CN109166106A - A kind of target detection aligning method and apparatus based on sliding window - Google Patents
A kind of target detection aligning method and apparatus based on sliding window Download PDFInfo
- Publication number
- CN109166106A CN109166106A CN201810871600.3A CN201810871600A CN109166106A CN 109166106 A CN109166106 A CN 109166106A CN 201810871600 A CN201810871600 A CN 201810871600A CN 109166106 A CN109166106 A CN 109166106A
- Authority
- CN
- China
- Prior art keywords
- sliding window
- target
- confidence level
- region
- candidate target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The target detection aligning method and apparatus based on sliding window that the invention discloses a kind of, are arranged the width and Jump step of sliding window, the image of target to be detected are divided using sliding window, obtain several candidate target regions；All candidate target regions feeding CNN neural network is trained processing, obtains the confidence level of all candidate target regions；Selection confidence level maximum value is worth on the basis of index region corresponding with the maximum value；Candidate target region is cut and combined with a reference value using aligning method, forms new target area.The present invention gives the localization method that can be combined, cut based on convolutional neural networks and sliding window for simple target in image, improves the accuracy and speed of target identification.
Description
Technical field
The present invention relates to field of image processings, and in particular to a kind of target detection aligning method based on sliding window
And device.
Background technique
It is well known that being currently the information age, acquisition, processing, processing and the application of information, which have, develops by leaps and bounds.People
To recognize the important Knowledge Source in the world be exactly image information, in many occasions, the information that image is transmitted is than other forms
Information is richer, vivid and specific.Human eye and brain cooperate so that people are available, handle and understand visual information,
The efficiency of human use's visual perception external environment information is very high.In fact, according to the statistics that some foreign scholars are done, Ren Leisuo
Obtaining external information to have 80% or so is the image from eyes intake.It can be seen that vision obtains external information as the mankind
Main carriers, computer will realize intelligence, just allow for processing image information.Especially in recent years, with figure, figure
The image real time transfer that the large capacities such as picture, video are characterized is widely used in the fields such as medicine, traffic, industrial automation.
In recent years, machine learning is by the extensive concern in academic and engineering.In machine learning, convolutional neural networks
(Convolutional Neural Network, CNN) is a kind of depth feed forward-fuzzy control, generally includes convolutional layer
(convolution layer), layer (normalization layer), pond layer (pooling layer) are normalized and is connected entirely
Layer (full-connected layer) is met, image recognition has been applied successfully to.Now, CNN has become numerous science necks
One of the research hotspot in domain, especially in pattern classification field, since the network avoids the pretreatment complicated early period to image,
Original image can be directly inputted and the operation such as classify is carried out to image, thus obtained more being widely applied.
Target detection is the important content of image procossing, field of target recognition, and main task is from a width given image
Middle positioning target and classification, wherein the method based on sliding window search is used widely in target detection.But traditional sliding window
(sliding window) there are disadvantages to have for search technique: (1) window size is fixed, and the size of segmented image will not be because of target
Size and change；(2) if there is multiple groups sliding window of different sizes to work at the same time, it will definitely increase calculation amount, influence efficiency；(3)
When sliding stride is intensive, data volume increases, and influences speed；When sliding stride is excessive, Detection accuracy is influenced.
In conclusion in the prior art for the accuracy rate of target detection and low efficiency the problem of, still lack effective solution
Certainly scheme.
Summary of the invention
In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a kind of target detection position based on sliding window
Antidote and device, for simple target in image, based on convolutional neural networks and sliding window, giving can group
The localization method for closing, cutting, improves the accuracy and speed of target identification.
The technical scheme adopted by the invention is that:
The first object of the present invention is to provide a kind of target detection aligning method based on sliding window, this method packet
Include following steps:
The width and Jump step of sliding window are set, the image of target to be detected is divided using sliding window, if obtaining
Dry candidate target region；
All candidate target regions feeding CNN neural network is trained processing, obtains all candidate target regions
Confidence level；
Selection confidence level maximum value is worth on the basis of index region corresponding with the maximum value；
Candidate target region is cut and combined with a reference value using aligning method, forms new target area
Domain.
Further, the width of sliding window is determined according to the mean size of all examined objects；The shifting of sliding window
Dynamic split is less than or equal to the half of sliding window width.
Further, described the step of all candidate target regions feeding CNN neural network is trained processing, includes:
It, will be with target area correlation ratio using the candidate target region with target area correlation ratio less than threshold value I as noise
Candidate target region greater than threshold value I is separately input to training in CNN neural network as target；
The confidence level of all candidate target regions is obtained using trained CNN neural network.
Further, it when noise region is excessive, using the multiple noise regions of arbitrary sampling method random erasure, or deletes
Except the picture of corresponding training set.
Further, in all confidence levels of CNN neural network output, confidence level maximum value is chosen, by the confidence level
Maximum value is worth on the basis of index region corresponding with the confidence level maximum value.
Further, according to the size of the width of sliding window and target to be detected, using the depth of breadth traversal as time
Constraint condition is gone through, carries out aligning when the depth capacity of breadth traversal is less than or equal to 2.
Further, described that cutting and combined method are carried out to candidate target region using aligning method and a reference value
Are as follows:
Region is put centered on maximum confidence manipulative indexing region；
Setting area power threshold value T1, confidence level activation threshold T2Inhibit threshold value T with confidence level3；
Using center point area as origin, four adjacent areas are current candidate diffusion zone up and down；
Based on breadth traversal algorithm, current diffusion zone confidence level and index region central point maximum confidence are done
Difference, and inhibit threshold value to be compared respectively diffusion zone confidence level and confidence level activation threshold, confidence level；
If the difference of current some direction confidence level of diffusion zone and maximum confidence is less than T1, and diffusion zone confidence level
Greater than T2, then central area is expanded into boundary to the corresponding direction of the diffusion zone；
If current some direction confidence level of diffusion zone is less than T3, then illustrate that target area does not extend in the direction, to
Target is detected in the corresponding index region of maximum value, central area corresponding direction opposite direction is reduced.
The second object of the present invention is to provide a kind of target detection device for aligning based on sliding window, the device packet
The computer program that includes memory, processor and storage on a memory and can run on a processor, the processor execute
Following steps are realized when described program, comprising:
The width and Jump step of sliding window are set, the image of target to be detected is divided using sliding window, if obtaining
Dry candidate target region；
All candidate target regions feeding CNN neural network is trained processing, obtains all candidate target regions
Confidence level；
Selection confidence level maximum value is worth on the basis of index region corresponding with the maximum value；
Candidate target region is cut and combined with a reference value using aligning method, forms new target area
Domain.
Compared with prior art, the beneficial effects of the present invention are:
(1) size of sliding window is arranged according to the mean size of examined object by the present invention, so that combining and cutting
There is preferable elasticity when target area, the region where target can be detected under less combination, trimming operation；And it will move
Dynamic stride is set below the half of the size of sliding window, under the premise of guaranteeing that window has greater overlap area, improves target inspection
The speed of survey；
(2) present invention is based on breadth traversal method, according to sliding window size and detection target actual size addition
The depth of breadth traversal is traversal constraint condition, i.e., breadth traversal depth capacity is less than or equal to be combined, cut when 2, in turn
It is described the height of target to be detected is fat or thin, rather than is an original square, effectively improve the accurate of target detection
Degree.
Detailed description of the invention
The accompanying drawings constituting a part of this application is used to provide further understanding of the present application, and the application's shows
Meaning property embodiment and its explanation are not constituted an undue limitation on the present application for explaining the application.
Fig. 1 is the target detection aligning method flow diagram based on sliding window of the embodiment of the present invention one；
Fig. 2 is the target detection aligning method flow diagram based on sliding window of the embodiment of the present invention two；
Fig. 3 is image candidate target area attribute value schematic diagram；
Fig. 4 is depth capacity schematic diagram；
Fig. 5 is to cut candidate region exemplary diagram；
Fig. 6 is combination candidate region exemplary diagram.
Specific embodiment
The invention will be further described with embodiment with reference to the accompanying drawing.
It is noted that following detailed description is all illustrative, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has usual with the application person of an ordinary skill in the technical field
The identical meanings of understanding.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular
Also it is intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or their combination.
As background technique is introduced, that there are window sizes is solid for existing sliding window (sliding window) search technique
Fixed, the size of segmented image will not change because of target sizes, if there is multiple groups sliding window of different sizes to work at the same time, be bound to
It will increase calculation amount, influence efficiency, when sliding stride is intensive, data volume increases, speed is influenced, when sliding stride is excessive,
Influence the deficiency of Detection accuracy.
In view of the above-mentioned deficiencies, the embodiment of the present invention one provides a kind of target detection aligning based on sliding window
Method.As shown in Figure 1, this method comprises the following steps:
S101 is arranged sliding window size and Jump step, utilizes sliding window segmented image.
First acquire the image of examined object；Then sliding window size and Jump step are set.
When sliding window size is arranged, usually default sliding window is rectangular, and the size of sliding window will be according to detectable substance
The mean size of body and determine, should not select it is excessive, although big window can reduce to a certain extent number of windows, improve
Speed is detected, but sliding window is excessive, causes noise in window excessive, influences recognition accuracy.
Choose Jump step s when, Jump step s will be lower than sliding window half, i.e. 0 < s≤slide.width,
Slide.width is the width of sliding window；Guarantee that sliding window has greater overlap area, improves recognition accuracy.
After setting sliding window size and Jump step, start the original graph of segmentation object detection using sliding window
Picture obtains several candidate target regions.
Obtained all candidate target regions feeding CNN neural network is trained processing, obtains all candidates by S102
The confidence level of target area.
In training CNN neural network weight, by candidate target area obtained by all sliding windows in addition to target area
Domain will be greater than all slidings of threshold value I as noise with target area correlation ratio IoU (Intersection over Union)
Candidate target region obtained by window is put into training in CNN neural network as target.
The expression formula of the target area correlation ratio IoU are as follows:
In formula, Area of Overlap indicates correct result region and the intersection of testing result region；Area of Union
Indicate correct result region and testing result region union；Detection result indicates target detection position；
GroundTruth indicates target actual position.
When noise region is excessive, it can use some noise regions of random sampling random () method random erasure, or
Person reduces some pictures of corresponding training set.
Wherein the selection needs of threshold value I (0 < I < 1) are adjusted according to hands-on situation, i.e., to the side having compared with high-accuracy
To adjusting.
Finally, as CNN neural network output be all candidate target regions obtained by sliding window confidence level score
[1]、score[2]、...、score[n]。
S103 selects maximum value max_score=max { score in all confidence levels of CNN neural network output
[1], [2] score ... score [n] } it is worth on the basis of corresponding with maximum value index region v ∈ (1, n).
In the present embodiment, maximum value is selected in all confidence levels of CNN neural network output, that is, selects target to be detected
The position being most likely at, the then point centered on this position, then be adjusted.
When sliding window sequence divides original image, having array bounds, also sequence has recorded each
The information of image block after segmentation is one-to-one with score [i].If socre [v] is maximum value, then bounds
[v] is exactly corresponding image block, referred to as index region.
S104, using aligning method BSF_Revise_bounds () and a reference value max_score, v to candidate target
Region is cut and is combined, and new target area is formed.
The present invention is based on BFS (Breadth-First Search) breadth traversal, according to sliding window size and inspection
The depth extent for surveying target actual size addition breadth traversal is traversal constraint condition, i.e. breadth traversal depth capacity extent
≤ 2 carry out aligning.
The cutting and combined method specifically:
Centered on maximum confidence max_score manipulative indexing region, spread around in breadth traversal BFS method；
By diffusion zone confidence level score [w] compared with the maximum confidence max_score in index region；Wherein this hair
It is bright to be provided with three threshold values, respectively T1、T2、T3, T1Indicate diffusion zone confidence level score [w] and maximum confidence max_
The difference of score indicates that the connection in two regions is strong and weak；T2For confidence level activation threshold, if current diffusion zone confidence level
Score [w] is greater than T2, indicate that current region confidence level is higher；T3Inhibit threshold value for confidence level, if current diffusion zone confidence level
Score [w] is less than T3, indicate that current region confidence level is lower；T1、T2、T3Size setting can be obtained according to CNN neural network
Actual degree of belief and detection target contact flexible setting.
If current diffusion zone confidence level score [w] and maximum region confidence level max_score difference are less than T1And it is current
Diffusion zone confidence level score [w] is greater than T2, illustrate diffusion zone and index region close relation, and diffusion zone w is that height can
Reliability region.Boundary should be expanded to diffusion zone w corresponding direction at this time, and consider that the peripheral region adjacent with diffusion zone w is
No also includes target area.
If current diffusion zone confidence level score [w] is less than T3, then it is weaker to illustrate that the region w is contacted with target area, and mesh
Mark region corresponds in candidate region in max_score, should reduce the boundary of diffusion zone w corresponding direction at this time, while abandoning examining
Consider the extension of diffusion zone w corresponding direction.
S105, target area coordinates after output correction.
The embodiment of the present invention proposes a kind of target detection aligning method based on sliding window, according to detection object
Size setting sliding window size so that segmented image size can because target size and change；And it will mobile cloth
Width is set below the half of the size of sliding window, guarantees that window has greater overlap area, improves recognition accuracy；According to sliding
The depth of window size and detection target actual size addition breadth traversal is traversal constraint condition, i.e. breadth traversal depth capacity
Aligning is carried out when less than or equal to 2, effectively improves the accuracy of target detection.
In order to allow those skilled in the art to be best understood from the present invention, second embodiment of the present invention provides a kind of based on cunning
The target detection aligning method of dynamic window, as shown in Fig. 2, method includes the following steps:
S201 chooses the width slide.width and Jump step s of sliding serial ports, treats detectable substance using sliding window
The image of body is handled, and the candidate region 1~n is obtained.
Obtained all candidate regions feeding CNN neural network is trained by S202, uses trained CNN nerve
Network obtains candidate region confidence level score [i], as shown in Figure 3.
S203 chooses maximum confidence using maximum value max method in all candidate region confidence level score [i]
Max_score and maximum confidence manipulative indexing region v；Using maximum confidence max_score and index region v as
A reference value.
The borderline region of S204, calling station antidote BFS_Revise_Bounds () correction candidate region.
BFS_Revise_bounds () aligning method specifically:
S2041 initializes queue Q；Initialize candidate region depth extent [1,2,3....n]=∞.
Queue Q is that an one-dimensional queue has the characteristics that first in first out, stores the central point of current breadth traversal, is just started
Maximum value corresponding region v, extent=0 are stored, then by circulation, the several points in left, up, right, down of v can be stored, combine,
After cutting, circulation terminates extent=1 at this time, into recycling next time；The point centered on left point is recycled next time, respectively
The left, up, right, down of left point are added, combination is cut, and centered on above, is separately added into the left, up, right, down of point, cutting group
It closes, circuits sequentially, completed until the left, up, right, down of lower point traverse, at this time extent=2；There is no new element in queue Q
, exit circulation.
Extent array stores distance of the index point apart from maximum value v, and initialization is positive infinite.
S2042 accesses maximum confidence manipulative indexing region point v；Visited [v]=1, extent [v]=0；By center
In point v enqueue Q.
Whether visited array representation region accessed, if visited [v]=1 expression accessed；Visited [v]=
0；Expression has not visited.
S2043, if continuing to execute when queue Q non-empty, otherwise jumping to S20413.
Team's head element of S2044, queue Q go out team, are assigned to temp.
S2045, w equal to temp it is left it is upper it is right under index point.
It is left it is upper it is right under mean access one every time, the sequence only accessed is different, for example is visited for the first time
Ask a left side, on back-call, third time access it is right, under the 4th access.This is in Do statement, if left for the first time do not deposit
Just in access.
S2046 executes downwards if w exists, otherwise jumps to S2044；
S2047 executes downwards if w is not access, otherwise jumps to S20412；
S2048 accesses w, sets 1, extent [w]=extent [w]+1 for visited [w]；As shown in Figure 4；
S2049, if the difference max_ of the confidence level score [w] and maximum confidence max_score of diffusion zone w
score-score[w]≤T1And confidence level score [w] >=T of diffusion zone w2, then from the corresponding region central point v it is left/upper/
The right side/lower boundary is expanded to the corresponding left/upper/right side/lower boundary of diffusion zone w, as shown in Figure 5；
S20410, if extent [w] < 2, w enqueue Q；
S20411, if confidence level score [w]≤T of diffusion zone w3, and diffusion zone w left/right/up/down boundary does not have
Expand；Then the corresponding region maximum confidence max_score left/right/up/down boundary reduces d1；As shown in Figure 6；
Next (upper right under) adjacent confidence point of S20412, w=v, jump to S2046；Until w is not present, jump
To S2044；
S20413 returns to target area boundaries coordinate after amendment.
Wherein, T1、T2、T3Size setting can be according to the connection for the actual degree of belief and detection target that CNN neural network obtains
It is flexible setting, d1It is arranged according to sliding window size and target object flexible in size.
The embodiment of the present invention proposes a kind of target detection aligning method based on sliding window, according to detection object
Mean size determine the size of sliding window, Jump step is lower than the half of window, guarantees that window has greater overlap area, improves
Recognition accuracy；It centered on maximum confidence corresponding region, spreads in BFS method, is obtained according to CNN neural network around
Actual degree of belief and detection target contact flexible setting threshold value T1、T2、T3Borderline region is corrected, improves target identification
Accuracy and speed.
Above-mentioned, although the foregoing specific embodiments of the present invention is described with reference to the accompanying drawings, not protects model to the present invention
The limitation enclosed, those skilled in the art should understand that, based on the technical solutions of the present invention, those skilled in the art are not
Need to make the creative labor the various modifications or changes that can be made still within protection scope of the present invention.
Claims (8)
1. a kind of target detection aligning method based on sliding window, characterized in that the following steps are included:
The width and Jump step of sliding window are set, the image of target to be detected is divided using sliding window, obtains several
Candidate target region；
All candidate target regions feeding CNN neural network is trained processing, obtains the confidence of all candidate target regions
Degree；
Selection confidence level maximum value is worth on the basis of index region corresponding with the maximum value；
Candidate target region is cut and combined with a reference value using aligning method, forms new target area.
2. the target detection aligning method according to claim 1 based on sliding window, characterized in that according to all
The mean size of examined object determines the width of sliding window；The mobile split of sliding window is less than or equal to sliding window width
Half.
3. the target detection aligning method according to claim 1 based on sliding window, characterized in that described by institute
There is candidate target region to be sent into CNN neural network the step of being trained processing to include:
Using the candidate target region with target area correlation ratio less than threshold value I as noise, will be greater than with target area correlation ratio
The candidate target region of threshold value I is separately input to training in CNN neural network as target；
The confidence level of all candidate target regions is obtained using trained CNN neural network.
4. the target detection aligning method according to claim 3 based on sliding window, characterized in that work as noise regions
When domain is excessive, using the multiple noise regions of arbitrary sampling method random erasure, or the picture of corresponding training set is deleted.
5. the target detection aligning method according to claim 1 based on sliding window, characterized in that in CNN mind
In all confidence levels through network output, confidence level maximum value is chosen, by the confidence level maximum value and the confidence level maximum value pair
It is worth on the basis of the index region answered.
6. the target detection aligning method according to claim 1 based on sliding window, characterized in that according to sliding
The size of the width of window and target to be detected, using the depth of breadth traversal as traversal constraint condition, when breadth traversal most
Big depth is less than or equal to carry out aligning when 2.
7. the target detection aligning method according to claim 1 based on sliding window, characterized in that the utilization
Aligning method and a reference value carry out cutting and combined method to candidate target region are as follows:
The point centered on maximum confidence manipulative indexing region；
Setting area power threshold value T1, confidence level activation threshold T2Inhibit threshold value T with confidence level3；
Using center point area as origin, four, upper and lower, left and right adjacent area is current candidate diffusion zone；
Based on breadth traversal algorithm, current diffusion zone confidence level and index region central point maximum confidence are made the difference,
And threshold value is inhibited to be compared respectively diffusion zone confidence level and confidence level activation threshold, confidence level；
If the difference of current some direction confidence level of diffusion zone and maximum confidence is less than T1, and diffusion zone confidence level is greater than
T2, then central area is expanded into boundary to the corresponding direction of the diffusion zone；
If current some direction confidence level of diffusion zone is less than T3, then illustrate that target area does not extend in the direction, mesh to be detected
It is marked in the corresponding index region of maximum value, central area corresponding direction opposite direction is reduced.
8. a kind of target detection device for aligning based on sliding window, characterized in that including memory, processor and storage
On a memory and the computer program that can run on a processor, the processor realize following step when executing described program
Suddenly, comprising:
The width and Jump step of sliding window are set, the image of target to be detected is divided using sliding window, obtains several
Candidate target region；
All candidate target regions feeding CNN neural network is trained processing, obtains the confidence of all candidate target regions
Degree；
Selection confidence level maximum value is worth on the basis of index region corresponding with the maximum value；
Candidate target region is cut and combined with a reference value using aligning method, forms new target area.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810871600.3A CN109166106B (en) | 2018-08-02 | 2018-08-02 | Target detection position correction method and device based on sliding window |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810871600.3A CN109166106B (en) | 2018-08-02 | 2018-08-02 | Target detection position correction method and device based on sliding window |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109166106A true CN109166106A (en) | 2019-01-08 |
| CN109166106B CN109166106B (en) | 2021-07-30 |
Family
ID=64898741
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810871600.3A Active CN109166106B (en) | 2018-08-02 | 2018-08-02 | Target detection position correction method and device based on sliding window |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109166106B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112751633A (en) * | 2020-10-26 | 2021-05-04 | 中国人民解放军63891部队 | Broadband spectrum detection method based on multi-scale window sliding |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106327507A (en) * | 2016-08-10 | 2017-01-11 | 南京航空航天大学 | Color image significance detection method based on background and foreground information |
| US20170147905A1 (en) * | 2015-11-25 | 2017-05-25 | Baidu Usa Llc | Systems and methods for end-to-end object detection |
| CN107154024A (en) * | 2017-05-19 | 2017-09-12 | 南京理工大学 | Dimension self-adaption method for tracking target based on depth characteristic core correlation filter |
| KR20170131081A (en) * | 2016-05-20 | 2017-11-29 | 주식회사 에스원 | Method and appterminal for detecting object |
| CN107665336A (en) * | 2017-09-20 | 2018-02-06 | 厦门理工学院 | Multi-target detection method based on Faster RCNN in intelligent refrigerator |
| CN107679469A (en) * | 2017-09-22 | 2018-02-09 | 东南大学—无锡集成电路技术研究所 | A kind of non-maxima suppression method based on deep learning |
| CN108062531A (en) * | 2017-12-25 | 2018-05-22 | 南京信息工程大学 | A kind of video object detection method that convolutional neural networks are returned based on cascade |
| CN108154520A (en) * | 2017-12-25 | 2018-06-12 | 北京航空航天大学 | A kind of moving target detecting method based on light stream and frame matching |
| CN108304808A (en) * | 2018-02-06 | 2018-07-20 | 广东顺德西安交通大学研究院 | A kind of monitor video method for checking object based on space time information Yu depth network |
| CN108319949A (en) * | 2018-01-26 | 2018-07-24 | 中国电子科技集团公司第十五研究所 | Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image |
-
2018
- 2018-08-02 CN CN201810871600.3A patent/CN109166106B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170147905A1 (en) * | 2015-11-25 | 2017-05-25 | Baidu Usa Llc | Systems and methods for end-to-end object detection |
| KR20170131081A (en) * | 2016-05-20 | 2017-11-29 | 주식회사 에스원 | Method and appterminal for detecting object |
| CN106327507A (en) * | 2016-08-10 | 2017-01-11 | 南京航空航天大学 | Color image significance detection method based on background and foreground information |
| CN107154024A (en) * | 2017-05-19 | 2017-09-12 | 南京理工大学 | Dimension self-adaption method for tracking target based on depth characteristic core correlation filter |
| CN107665336A (en) * | 2017-09-20 | 2018-02-06 | 厦门理工学院 | Multi-target detection method based on Faster RCNN in intelligent refrigerator |
| CN107679469A (en) * | 2017-09-22 | 2018-02-09 | 东南大学—无锡集成电路技术研究所 | A kind of non-maxima suppression method based on deep learning |
| CN108062531A (en) * | 2017-12-25 | 2018-05-22 | 南京信息工程大学 | A kind of video object detection method that convolutional neural networks are returned based on cascade |
| CN108154520A (en) * | 2017-12-25 | 2018-06-12 | 北京航空航天大学 | A kind of moving target detecting method based on light stream and frame matching |
| CN108319949A (en) * | 2018-01-26 | 2018-07-24 | 中国电子科技集团公司第十五研究所 | Mostly towards Ship Target Detection and recognition methods in a kind of high-resolution remote sensing image |
| CN108304808A (en) * | 2018-02-06 | 2018-07-20 | 广东顺德西安交通大学研究院 | A kind of monitor video method for checking object based on space time information Yu depth network |
Non-Patent Citations (4)
| Title |
|---|
| JUFENG ZHAO等: "Real-time automatic small target detection using saliency extraction and morphological theory", 《OPTICS & LASER TECHNOLOGY》 * |
| WEI ZHANG等: "Real-time Accurate Object Detection using Multiple Resolutions", 《2007 IEEE 11TH INTERNATIONAL CONFERENCE ON COMPUTER VISION》 * |
| 岳国伟: "基于动态窗口的图像线性目标快速检测算法", 《山东科技大学学报（自然科学版）》 * |
| 温捷文等: "实时目标检测算法YOLO的批再规范化处理", 《计算机应用研究》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112751633A (en) * | 2020-10-26 | 2021-05-04 | 中国人民解放军63891部队 | Broadband spectrum detection method based on multi-scale window sliding |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109166106B (en) | 2021-07-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | Matching-cnn meets knn: Quasi-parametric human parsing | |
| CN111291739B (en) | Face detection and image detection neural network training method, device and equipment | |
| CN100561503C (en) | A kind of people's face canthus and corners of the mouth location and method and the device followed the tracks of | |
| CN109902798A (en) | The training method and device of deep neural network | |
| CN106557173B (en) | Dynamic gesture identification method and device | |
| CN107145845A (en) | The pedestrian detection method merged based on deep learning and multi-characteristic points | |
| CN109685037A (en) | A kind of real-time action recognition methods, device and electronic equipment | |
| CN107818333A (en) | Robot obstacle-avoiding action learning and Target Searching Method based on depth belief network | |
| CN110298330A (en) | A kind of detection of transmission line polling robot monocular and localization method | |
| CN106383912A (en) | Picture retrieval method and apparatus | |
| CN109902761A (en) | A kind of fishing feelings prediction technique based on marine environment Factor Fusion and deep learning | |
| CN109614990A (en) | A kind of object detecting device | |
| CN110163864A (en) | Image partition method, device, computer equipment and storage medium | |
| CN107886711A (en) | A kind of pedestrian running red light caution system and implementation method | |
| CN116934807A (en) | Twin network single target tracking method and device | |
| CN109325407A (en) | Optical remote sensing video object detection method based on F-SSD network filtering | |
| CN109166106A (en) | A kind of target detection aligning method and apparatus based on sliding window | |
| CN109460820A (en) | A kind of neural network training method, device, computer equipment and storage medium | |
| Ahmed et al. | Robust Object Recognition with Genetic Algorithm and Composite Saliency Map | |
| CN109299641A (en) | A kind of train dispatcher's fatigue monitoring image adaptive Processing Algorithm | |
| WO2021051268A1 (en) | Machine vision-based tree type identification method and apparatus | |
| CN111444313A (en) | Knowledge graph-based question and answer method and device, computer equipment and storage medium | |
| CN114693966A (en) | Target detection method based on deep learning | |
| CN109858402A (en) | A kind of image detecting method, device, terminal and storage medium | |
| CN113762051B (en) | Model training method, image detection device, storage medium and equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |