CN107194348A - The domain color recognition methods of target object in a kind of image - Google Patents
The domain color recognition methods of target object in a kind of image Download PDFInfo
- Publication number
- CN107194348A CN107194348A CN201710357004.9A CN201710357004A CN107194348A CN 107194348 A CN107194348 A CN 107194348A CN 201710357004 A CN201710357004 A CN 201710357004A CN 107194348 A CN107194348 A CN 107194348A
- Authority
- CN
- China
- Prior art keywords
- color
- target object
- image
- domain
- region
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
-
- 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
- G06V10/56—Extraction of image or video features relating to colour
Abstract
The domain color recognition methods of target object, is related to field of image recognition in a kind of image proposed by the present invention.This method determines the classification number of RGB color color value division and clustered, the color mapping table of generation color value and colour type first；Then to every image of specified data set, the image in the region is transformed into HSV space, reconvert returns rgb space after being clustered by the region in positioning selected digital image where target object；According to statistical law, different weights are assigned for the pixel of each diverse location in target object region, obtain the Weighted color histogram in the region, take the domain color of the target object that the corresponding color of highest rectangle is obtained as identification in histogram, and the corresponding colour type of domain color is searched in color mapping table, record annotation results.The present invention improves the degree of accuracy of colour recognition and the scalability of color mark.
Description
Technical field
The present invention relates to field of image recognition, the domain color recognition methods of target object in more particularly to a kind of image.
Background technology
Color is a underlying attribute of target object itself, and it is that machine and the mankind distinguish the main of different target object
One of factor.The color of target object in image is recognized, can be can be applied in for more accurately distinguishing generic object
In Data Preparation Process in machine learning, the speed of data mark is substantially improved, save time for manually marking and manpower into
This.According to material whether can resistance to deformation, target object can be divided into two kinds of rigid objects and flexible article.For target
Object, for the flexible article being made up of various material, multiple element, on the one hand due to target object or its
The characteristics of background self color is abundant causes target object to show multiple color in the picture；On the other hand, target object
Color influenceed also very big by factors such as object self-deformation, surrounding environment light, image taking angle, camera exposure parameters,
Even same color can also show a variety of visually differentiated colors in the picture.In addition, color space is huge,
The features such as color Distribution value is very scattered in color value quantity very abundant, true environment also gives the identification of the domain color of target object
Very big difficulty is brought with mark.
The identification of target object domain color refers to, for one group of given image sequence, extract in image sequence in image
In every image the primary color information of target object and in certain specific color space specify its belonging to colour type, its
The primary color of middle target object refers to the maximum color of pixel quantity accounting in target object.Special instruction, the sense of image is emerging
Interesting region refers to the region where target object in image, the region can be oriented by the detection of artificial or machine come.
At present, the identification of existing domain color and mask method can be divided into artificial mark, Octree conflation algorithm and color away from
Three kinds from matching method methods, will describe in detail to existing method below：
1 artificial mark；
Artificial mark refers to the recognition capability by people, and object is manually specified by mark personnel by artificial labeling system
Colour type belonging to body.The core concept of this method is, in advance by manually probably counting target in image sequence to be marked
The distributed number of the color attribute of object, therefrom selects several common colour types as colour type to be selected；Annotation process
In the thumbnail of the area-of-interest of every image is presented to mark personnel by labeling system, by mark, personnel are sentenced by subjectivity
It is disconnected, selection and target object domain color in image the most close colour type conduct from the colour type to be selected of above-mentioned determination
The domain color classification of the target object.This method still receives on small-scale data set and color discrimination large data sets,
But it is due to completely, by manually operating, certainly to be had the following disadvantages under big data environment now：
(1) when data scale is big, artificial mark cycle length, cost are high；
(2) need to know in advance the distribution of color of target object in data set to be marked, to data in itself and before mark
Preparation require it is too high；The error of statistics can largely effect on the quality of mark；
(3) easily there is marking error.When image down is to very little, artificial mark is easy to by target in image
The influence of background color around object, and by mistake not being that the color of target object domain color is labeled as the main face of target object
Color.
(4) standard of mark differs, and as a result there may be conflicting.When image target object includes multiple color
Wait, mark personnel are difficult to determine that domain color is any；The even single object of color, the result of different people's marks also may be used
Can be different.The subjectivity of mark personnel is larger in whole annotation process, and annotation results are uncontrollable.
2 Octree conflations algorithm
The main thought of this method is to have chosen the exact amount of the data set color category to be marked first.Then for
Each image, the approximate region where target object is oriented by certain method, is existed by calculating this block image-region
The color histogram of rgb space, will appear from frequency highest color as the domain color of target object in image.Only consider face
Color tablets degree is 256*256*256 rgb space, just there is 16777216 kinds of color values, so marks out the domain color come such
Distribution in big color space obviously can be very sparse.In order to give color designated color classification, this method utilizes the spy of Octree
Property, the domain color of all images is carried out to merger from bottom to top in Octree, colour type quantity convergence mesh to the last
Untill marking colour type quantity.So each color can find an one's own color from bottom to top in Octree
Classification.
This method major advantage is that thinking is succinct, it is easy to accomplish.But have the following disadvantages:
(1) standard of domain color classification mark is related to data set.When color merger being carried out in Octree, it is impossible to
The classification of color is determined in advance, only waits until just to can determine that after the completion of merger.Then the color of the classification of color just with data set
Distribution it is closely related, even the standard that the colour type of same mark task different batches data set is divided is completely not yet
Together.
(2) annotation results of identical category different batches can not simply merge.Because domain color classification is related to data set,
The color annotation results of the data set of different batches can not directly merge.Want merging data collection, it is necessary to again all
Data are marked one time, and this expands to the increment of data set brings many inconvenience.
(3) Octree easily cause colour type granularity it is uneven.When the frequency that a certain class color occurs in data set
When very high, this kind of color granularity can be divided very thin；And the not high color granularity of the frequency of occurrences is then divided very thick, cause
Same image is widely different in the annotation results of different pieces of information collection.
3 color distance matching methods
The main thought of color distance matching method is as follows：Some common colour types are chosen first as color to be selected,
Then for each image, the position of target object substantially is oriented, in certain particular color space, the object is calculated
The color histogram of body, will appear from frequency highest color as the domain color classification of the object.Afterwards in a certain particular color
In space, the color distance between the domain color and each colour type for choosing in advance of object is calculated as fraction, is taken point
Minimum colour type is counted as the classification of the domain color of the object.
The standard of the algorithm is unified, the annotation results of different pieces of information collection can with general, very convenient data set annotation results
Fusion.But have the following disadvantages：
(1) classification of the color of color hop region easily specifies error in color space.Although the phase in color space
The distance between near color very little, but actually in any color space, color distance is small can not to draw color really
Similar conclusion.The distance of any two color distribution overall due to not accounting for color space so that in some critical bars
Under part, such as color change region, the colour type chosen by small distance differs greatly with actual domain color.
(2) quantity of colour type, coverage and granularity are difficult to be determined in advance.Manually it is determined that colour type to be marked
When, there is larger subjectivity, it is incomplete distribution of color extremely easily occur, the problems such as color granularity is uneven.
In brief, there are some unavoidable shortcomings in the domain color identification technology of existing complex target object.It is multiple
The characteristics of domain color identification problem of miscellaneous target object, is that color space is huge but data images distribution of color uncertain, figure
The domain color and non-master distribution of color of target object color category complexity and target object region as in are not known, image
Illumination variation is various, and the mark colour type of domain color is difficult determination etc..
Clustering algorithm is a special kind of skill for analysis of statistical data, in many fields by extensive use.Cluster be with
Based on similitude, similar object is divided into by the method for static classification different groups or more subsets, so
The member object allowed in same subset has some similar attributes.Current clustering algorithm has on the color quantizing of image
Using the pixel of image such as is collected as into several classifications.
Convolutional neural networks are that the artificial neuron in a kind of feedforward neural network, network can respond surrounding cells, main
To include convolutional layer and pond layer, possess good nonlinear fitting ability, possess at present in object detection, classification and identification
It is widely applied.
The content of the invention
The purpose of the present invention is to overcome the weak point of prior art, propose a kind of domain color of target object in image
Recognition methods.This method has good ease for use compared with the conventional method, and what the degree of accuracy, color in colour recognition were marked can
All have greatly improved in the efficiency of autgmentability and entirety.
The domain color recognition methods of target object in a kind of image proposed by the present invention, it is characterised in that including following step
Suddenly：
1) color mapping table is generated；Comprise the following steps that：
1-1) according to the requirement to Color-sensitive degree, the classification number X that RGB color color value is divided is determined, wherein 0
<X<16777216,16777216 be the sum of color value in RGB color；
1-2) 16777216 color values of RGB color are gathered for X different colour types using clustering method；
1-3) according to step 1-2) cluster result, by the color value that each classification is included in X colour type after cluster
Mapped one by one with the title of each classification, one group of 16777216 color value of generation reflect to X the incomplete of colour type
Penetrate the color mapping table Ω of relation, generation color value and colour type；
2) any one image in specified data set is selected, the region of target object in the picture, i.e. target is oriented
Region where object；
3) by step 2) in the image in region where the target object that determines arrived by color space conversion method migration
HSV space, poly- that the image is converted back into rgb space again after C colour type by clustering algorithm, wherein C represents object
The color category of body, 0 ＜ C≤16；
4) for step 3) image completed is clustered, step 2 is according to target object) the middle target object place determined
The statistical law of position distribution in region, is that the pixel of each diverse location (x, y) in the region where target object is assigned
Give corresponding difference weight f (x, y)；
5) to step 3) pixel of image that completes of cluster travels through, the frequency and step 4 occurred according to color value)
In the weight that is assigned to each pixel, calculate and obtain the weighted color of a target object region containing C rectangle
Histogram, each rectangle corresponds to step 3 respectively) cluster an obtained colour type；Weighted color histogram is calculated,
Take the domain color ω of the present image target object that the corresponding color of highest rectangle is obtained as identification in histogram；
6) in step 1) the corresponding colour type Φ of domain color ω are searched in obtained color mapping table Ω and record result；
7) repeat step 2) to step 6), complete to know the domain color for specifying target object in all images in data set
Not.
The features of the present invention and beneficial effect are：
1 target object domain color is influenceed to significantly reduce by external factor such as color value distortion and illumination.By region of interest
The image in domain has been transformed into HSV space and processed again so that influence of the illumination to image is greatly reduced；Made using the method for cluster
On image original identical color caused visually different pixel also to get together due to various factors, as a class face
Color.
The color of 2 target objects is more protruded.It is densely distributed sparse with distribution that area-of-interest can be divided into target object
Part, the present invention by giving the colouring information in the densely distributed region of target object higher importance so that
Inside the color histogram entirely weighted, the color close to the densely distributed region of target object is more protruded, that is, object
The color of body is more prominent.
3 color granularities are controllable, and coverage rate is complete.The demand that the present invention can be marked according to actual color, sets different
The number of cluster.The number of cluster is more, and the discrimination of color is more obvious；When clusters number is few, it can also be included per class
More more close color relatively.No matter color granularity, all colours classification is all the completely whole color space of covering
's.
The division of 4 colour types is the optimal solution of color space, and it is objective to divide, unrelated with data set.Divide the kind of color
Using the method for cluster when class, it is contemplated that the correlation between color space color.Cluster the mistake of such a iteration
What journey was obtained is a kind of global optimal solution so that color of the color got together per cluster all than being got together in other clusters
It is visually closer.Such division is the objective reality according to color space, without human intervention and uncertain factor.
5 are with good expansibility, and are conducive to the expansion stage by stage of data set.The colour type criteria for classifying is consistent,
The annotation results of different batches can directly merge in the case of granularity identical, and face can also be passed through when granularity is different
The mapping relations of color classification are merged.
6 annotating efficiencies are high.The process entirely marked does not have artificial participation, when to every image clustering number x=8,
The mark of average every image only needs to 0.259S, and test platform is the system of Intel Core i7-4790 CPU, Centos 7.
Multi-process can cause annotating efficiency higher.
Brief description of the drawings
Fig. 1 is the flow chart of the inventive method.
Fig. 2 is the effect diagram of the embodiment of the present invention.
Embodiment
The domain color recognition methods of target object in a kind of image proposed by the present invention, below in conjunction with the accompanying drawings and specific implementation
That the present invention is described in more detail is as follows for example.Embodiment of the present invention is exemplary, is only used for explaining the present invention, and
It is not construed as limiting the claims.
The domain color recognition methods of target object, flow are as shown in figure 1, this method bag in a kind of image proposed by the present invention
Include following steps：
1) color mapping table is generated；Comprise the following steps that：
1-1) according to the requirement to Color-sensitive degree, the classification number X that RGB color color value is divided is determined, wherein 0
<X<16777216,16777216 be the sum of color value in RGB color；
Specifically, when exigent to Color-sensitive degree when, X=10000 can be taken, now red can be divided
For the classification of dark red and pale red etc. different colours grade；When less demanding to Color-sensitive degree when, X=can be taken
500, mainly it is to discriminate between red and green etc..
1-2) RGB color is gathered for X different colour types using clustering method；
Specifically, RGB color, 256*256*256=16777216 color value is clustered altogether.During cluster,
Consider the speed and accuracy of cluster, the method that the different clustering algorithms initialization cluster center different with selection can be chosen.
The present embodiment selects kmeans algorithms, and the cluster center of cluster is initialized using kmeansplusplus methods, and so cluster is obtained
The color value quantity that includes of different colour types it is very uniform.
1-3) according to step 1-2) cluster result generation color value and colour type color mapping table Ω；
Specifically, according to the result of cluster, by each classification is included in X colour type after cluster color value and each
The title of classification is mapped one by one, the incomplete mapping relations of one group of 16777216 color value of generation to X colour type.
So for some color value, it is possible to obtain the corresponding colour type of the color value by mapping relations.
2) any one image in specified data set is selected, the region of target object in the picture, i.e. target is oriented
Area-of-interest residing for object.Specified data set refers to one group of picture sequence that colour recognition and mark are carried out using this method
Picture number is not limited in row, data set；
Specifically, the localization method of area-of-interest residing for target object can use existing object detection algorithms.This
Embodiment uses the object detection algorithms based on depth convolutional neural networks, can be quickly and accurately positioned target object in figure
Position as in.
3) by step 2) in the image of area-of-interest residing for the target object that determines pass through color space conversion method
HSV space is transformed into, it is poly- that the image is converted back into rgb space again after C colour type by clustering algorithm, wherein 0 ＜ C≤
16；
Specifically, color space conversion method can use existing algorithm, and the present embodiment color space conversion is used
OpenCV color space conversion algorithm.Due to HSV space by illumination separately as a passage, such conversion subtracts significantly
Influence of the illumination to image is lacked.Limited in view of the number of color expressed by an image, we select cluster numbers C in 1-16
Between.If C is too big, the effect that close color is got together can not be played.By by the face of the area-of-interest of image
Color is clustered in HSV space, and close color value is got together in region, is gathered for C class.Actually these close face
Many originally same color values of pixel in colour, are simply influenceed to cause color value to be sent out in color space by various factors
Give birth to small skew and become different colors.We are greatly reduced by the method for cluster because the influence of these factors is led
The problem of colouring information mistake and mixed and disorderly colouring information of cause.
4) for step 3) cluster complete image, step 2 is according to target object) in determine area-of-interest in
Position distribution statistical law, be that the pixel of each diverse location (x, y) in area-of-interest assigns corresponding different power
Weight f (x, y).
Specifically, it is believed that the color of target object is domain color interested.The position statistical law of target object is
Refer to the position regularity of distribution generally of the target object in area-of-interest, can be determined by Statistics-Based Method, no
The statistical law of same position distribution has different influences to last identification and annotation results.Position distribution in the present embodiment
Statistical law be adjusted so as to after artificial determine Lai.Because the area-of-interest where target object in the present embodiment is by machine
Detect, so target object is typically in area-of-interest center, then domain color interested, which is mainly distributed on, feels emerging
Interesting region lean on ectocentral position, in other words the statistical law of the position distribution of the target object of the present embodiment be region center
Intensive, surrounding is sparse.By assigning higher weight to the pixel positioned at area-of-interest center, area-of-interest edge is given
Pixel assign lower weight so that the color value of target object is more protruded in area-of-interest, greatly reduce by
The mistake for causing object domain color in the influence of the colouring information at area-of-interest edge is marked.
5) to step 3) pixel of image that completes of cluster travels through, it is considered to the frequency and step 4 of color value appearance)
The weight assigned to each pixel, calculates the Weighted color histogram for obtaining an area-of-interest containing C rectangle, often
Individual rectangle corresponds to step 3 respectively) cluster an obtained colour type；Weighted color histogram is calculated, histogram is taken
The domain color ω for the present image target object that the corresponding color of middle highest rectangle is obtained as identification.
6) in step 1) the corresponding colour type Φ of domain color ω are searched in obtained color mapping table Ω and record result.
Specifically, by step 1) mapping table set up, the domain color ω institute that can directly find target object is right
The colour type Φ answered.Then the colour type Φ of the target object domain color of the image and the target object of image are recorded
Domain color ω.When its domain color is recorded mainly in view of follow-up data set merging, when the color grain of two datasets
When spending different, it can directly set up new mapping relations and the domain color of the target object of image is mapped to new color grain
Colour type under degree.
7) repeat step 2) to step 6), complete to know the domain color for specifying target object in all images in data set
Not.
The specific embodiment of the present invention, effect diagram is as shown in Figure 2.Mesh in the test pictures of the embodiment of the present invention
The rectangle that mark object region (area-of-interest) has different colours in 8 colour types, histogram refers to different face
Color classification, is arranged in order according to the order of quantity from high to low.Fig. 2 (a) is by step 3) processing after embodiment test
In image in the color histogram of area-of-interest, Fig. 2 (a), rectangle is arranged in order as color 1 to color 8 from high to low, now
Domain color be color 1；Fig. 2 (b) be Fig. 2 (a) histogram pass through step 4) processing after Weighted color histogram, processing
Afterwards, each rectangle putting in order from high to low is changed into：3-1-5-4-2-8-6-7, domain color now is color 3.From figure
As can be seen that the domain color of real target object comes the 3rd in Fig. 2 (a) in embodiment test pictures, it is not qualified as
The domain color of target object, recognizes mistake；And the domain color of real target object comes the 1st in Fig. 2 (b), it is considered as
It is the domain color of target object, identification is correct.
Claims (2)
1. the domain color recognition methods of target object in a kind of image, it is characterised in that comprise the following steps：
1) color mapping table is generated；Comprise the following steps that：
1-1) according to the requirement to Color-sensitive degree, the classification number X that RGB color color value is divided is determined, wherein 0<X<
16777216,16777216 be the sum of color value in RGB color；
1-2) 16777216 color values of RGB color are gathered for X different colour types using clustering method；
1-3) according to step 1-2) result of cluster, by each classification is included in X colour type after cluster color value with respectively
The title of individual classification is mapped one by one, and the incomplete mapping of one group of 16777216 color value of generation to X colour type is closed
The color mapping table Ω of system, generation color value and colour type；
2) any one image in specified data set is selected, the region of target object in the picture, i.e. target object is oriented
The region at place；
3) by step 2) in determine target object where region image pass through color space conversion method migration to HSV sky
Between, poly- that the image is converted back into rgb space again after C colour type by clustering algorithm, wherein C represents the face of target object
Color species, 0 ＜ C≤16；
4) for step 3) image completed is clustered, step 2 is according to target object) the middle target object region determined
The statistical law of interior position distribution, is that the pixel of each diverse location (x, y) in the region where target object assigns phase
The different weight f (x, y) answered；
5) to step 3) pixel of image that completes of cluster travels through, the frequency and step 4 occurred according to color value) in it is right
The weight that each pixel is assigned, calculates the weighted color Nogata for obtaining a target object region containing C rectangle
Figure, each rectangle corresponds to step 3 respectively) cluster an obtained colour type；Weighted color histogram is calculated, cut-off
The domain color ω for the present image target object that the corresponding color of highest rectangle is obtained as identification in square figure；
6) in step 1) the corresponding colour type Φ of domain color ω are searched in obtained color mapping table Ω and record result；
7) repeat step 2) to step 6), complete to recognize the domain color for specifying target object in all images in data set.
2. the method as described in claim 1, it is characterised in that the step 2) in orient the area of target object in the picture
Domain, the localization method used is the object detection algorithms based on depth convolutional neural networks.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710357004.9A CN107194348A (en) | 2017-05-19 | 2017-05-19 | The domain color recognition methods of target object in a kind of image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710357004.9A CN107194348A (en) | 2017-05-19 | 2017-05-19 | The domain color recognition methods of target object in a kind of image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107194348A true CN107194348A (en) | 2017-09-22 |
Family
ID=59874839
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710357004.9A Pending CN107194348A (en) | 2017-05-19 | 2017-05-19 | The domain color recognition methods of target object in a kind of image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107194348A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108805180A (en) * | 2018-05-24 | 2018-11-13 | 北京嘀嘀无限科技发展有限公司 | The detection method and device of target object |
| CN110020673A (en) * | 2019-03-12 | 2019-07-16 | 深圳市象形字科技股份有限公司 | A kind of method of HSV color space color identification and noise filtering |
| CN110349223A (en) * | 2018-04-08 | 2019-10-18 | 中兴通讯股份有限公司 | A kind of method and device of image procossing |
| CN110826418A (en) * | 2019-10-15 | 2020-02-21 | 深圳和而泰家居在线网络科技有限公司 | Face feature extraction method and device |
| CN110969170A (en) * | 2019-12-03 | 2020-04-07 | 北京奇艺世纪科技有限公司 | Image theme color extraction method and device and electronic equipment |
| CN111353536A (en) * | 2020-02-28 | 2020-06-30 | 北京字节跳动网络技术有限公司 | Image annotation method and device, readable medium and electronic equipment |
| WO2020151529A1 (en) * | 2019-01-23 | 2020-07-30 | 阿里巴巴集团控股有限公司 | Clothing image classification and image classification methods and apparatuses, and device |
| CN112561976A (en) * | 2020-12-09 | 2021-03-26 | 齐鲁工业大学 | Image dominant color feature extraction method, image retrieval method, storage medium and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1300804A2 (en) * | 2001-10-05 | 2003-04-09 | Lg Electronics Inc. | Face detecting method by skin color recognition |
| CN101763429A (en) * | 2010-01-14 | 2010-06-30 | 中山大学 | Image retrieval method based on color and shape features |
-
2017
- 2017-05-19 CN CN201710357004.9A patent/CN107194348A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1300804A2 (en) * | 2001-10-05 | 2003-04-09 | Lg Electronics Inc. | Face detecting method by skin color recognition |
| CN101763429A (en) * | 2010-01-14 | 2010-06-30 | 中山大学 | Image retrieval method based on color and shape features |
Non-Patent Citations (1)
| Title |
|---|
| 方珍红等: "一种新的基于加权颜色特征的图像检索算法", 《激光与红外》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110349223A (en) * | 2018-04-08 | 2019-10-18 | 中兴通讯股份有限公司 | A kind of method and device of image procossing |
| CN110349223B (en) * | 2018-04-08 | 2021-04-30 | 中兴通讯股份有限公司 | Image processing method and device |
| CN108805180A (en) * | 2018-05-24 | 2018-11-13 | 北京嘀嘀无限科技发展有限公司 | The detection method and device of target object |
| WO2020151529A1 (en) * | 2019-01-23 | 2020-07-30 | 阿里巴巴集团控股有限公司 | Clothing image classification and image classification methods and apparatuses, and device |
| CN110020673A (en) * | 2019-03-12 | 2019-07-16 | 深圳市象形字科技股份有限公司 | A kind of method of HSV color space color identification and noise filtering |
| CN110826418A (en) * | 2019-10-15 | 2020-02-21 | 深圳和而泰家居在线网络科技有限公司 | Face feature extraction method and device |
| CN110969170A (en) * | 2019-12-03 | 2020-04-07 | 北京奇艺世纪科技有限公司 | Image theme color extraction method and device and electronic equipment |
| CN110969170B (en) * | 2019-12-03 | 2024-03-08 | 北京奇艺世纪科技有限公司 | Image theme color extraction method and device and electronic equipment |
| CN111353536A (en) * | 2020-02-28 | 2020-06-30 | 北京字节跳动网络技术有限公司 | Image annotation method and device, readable medium and electronic equipment |
| CN112561976A (en) * | 2020-12-09 | 2021-03-26 | 齐鲁工业大学 | Image dominant color feature extraction method, image retrieval method, storage medium and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107194348A (en) | The domain color recognition methods of target object in a kind of image | |
| Liu et al. | Person re-identification: What features are important? | |
| CN104134071B (en) | A kind of deformable part model object detecting method based on color description | |
| CN109447169A (en) | The training method of image processing method and its model, device and electronic system | |
| CN106446933B (en) | Multi-target detection method based on contextual information | |
| CN104182765B (en) | Internet image driven automatic selection method of optimal view of three-dimensional model | |
| CN109214420A (en) | The high texture image classification method and system of view-based access control model conspicuousness detection | |
| CN109635875A (en) | A kind of end-to-end network interface detection method based on deep learning | |
| CN107330892A (en) | A kind of sunflower disease recognition method based on random forest method | |
| CN105740915B (en) | A kind of collaboration dividing method merging perception information | |
| CN108537239A (en) | A kind of method of saliency target detection | |
| CN110334687A (en) | A kind of pedestrian retrieval Enhancement Method based on pedestrian detection, attribute study and pedestrian's identification | |
| CN106845513B (en) | Manpower detector and method based on condition random forest | |
| CN109740572A (en) | A kind of human face in-vivo detection method based on partial color textural characteristics | |
| CN109271932A (en) | Pedestrian based on color-match recognition methods again | |
| CN109492534A (en) | A kind of pedestrian detection method across scene multi-pose based on Faster RCNN | |
| CN104598889B (en) | The method and apparatus of Human bodys' response | |
| CN112001251A (en) | Pedestrian re-identification method and system based on combination of human body analysis and clothing color | |
| CN104966075B (en) | A kind of face identification method and system differentiating feature based on two dimension | |
| CN108846404A (en) | A kind of image significance detection method and device based on the sequence of related constraint figure | |
| CN108388905A (en) | A kind of Illuminant estimation method based on convolutional neural networks and neighbourhood context | |
| CN109377441A (en) | Tongue with privacy protection function is as acquisition method and system | |
| CN105825168A (en) | Golden snub-nosed monkey face detection and tracking algorithm based on S-TLD | |
| CN114863263B (en) | Snakehead fish detection method for blocking in class based on cross-scale hierarchical feature fusion | |
| CN108710916A (en) | The method and device of picture classification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170922 |
|
| WD01 | Invention patent application deemed withdrawn after publication |