CN105469428B - A kind of detection method of small target based on morphologic filtering and SVD - Google Patents
A kind of detection method of small target based on morphologic filtering and SVD Download PDFInfo
- Publication number
- CN105469428B CN105469428B CN201510837688.3A CN201510837688A CN105469428B CN 105469428 B CN105469428 B CN 105469428B CN 201510837688 A CN201510837688 A CN 201510837688A CN 105469428 B CN105469428 B CN 105469428B
- Authority
- CN
- China
- Prior art keywords
- image
- max
- image sequence
- svd
- morphologic filtering
- 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.)
- Expired - Fee Related
Links
Classifications
-
- 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
- G06T2207/10016—Video; Image sequence
-
- 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
- G06T2207/20036—Morphological image processing
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30212—Military
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of detection method of small target based on morphologic filtering and SVD, step 1:Background suppression, noise remove, image sequence after being pre-processed are carried out by morphologic filtering target enhancing algorithm;Step 2:Read in NmaxImage sequence obtained by width, carries out frame number estimation, obtains frame number N to be treated;Step 3:The image that N+1 width image forms is merged into 2-D data, its autocorrelation matrix is sought and SVD is carried out to its autocorrelation matrix;Step 4:Suitable feature vector reconstructed image sequence is selected, obtains new characteristic image sequence;Step 5:To reconstructed image sequence into row threshold division, from background in isolated original image Weak target position;Step 6:Each image in sequence is modified respectively;Step 7:N is replaced into NmaxAfterwards, repeat steps 2 through 7.The method of morphologic filtering and singular value decomposition is effectively combined by the present invention is detected the Weak target in video, and the calculating time is short, and detection efficiency is high, and accuracy and robustness are all relatively good.
Description
Technical field
This method belongs to video analysis field, and in particular to a kind of Dim targets detection based on morphologic filtering and SVD
Method.
Background technology
Status of the detection of Weak target in modern war is self-evident, becomes satellite remote sensing, high energy at present
The core technology of the realm informations such as physics, low latitude early warning and precise guidance processing.Since the pixel number of Weak target is seldom,
Lack the structural information of target, it is seldom for splitting the information utilized with detection algorithm.And sensor receive target strength compared with
Weak, noise and background clutter interference are stronger, make the signal-to-noise ratio of image reduce, so we should make good use of target in sequence image
Continuity and systematicness detect target.All the time, how Weak target inter-frame information is preferably utilized, improves detection
Reliability and efficiency, are the emphasis of Dim targets detection.
Suppress mainly to concentrate on Dynamic Programming and shape with Dim targets detection problem, a few thing of early stage for background
State estimation technique increases the detectability of target, still, poor performance may be presented in the case of low signal-to-noise ratio.Mesh
Before, someone will be used for Dim targets detection the methods of morphologic filtering, genetic algorithm, neural network algorithm, wavelet transformation.
But under complex background, target point is easily flooded by noise, realize that the reliability detection of target and identification difficulty are larger.This
Outside, it is difficult to meet good detection performance under conditions of data throughout is big, requirement of real-time is high.
The content of the invention
Goal of the invention:In view of the problems of the existing technology, the present invention provides one kind calculate the time it is short, detection accurately
Detection method of small target based on morphologic filtering and singular value decomposition (hereinafter referred SVD).
The content of the invention:The present invention provides a kind of detection method of small target based on morphologic filtering and SVD, including with
Lower step:
Step 1:Video sequence to be detected is inputted, background suppression is carried out by morphologic filtering target enhancing algorithm, is made an uproar
Sound removes, image sequence after being pre-processed;
Step 2:From pretreated image sequence, N is read inmaxThe image sequence of width image composition, carries out frame number and estimates
Meter, obtains frame number N to be treated;
Step 3:Read in N+1 width images, including the latter picture of N two field pictures to be treated and N two field pictures, and by N+
The image of 1 width image composition is merged into 2-D data, seeks its autocorrelation matrix and carries out SVD to its autocorrelation matrix;
Step 4:Intermediate characteristic value character pair vector reconstruction image sequence is selected, obtains reconstructed image sequence;
Step 5:To reconstructed image sequence into row threshold division, from background in isolated original image Weak target position
Put;
Step 6:Location between frames amendment and frame are carried out respectively to each image in the reconstructed image sequence that is obtained in step 4
Interior position correction;
Step 7:N is replaced into NmaxAfterwards, repeat steps 2 through 7, until NmaxLast width in the image sequence of width image composition
Result is exported after the completion of image procossing.
Further, morphologic filtering target enhancing algorithm is used as structural element using circle in the step 1.This is for extensive
Multiple image polluted by noise can produce preferable filter effect.More preferable filter effect can so be obtained.
Further, morphologic filtering target enhancing algorithm is utilized in the step 1 first to each frame in video to be detected
Image carries out closed operation, then to have inside the image that can insert between sand holes noise or will not be formed the region of degeneration rectangle into
Row opening operation.
Further, the method for frame number estimation is in the step 2:Read in NmaxThe image sequence of width image composition, to reading in
All images carry out two-by-two difference operation obtain Nmax- 1 frame image sequence, then to NmaxEach width in -1 frame image sequence
The summation of image total pixel, then with the N tried to achievemax- 1 pixel and respectively divided by correspondence image width and height obtain Nmax- 1 assessment
Value, by Nmax- 1 assessed value is averaging to NmaxTwo field picture changes the value X of severe degree, finally by image change severe degree X
Substitute into formulaIn, it is to be treated to obtain this
Number of image frames N, wherein, NmaxRepresent maximum processing frame number, NminRepresent minimum treat frame number.
Further, the maximum processing frame number Nmax25 are set to, minimum treat frame number NminIt is set to 5.
Further, in the step 5 using maximum entropy method to image into row threshold division.
Further, modification method is in the step 6:Gray scale maximum is found in current search window, then by ash
Center iterative search of the coordinate of maximum as next search window is spent, untill final search window is constant, this
When search window center position be target location.
Beneficial effect:Compared with prior art, the present invention effectively ties the method for morphologic filtering and singular value decomposition
Conjunction is detected the Weak target in video, and not only the calculating time is short, and detection efficiency is high, and accuracy and robustness all compare
Preferably.
Brief description of the drawings
Fig. 1 is the work flow diagram of the present invention.
Fig. 2 (a) and Fig. 2 (b) is respectively to detect the 20th frame and the 50th two field picture in video;
Fig. 3 (a) and Fig. 3 (b) is respectively to the 20th frame and the 50th two field picture in detection video based on morphologic filtering method
The result of middle Dim targets detection;
Fig. 4 (a) and Fig. 4 (b) is respectively that Weak target in the 20th frame in detection video and the 50th two field picture is examined based on SVD
The result of survey;
Fig. 5 (a) and Fig. 5 (b) is respectively that the present invention examines Weak target in the 20th frame in detection video and the 50th two field picture
The result of survey.
Embodiment
Below in conjunction with the accompanying drawings, the present invention is described in detail.
As shown in Figure 1, the present invention provides a kind of detection method of small target based on morphologic filtering and SVD, including
Following steps:
Step 1:Video sequence to be detected is inputted, background suppression is carried out by morphologic filtering target enhancing algorithm, is made an uproar
Sound removes, image sequence after being pre-processed;
In view of the detection of Weak target, it can more or less run into noise image and overlap to form knot with non-noise image
The radius of group or some noise particles has exceeded the radius of non-noise particle, then the circle of correct radial may be selected in such case
For shape as structural element, this can produce preferable filter effect for recovering image polluted by noise, be because:(1) circular circle
Change acts on the effect that can play low-pass filtering;(2) rotating influence may not necessarily be considered using circular filter.
When determining the radius of circular configuration element, optimization method can be used, image and noise are considered as random process, led to
Statistical analysis is crossed, to carrying out quantitative analysis by the Granule Images of noise pollution, statistical distribution parameter is asked for, obtains probability of occurrence most
Big noise particles and not by the radius of noise pollution particle, it is not structural element half by the radius of noise pollution particle to choose
Footpath, obtains optimum results.
In addition, may be both there are pepper shape noise (poor noise) in the sequence of Weak target, there is also sand holes noise (and to make an uproar
Sound), it is so simple using opening operation or closed operation effect all without good, for compatible smooth noise and retain image border
And other significant features, the present invention first carry out closed operation to each two field picture in video to be detected, then to can be into
The region opening operation of row opening operation, can obtain more preferable filter effect, wherein the region that can carry out opening operation is to have and can fill out
Enter inside the image between sand holes noise, the region of serious degeneration rectangle will not be formed.According to formula and definition, we are not rare
Know, which has good translation invariance, incremental, idempotence and duality.These properties cause such a
Wave filter practicality is closed-opened to open-close and feasibility is stronger, and the effect of acquirement is also more preferably.
Step 2:From pretreated image sequence, N is read inmaxThe image sequence of width image composition, carries out frame number and estimates
Meter, obtains frame number N to be treated;
In Weak target sequence, different target needs the frame number disposably handled and differs, so needing to pass through certainly
Adaptation is determined to obtain frame number N to be treated.When frame number is selected, first have to consider is image change speed, i.e.,
Less frame number is selected to handle when changing fast, on the contrary the more frame number of selection.So it needs to be determined that maximum when frame number is estimated
Manage frame number NmaxWith minimum treat frame number Nmin, wherein, maximum processing frame number NmaxAs read in from pretreated image sequence
Picture number Nmax.If NmaxVery big then algorithm does not have real-time, equally, if Nmin≤ 0, then carrying out singular value decomposition to it is also
Not in all senses.In view of real-time and feasibility, maximum processing frame number Nmax25 are set to, minimum treat frame number NminIt is set to
5, calculate N with cosine functionmaxAnd NminBetween frame number estimation, independent variable is image change severe degree X.
Handled, handled as X > 1.5 using maximum frame number, as 0.5 < X < 1.5 using minimum frame number as X < 0.5
Handled using cosine function transient mode.Using trigonometric function knowledge, obtain image change severe degree X and handle frame number N's
Formula is as follows:
Wherein, specific assessment algorithm is:In Nmax=25 and NminOn the premise of=5, the image of maximum frame number is read in first,
Difference operation obtains N two-by-two for image progress to these framesmax- 1 frame image sequence, then to NmaxIt is every in -1 frame image sequence
The summation of piece image total pixel, then with the N tried to achievemax- 1 pixel and respectively divided by correspondence image width and height obtain Nmax- 1
Assessed value, by Nmax- 1 assessed value, which is averaging, can be obtained by NmaxTwo field picture changes the value X of severe degree, finally substitutes into X
Formula (1) obtains this number of image frames N to be treated.
Step 3:Read in N+1 width images, including the latter picture of N two field pictures to be treated and N two field pictures, and by N+
The image of 1 width image composition is merged into 2-D data, seeks its autocorrelation matrix and carries out SVD to its autocorrelation matrix;
The image sequence A of N+1 width image composition is the three-dimensional data set of one { x, y, z }, and wherein x represents the row of image
Number, y represent the columns of image, and z represents the frame number of sequence.If matrix Bq×zFor the two-dimensional development of A, wherein q=x × y, then B is x
The matrix of × y rows z row.Under normal circumstances, the value of x × y is bigger, so matrix Bq×zData volume be not suitable for directly using
To carry out singular value decomposition.If Bq×zAutocorrelation matrix be Cq×z, then Cq×z=B 'q×zBq×z, wherein, B 'q×zIt is Bq×zTurn
Put, so Matrix Cq×zIt is the square formation that row and column is equal to z.At this time, to Matrix Cq×zCarry out SVD.In order to realize to Weak target
Detection, image sequence to every piece image and its before does above-mentioned singular value decomposition.
Step 4:Intermediate characteristic value character pair vector reconstruction image sequence is selected, obtains reconstructed image sequence;
According to singular value decomposition as a result, selecting suitable feature vector to carry out corresponding image sequence reconstruct, after reconstruct
Image sequenceBe q rows i row matrix, Vz×iIt is that intermediate characteristic value corresponds to spy
Sign vector, can incite somebody to actionRemap as image arrayR represents the sum of the singular value obtained in step 3, and i represents strange
The label of different value, selects different i to obtain different reconstructed images.Wherein, background image sequence accounts for whole image to be detected
The overwhelming majority of sequence, the image of the obtained corresponding feature vector reconstruct of main characteristic value will reflect the information of background.
Noise image sequence is the random process of one group of two dimension, the image of the corresponding feature vector reconstruct of obtained minimum characteristic value because
And it can reflect noise information.Target image sequence is continuous motion change in whole image sequence, obtained intermediate features
The image of value reconstruct will be strengthened mobile Weak target.
Step 5:To the reconstructed image sequence that is obtained in step 4 into row threshold division, the isolated original image from background
The position of middle Weak target;
Wherein, use maximum entropy method to image into row threshold division in step 5, mainly using in reconstructed image sequence
The intensity profile density function of each image defines the comentropy of image, and difference is proposed according to the difference of hypothesis or the different of visual angle
Entropy criterion, finally by optimize the criterion obtain threshold value.The comentropy of image has reacted the overall looks of image.If in image
It is comprising target, then maximum in target and the divisible place's information content (i.e. entropy) of background.Contain the multi-grey image of multiple target in a width
In, a gray scale is certainly existed, using this gray scale as threshold value, image can be made to obtain optimal binarization segmentation.So as to fulfill mesh
The separation of mark and background, obtains its approximate location.
Step 6:Location between frames amendment and frame are carried out respectively to each image in the reconstructed image sequence that is obtained in step 4
Interior position correction, to achieve the purpose that to detect Weak target;
In order to improve the accuracy of target location estimation, a two field picture and its need during image sequence to use this under treatment
Secondary testing result is modified target location.
Specific method is the difference of different twice testing results before and after, if in certain threshold value, updates mesh
Test position is marked, otherwise it is assumed that this processing is exceptional value, testing result is given up.Threshold value herein passes through maximum entropy method
Obtain.
Singular value decomposition is image sequence, therefore not only includes the data of present image in whole data space, together
When also include before N width view data.The target location obtained after reconstruct can reflect that current goal is gone over to a certain extent
Position, rather than in present image target position, so also need in frame carry out target location amendment.
The basic principle of correction algorithm is:Gray scale maximum is found in current search window, then coordinate will be made herein
For the center iterative search of next search window, untill final search window is constant, at this moment search window center
Position is target location.
Combining interframe after SVD processing can remain the Dim targets detection rate of image sequence with the joint-detection in frame
90% or so, this largely meets the real-time and feasibility requirement of image procossing.
Step 7:N is replaced into NmaxAfterwards, repeat steps 2 through 7, until NmaxLast width in the image sequence of width image composition
Result is exported after the completion of image procossing.
As shown in Fig. 2 (a) and Fig. 2 (b), to the infrared weak of actual photographed under vs2010+opencv2.4.3 running environment
Small object sequence image is tested, and totally 100 frame, image size are 200 × 256 pixels to sequence image, and target sizes are about 2
× 2 pixels, signal-to-noise ratio are contrast 7% within 2, and background be sky and cloud layer the 20th two field pictures of selection (for Fig. 2 (a)) and the
50 frames (Fig. 2 (b)) are based respectively on morphologic filtering, are detected based on SVD and method provided by the invention.
As shown in Fig. 3 (a)~Fig. 5 (b), morphologic filtering method processing speed is very fast, but can not accurately reach weak
The purpose of small target deteection, also detects decoy;SVD can preferably realize prominent Weak target, and can keep higher inspection
Survey rate, but the tracking to target location is not too preferable, and in addition detection calculations amount is big, and speed is slower.And side proposed by the present invention
Not only verification and measurement ratio is high for method, but also the speed handled is fast, and robustness and real-time are all relatively good.
Claims (6)
- A kind of 1. detection method of small target based on morphologic filtering and SVD, it is characterised in that:Comprise the following steps:Step 1:Video sequence to be detected is inputted, background suppression is carried out by morphologic filtering target enhancing algorithm, noise is gone Remove, image sequence after being pre-processed;Step 2:From pretreated image sequence, N is read inmaxThe image sequence of width image composition, carries out frame number estimation, obtains To frame number N to be treated;Step 3:Read in N+1 width images, including the latter picture of N two field pictures to be treated and N two field pictures, and by N+1 width The image of image composition is merged into 2-D data, seeks its autocorrelation matrix and carries out SVD to its autocorrelation matrix;Step 4:Intermediate characteristic value character pair vector reconstruction image sequence is selected, obtains reconstructed image sequence;Step 5:To reconstructed image sequence into row threshold division, from background in isolated original image Weak target position;Step 6:Location between frames amendment and position in frame are carried out respectively to each image in the reconstructed image sequence that is obtained in step 4 Put amendment;Step 7:N is replaced into NmaxAfterwards, repeat steps 2 through 7, until NmaxLast piece image in the image sequence of width image composition Result is exported after the completion of processing;The method of frame number estimation is in the step 2:Read in NmaxThe image sequence of width image composition, to all images of reading Carry out difference operation two-by-two and obtain Nmax- 1 frame image sequence, then to NmaxThe total pixel of every piece image in -1 frame image sequence Summation, then with the N tried to achievemax- 1 pixel and respectively divided by correspondence image width and height obtain Nmax- 1 assessed value, by Nmax-1 A assessed value is averaging to NmaxTwo field picture changes the value X of severe degree, and image change severe degree X finally is substituted into formulaIn, obtain this number of image frames to be treated N, wherein, NmaxRepresent maximum processing frame number, NminRepresent minimum treat frame number.
- 2. the detection method of small target according to claim 1 based on morphologic filtering and SVD, it is characterised in that:Institute State morphologic filtering target enhancing algorithm in step 1 and structural element is used as using circle.
- 3. the detection method of small target according to claim 1 based on morphologic filtering and SVD, it is characterised in that:Institute State in step 1 and closed operation first is carried out to each two field picture in video to be detected using morphologic filtering target enhancing algorithm, so Opening operation is carried out to the region that has inside the image that can be inserted between sand holes noise or will not be formed degeneration rectangle afterwards.
- 4. the detection method of small target according to claim 1 based on morphologic filtering and SVD, it is characterised in that:Institute State maximum processing frame number Nmax25 are set to, minimum treat frame number NminIt is set to 5.
- 5. the detection method of small target according to claim 1 based on morphologic filtering and SVD, it is characterised in that:Institute State in step 5 using maximum entropy method to image into row threshold division.
- 6. the detection method of small target according to claim 1 based on morphologic filtering and SVD, it is characterised in that:Institute Stating modification method in step 6 is:In current search window find gray scale maximum, then using the coordinate of gray scale maximum as The center iterative search of next search window, untill final search window is constant, the at this moment position at search window center Put as target location.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510837688.3A CN105469428B (en) | 2015-11-26 | 2015-11-26 | A kind of detection method of small target based on morphologic filtering and SVD |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510837688.3A CN105469428B (en) | 2015-11-26 | 2015-11-26 | A kind of detection method of small target based on morphologic filtering and SVD |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105469428A CN105469428A (en) | 2016-04-06 |
CN105469428B true CN105469428B (en) | 2018-05-11 |
Family
ID=55607087
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510837688.3A Expired - Fee Related CN105469428B (en) | 2015-11-26 | 2015-11-26 | A kind of detection method of small target based on morphologic filtering and SVD |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105469428B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106570889A (en) * | 2016-11-10 | 2017-04-19 | 河海大学 | Detecting method for weak target in infrared video |
CN108010064A (en) * | 2017-11-03 | 2018-05-08 | 河海大学 | Motor cell tracking based on active profile and Kalman filter |
CN111047624A (en) * | 2019-12-27 | 2020-04-21 | 成都英飞睿技术有限公司 | Image dim target detection method, device, equipment and storage medium |
CN112381738B (en) * | 2020-11-23 | 2023-12-05 | 深圳企业云科技股份有限公司 | Perspective image self-adaptive correction algorithm based on correlation operation |
CN112686890B (en) * | 2021-02-09 | 2024-02-20 | 浙江师范大学 | Artificial board surface defect detection method based on singular value decomposition |
CN113971815B (en) * | 2021-10-28 | 2024-07-02 | 西安电子科技大学 | Few-sample target detection method based on singular value decomposition feature enhancement |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102750533A (en) * | 2012-07-05 | 2012-10-24 | 重庆大学 | Infrared small and weak target detection method based on morphological component sparse representation |
CN103440502A (en) * | 2013-09-06 | 2013-12-11 | 重庆大学 | Infrared small-target detection method based on mixing Gauss and sparse representation |
-
2015
- 2015-11-26 CN CN201510837688.3A patent/CN105469428B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102750533A (en) * | 2012-07-05 | 2012-10-24 | 重庆大学 | Infrared small and weak target detection method based on morphological component sparse representation |
CN103440502A (en) * | 2013-09-06 | 2013-12-11 | 重庆大学 | Infrared small-target detection method based on mixing Gauss and sparse representation |
Non-Patent Citations (3)
Title |
---|
Infrared dim small target detection algorithm based on NSCT and SVD;Ying Zhao et al;《AOPC 2015: Image Processing and Analysis》;20151014;第9675卷;全文 * |
基于SVD背景抑制和粒子滤波的弱小目标检测;崔丽洁 等;《计算机应用研究》;20110430;第28卷(第4期);全文 * |
基于背景抑制和灰度形态学的弱小目标检测;董宇星 等;《图像图形技术研究与应用》;20101231;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN105469428A (en) | 2016-04-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105469428B (en) | A kind of detection method of small target based on morphologic filtering and SVD | |
CN109934200B (en) | RGB color remote sensing image cloud detection method and system based on improved M-Net | |
Gao et al. | Robust infrared small target detection using multiscale gray and variance difference measures | |
CN110119728A (en) | Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network | |
Chen et al. | MFFN: An underwater sensing scene image enhancement method based on multiscale feature fusion network | |
Li et al. | Road lane detection with gabor filters | |
Gao et al. | Infrared small target detection using a temporal variance and spatial patch contrast filter | |
CN113591968A (en) | Infrared weak and small target detection method based on asymmetric attention feature fusion | |
CN104504652A (en) | Image denoising method capable of quickly and effectively retaining edge and directional characteristics | |
CN107507209B (en) | Printogram extraction method of polarized SAR image | |
CN106469313A (en) | Weak and small target detection method for pipe diameter self-adaptive time-space domain filtering | |
Nair et al. | Fuzzy logic-based automatic contrast enhancement of satellite images of ocean | |
Tan et al. | A CNN-based self-supervised synthetic aperture radar image denoising approach | |
CN107944354A (en) | A kind of vehicle checking method based on deep learning | |
CN111797841A (en) | Visual saliency detection method based on depth residual error network | |
Zhao et al. | An adaptation of CNN for small target detection in the infrared | |
CN106845343A (en) | A kind of remote sensing image offshore platform automatic testing method | |
Zhang et al. | Infrared small target detection based on gradient correlation filtering and contrast measurement | |
Xu et al. | COCO-Net: A dual-supervised network with unified ROI-loss for low-resolution ship detection from optical satellite image sequences | |
Donate et al. | Improved reconstruction of images distorted by water waves | |
CN106570889A (en) | Detecting method for weak target in infrared video | |
Xu et al. | Extended non-local feature for visual saliency detection in low contrast images | |
Liu et al. | A multi-scale feature pyramid SAR ship detection network with robust background interference | |
CN107748885B (en) | Method for recognizing fuzzy character | |
CN107529647B (en) | Cloud picture cloud amount calculation method based on multilayer unsupervised sparse learning network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180511 Termination date: 20201126 |