, are stored thereon with computer program, and the program is by processor
A kind of detection method of above-mentioned infrared image Small object is realized when execution.
As shown from the above technical solution, a kind of computer readable storage medium provided in this embodiment, is stored thereon with meter
Calculation machine program, when which is executed by processor, by detecting candidate in the way of end-to-end by regression depth residual error network
Target executes disposable residual error network operations to the source images of whole picture input, the notable figure after obtaining clutter reduction background, then mentions
The candidate target in notable figure is taken, the input data amount of subsequent classification moldeed depth degree residual error network can be greatly lowered, using classification
Moldeed depth degree residual error network is able to achieve the high-precision identification of candidate target.By by regression depth residual error network and classifying type depth
Residual error network integration gets up to handle source images, can preferably inhibit background clutter and noise jamming, obtains object edge more
Clearly notable figure.
It is corresponding based on inventive concept identical with a kind of above-mentioned detection method of infrared image Small object, this
Inventive embodiments additionally provide a kind of detection device of infrared image Small object, comprising: memory, processor and are stored in storage
On device and the computer program that can run on a processor, the processor realize a kind of above-mentioned infrared figure when executing described program
As the detection method of Small object.
As shown from the above technical solution, the detection device of a kind of infrared image Small object provided in this embodiment, passes through benefit
Candidate target is detected in a manner of end-to-end with regression depth residual error network, and disposable residual error is executed to the source images of whole picture input
Network operations, the notable figure after obtaining clutter reduction background, then the candidate target in notable figure is extracted, it can be greatly lowered subsequent
The input data amount of classifying type depth residual error network is known using the high-precision that classifying type depth residual error network is able to achieve candidate target
Not.It, can be preferably by the way that regression depth residual error network and classifying type depth residual error network integration get up to handle source images
Inhibit background clutter and noise jamming, obtains the notable figure that object edge is more clear.
In specification of the invention, numerous specific details are set forth.It is to be appreciated, however, that the embodiment of the present invention can be with
It practices without these specific details.In some instances, well known method, structure and skill is not been shown in detail
Art, so as not to obscure the understanding of this specification.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example
Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not
It must be directed to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be in office
It can be combined in any suitable manner in one or more embodiment or examples.In addition, without conflicting with each other, the skill of this field
Art personnel can tie the feature of different embodiments or examples described in this specification and different embodiments or examples
It closes and combines.
Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations；To the greatest extent
Pipe present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: its according to
So be possible to modify the technical solutions described in the foregoing embodiments, or to some or all of the technical features into
Row equivalent replacement；And these are modified or replaceed, various embodiments of the present invention technology that it does not separate the essence of the corresponding technical solution
The range of scheme should all cover within the scope of the claims and the description of the invention.
Claims (10)
1. a kind of detection method of infrared image Small object characterized by comprising
Extract the characteristics of image of source images；
Residual error network operations are carried out to described image feature using regression depth residual error network, after obtaining clutter reduction background
Notable figure；
Extract the candidate target in the notable figure；
Classified using classifying type depth residual error network to the candidate target, obtains target detection notable figure.
2. the method according to claim 1, wherein the characteristics of image for extracting source images, comprising:
Using multi-channel Gabor filtering, the multidimensional image feature of source images is extracted.
3. the method according to claim 1, wherein the regression depth residual error network is free of pond layer；Institute
Classifying type depth residual error network is stated without pond layer.
4. the method according to claim 1, wherein utilizing regression depth residual error network to the figure described
Before the step of carrying out residual error network operations as feature, obtain the notable figure of clutter reduction background, further includes:
Obtain sample infrared image；
Using sliding window mode, traversal cutting is carried out to the sample infrared image, obtains training image collection；
The scene information that each training image is concentrated according to the training image carries out classification annotation to the training image collection；
According to the training image collection after classification annotation, regression depth residual error network and classifying type depth residual error network are carried out pre-
Training, the network parameter of optimized regression moldeed depth degree residual error network and classifying type depth residual error network.
5. the method according to claim 1, wherein the candidate target extracted in the notable figure, comprising:
Using threshold processing method, the candidate target in the notable figure is extracted.
6. the method according to claim 1, wherein described use classifying type depth residual error network to the candidate
Object is classified, and target detection notable figure is obtained, comprising:
Using four neighborhood clustering methods, the neighborhood image data of each candidate target are extracted；
According to the neighborhood image data, classified using classifying type depth residual error network to the candidate target, obtains mesh
Mark detection notable figure.
7. the method according to claim 1, wherein the output layer in the regression depth residual error network is external
Mean square error loss function；
According to the loss function, the quality for predicting the regression depth residual error network is measured.
8. a kind of detection device of infrared image Small object characterized by comprising
Characteristic extracting module, for extracting the characteristics of image of source images；
Clutter recognition module is obtained for being handled using regression depth residual error network described image feature and inhibits miscellaneous
Notable figure after wave background；
Object extraction module, for extracting the candidate target in the notable figure；
It is aobvious to obtain target detection for classifying using classifying type depth residual error network to the candidate target for categorization module
Write figure.
9. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is held by processor
Method described in one of claim 1-7 is realized when row.
10. a kind of detection device of infrared image Small object, comprising: memory, processor and storage are on a memory and can be
The computer program run on processor, which is characterized in that the processor realized when executing described program claim 1-7 it
Method described in one.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811166867.9A CN109389593A (en) | 2018-09-30 | 2018-09-30 | A kind of detection method, device, medium and the equipment of infrared image Small object |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811166867.9A CN109389593A (en) | 2018-09-30 | 2018-09-30 | A kind of detection method, device, medium and the equipment of infrared image Small object |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109389593A true CN109389593A (en) | 2019-02-26 |
Family
ID=65426647
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811166867.9A Pending CN109389593A (en) | 2018-09-30 | 2018-09-30 | A kind of detection method, device, medium and the equipment of infrared image Small object |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109389593A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110232378A (en) * | 2019-05-30 | 2019-09-13 | 苏宁易购集团股份有限公司 | A kind of image interest point detecting method, system and readable storage medium storing program for executing |
| CN110263638A (en) * | 2019-05-16 | 2019-09-20 | 山东大学 | A kind of video classification methods based on significant information |
| CN110910421A (en) * | 2019-11-11 | 2020-03-24 | 西北工业大学 | Weak and small moving object detection method based on block characterization and variable neighborhood clustering |
| CN112288778A (en) * | 2020-10-29 | 2021-01-29 | 电子科技大学 | Infrared small target detection method based on multi-frame regression depth network |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718889A (en) * | 2016-01-21 | 2016-06-29 | 江南大学 | Human face identity recognition method based on GB(2D)2PCANet depth convolution model |
| CN106446937A (en) * | 2016-09-08 | 2017-02-22 | 天津大学 | Multi-convolution identifying system for AER image sensor |
| CN107392183A (en) * | 2017-08-22 | 2017-11-24 | 深圳Tcl新技术有限公司 | Face classification recognition methods, device and readable storage medium storing program for executing |
| CN107909027A (en) * | 2017-11-14 | 2018-04-13 | 电子科技大学 | It is a kind of that there is the quick human body target detection method for blocking processing |
| CN108133218A (en) * | 2017-12-14 | 2018-06-08 | 内蒙古科技大学 | Infrared target detection method, equipment and medium |
| CN108229319A (en) * | 2017-11-29 | 2018-06-29 | 南京大学 | The ship video detecting method merged based on frame difference with convolutional neural networks |
| CN108288271A (en) * | 2018-02-06 | 2018-07-17 | 上海交通大学 | Image detecting system and method based on three-dimensional residual error network |
-
2018
- 2018-09-30 CN CN201811166867.9A patent/CN109389593A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718889A (en) * | 2016-01-21 | 2016-06-29 | 江南大学 | Human face identity recognition method based on GB(2D)2PCANet depth convolution model |
| CN106446937A (en) * | 2016-09-08 | 2017-02-22 | 天津大学 | Multi-convolution identifying system for AER image sensor |
| CN107392183A (en) * | 2017-08-22 | 2017-11-24 | 深圳Tcl新技术有限公司 | Face classification recognition methods, device and readable storage medium storing program for executing |
| CN107909027A (en) * | 2017-11-14 | 2018-04-13 | 电子科技大学 | It is a kind of that there is the quick human body target detection method for blocking processing |
| CN108229319A (en) * | 2017-11-29 | 2018-06-29 | 南京大学 | The ship video detecting method merged based on frame difference with convolutional neural networks |
| CN108133218A (en) * | 2017-12-14 | 2018-06-08 | 内蒙古科技大学 | Infrared target detection method, equipment and medium |
| CN108288271A (en) * | 2018-02-06 | 2018-07-17 | 上海交通大学 | Image detecting system and method based on three-dimensional residual error network |
Non-Patent Citations (2)
| Title |
|---|
| 林两魁 等: "基于深度卷积神经网络的红外过采样扫描图像点目标检测方法", 《红外与毫米波学报》 * |
| 王丽娜 等: "《信息隐藏技术与应用》", 31 December 2012, 武汉大学出版社 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110263638A (en) * | 2019-05-16 | 2019-09-20 | 山东大学 | A kind of video classification methods based on significant information |
| CN110232378A (en) * | 2019-05-30 | 2019-09-13 | 苏宁易购集团股份有限公司 | A kind of image interest point detecting method, system and readable storage medium storing program for executing |
| CN110910421A (en) * | 2019-11-11 | 2020-03-24 | 西北工业大学 | Weak and small moving object detection method based on block characterization and variable neighborhood clustering |
| CN112288778A (en) * | 2020-10-29 | 2021-01-29 | 电子科技大学 | Infrared small target detection method based on multi-frame regression depth network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109583369B (en) | Target identification method and device based on target area segmentation network | |
| CN109389593A (en) | A kind of detection method, device, medium and the equipment of infrared image Small object | |
| CN112818903A (en) | Small sample remote sensing image target detection method based on meta-learning and cooperative attention | |
| CN108021890B (en) | High-resolution remote sensing image port detection method based on PLSA and BOW | |
| Zhang et al. | Self-attention guidance and multiscale feature fusion-based UAV image object detection | |
| CN116168240A (en) | Arbitrary-direction dense ship target detection method based on attention enhancement | |
| Yang et al. | A super lightweight and efficient SAR image ship detector | |
| Kong et al. | Lightweight algorithm for multi-scale ship detection based on high-resolution SAR images | |
| Yasir et al. | ShipGeoNet: SAR image-based geometric feature extraction of ships using convolutional neural networks | |
| Liu et al. | A multi-scale feature pyramid SAR ship detection network with robust background interference | |
| Li et al. | A computational model of vision attention for inspection of surface quality in production line | |
| CN117079097A (en) | Sea surface target identification method based on visual saliency | |
| CN103455798A (en) | Human detection method based on maximum geometric flow direction column diagram | |
| Yu et al. | A deep learning-based SAR ship detection | |
| CN109977892A (en) | Ship Detection based on local significant characteristics and CNN-SVM | |
| CN114005045A (en) | Rotating frame remote sensing target detection method based on lightweight deep neural network | |
| Dong et al. | Accurate and real-time visual detection algorithm for environmental perception of USVS under all-weather conditions | |
| Liang et al. | Cross-layer triple-branch parallel fusion network for small object detection in uav images | |
| Wu et al. | An industrial meter detection method based on lightweight YOLOX-CAlite | |
| Fu et al. | Ship detection based on Improved YOLO Algorithm | |
| Pan et al. | The Crime Scene Tools Identification Algorithm Based on GVF-Harris-SIFT and KNN. | |
| Zheng et al. | Segmentation Method for Whole Vehicle Wood Detection Based on Improved YOLACT Instance Segmentation Model | |
| Dong et al. | A Contrastive Learning Transformer for Change Detection in High-Resolution SAR Images | |
| Chen et al. | Improved Faster R-CNN identification method for containers | |
| Chen et al. | Real-Time Ship Detection Algorithm Based on Improved YOLOv8 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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190226 |