CN114022747A - Salient object extraction method based on feature perception - Google Patents

Salient object extraction method based on feature perception Download PDF

Info

Publication number
CN114022747A
CN114022747A CN202210015109.7A CN202210015109A CN114022747A CN 114022747 A CN114022747 A CN 114022747A CN 202210015109 A CN202210015109 A CN 202210015109A CN 114022747 A CN114022747 A CN 114022747A
Authority
CN
China
Prior art keywords
feature
characteristic
channel
matrix
feature matrix
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
Application number
CN202210015109.7A
Other languages
Chinese (zh)
Other versions
CN114022747B (en
Inventor
左承林
何苗
易贤
熊浩
赵荣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Low Speed Aerodynamics Institute of China Aerodynamics Research and Development Center
Original Assignee
Low Speed Aerodynamics Institute of China Aerodynamics Research and Development Center
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Low Speed Aerodynamics Institute of China Aerodynamics Research and Development Center filed Critical Low Speed Aerodynamics Institute of China Aerodynamics Research and Development Center
Priority to CN202210015109.7A priority Critical patent/CN114022747B/en
Publication of CN114022747A publication Critical patent/CN114022747A/en
Application granted granted Critical
Publication of CN114022747B publication Critical patent/CN114022747B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a significant target extraction method based on feature perception. Processing the original picture through the characteristic MSCN factor, the characteristic image entropy, the characteristic dark channel, the H channel of the characteristic HSV channel and the S channel of the characteristic HSV channel to obtain 5 characteristic layers and a characteristic matrix
Figure 538296DEST_PATH_IMAGE001
(ii) a Feature matrix
Figure 596382DEST_PATH_IMAGE001
Carrying out down-sampling by 4 times to obtain a down-sampling feature matrix
Figure 907277DEST_PATH_IMAGE002
(ii) a Down-sampling feature matrix
Figure 709011DEST_PATH_IMAGE002
Normalization processing is carried out to obtain a normalized down-sampling feature matrix
Figure 54542DEST_PATH_IMAGE003
(ii) a To the normalized down-sampling feature matrix
Figure 283529DEST_PATH_IMAGE003
Extracting the single-feature significant target to obtain a feature matrix
Figure 753825DEST_PATH_IMAGE004
(ii) a Carrying out weight fusion on the single-feature significant target to obtain a feature matrix
Figure 483883DEST_PATH_IMAGE005
Feature matrix
Figure 559287DEST_PATH_IMAGE005
The position of the original image is a first obvious target; the method and the system fuse multi-source complementary information of different levels such as features, decisions and the like, and improve the accuracy of extracting the obvious target.

Description

Salient object extraction method based on feature perception
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to a significant target extraction method based on feature perception.
Background
After the wide and close attention of the machine vision discipline in the last 60 s, the technical staff is consistently dedicated to the project topic of researching the self-adaptive ability of the computer to the external environment like human. In order to enable a computer to perform high-level semantic analysis understanding, an object extraction link in an image needs to be solved first. Because object extraction becomes a fundamental problem of computer vision, great progress is made in the research result of the computable model based on the significance-based visual attention.
How to effectively capture salient objects of an image. The saliency calculation according to the calculated object may be divided into a saliency calculation based on a gaze point and a saliency region calculation. The salient region obtained by the salient calculation based on the gazing point is a small number of human eye attention points in the image, and further the salient object is a key problem to be solved in a salient object extraction algorithm adopting a salient map based on the gazing point. The salient region calculation can highlight the salient region in the image, so that the extraction efficiency of the salient object is greatly improved. However, the salient region is not considered with sufficient salient attributes, and is easy to generate highlight errors, and the effect of salient object extraction is greatly reduced if the highlight errors occur. Therefore, how to acquire a valid salient region is a key problem to be solved in a salient object extraction algorithm that adopts a salient map calculated based on the salient region.
Disclosure of Invention
In order to obtain an effective salient region, the invention provides a salient object extraction method based on feature perception, different features are used for fusion, different weights are given to different features, and the fusion is carried out from multi-source complementary information of different levels such as features, decisions and the like, so that the extraction accuracy of the detected salient object is improved.
The invention is realized by the following technical scheme:
the invention provides a significant target extraction method based on feature perception, which comprises the following steps:
s1, processing the original picture through the characteristic MSCN factor, the characteristic image entropy, the characteristic dark channel, the H channel of the characteristic HSV channel and the S channel of the characteristic HSV channel to obtain 5 characteristic layers and obtain a characteristic matrix
Figure 52466DEST_PATH_IMAGE001
(ii) a In the processing process, after an original picture is imported, 5 feature layers with the same size as the original picture are generated, wherein MSCN factors represent texture features, and a feature matrix is obtained by the feature layers
Figure 131280DEST_PATH_IMAGE002
The image entropy expresses detail features, and the feature layer obtains a feature matrix
Figure 894837DEST_PATH_IMAGE003
The dark channel feature layer obtains a feature matrix
Figure 604167DEST_PATH_IMAGE004
The dark channel can act when the color of the foreground and the background of the original picture are similar, the H channel of the HSV channel represents the color degree characteristic, and the characteristic layer obtains a characteristic matrix
Figure 85964DEST_PATH_IMAGE005
The S channel of the HSV channel represents the color tone characteristic, and the characteristic layer obtains a characteristic matrix
Figure 968469DEST_PATH_IMAGE006
；
S2: feature matrix
Figure 648849DEST_PATH_IMAGE007
Carrying out 4-time down-sampling to obtain a down-sampling feature matrix
Figure 856977DEST_PATH_IMAGE008
；
S3: down-sampling feature matrix
Figure 498174DEST_PATH_IMAGE008
Normalization processing is carried out to obtain a normalized down-sampling feature matrix
Figure 918791DEST_PATH_IMAGE009
；
S4, the normalized down-sampling feature matrix
Figure 656939DEST_PATH_IMAGE009
Extracting the single-feature significant target to obtain a feature matrix
Figure 832706DEST_PATH_IMAGE010
；
S5: carrying out weight fusion on the single-feature significant target to obtain a feature matrix
Figure 23516DEST_PATH_IMAGE011
Feature matrix
Figure 451086DEST_PATH_IMAGE012
The position of the original image is a first obvious target;
wherein:
Figure 778162DEST_PATH_IMAGE013
=1, 2, 3, 4, 5,1 denotes the characteristic MSCN factor, 2 denotes the characteristic image entropy, 3 denotes the characteristic dark channel, 4 denotes the H channel of the characteristic HSV channel, 5 denotes the S channel of the characteristic HSV channel.
Further, the feature matrix is down-sampled in step S2
Figure 328092DEST_PATH_IMAGE014
The obtaining method is as follows:
Figure 68515DEST_PATH_IMAGE015
；
wherein:
Figure 96514DEST_PATH_IMAGE016
the 4X4 block in the feature layer corresponds to the coordinates of the down-sampled point,
Figure 481359DEST_PATH_IMAGE017
representing the coordinates of a point in a 4X4 block in the downsampled feature layer, p being the number of rows in the 4X4 block, q being the number of columns in the 4X4 block,
Figure 936611DEST_PATH_IMAGE018
representing coordinates
Figure 102013DEST_PATH_IMAGE019
The characteristic value of the corresponding point is calculated,
Figure 7739DEST_PATH_IMAGE020
includes all that
Figure 43828DEST_PATH_IMAGE021
。
The feature matrix can be reduced by downsampling, and the function of improving the operation speed is achieved. The method is simple in obtaining mode and high in operation efficiency.
Further, the feature matrix is normalized in step S3
Figure 935560DEST_PATH_IMAGE022
The obtaining method is as follows:
Figure 525942DEST_PATH_IMAGE023
；
wherein:
Figure 630164DEST_PATH_IMAGE024
to represent
Figure 51918DEST_PATH_IMAGE008
The minimum value of the overall feature matrix is,
Figure 176869DEST_PATH_IMAGE025
to represent
Figure 51284DEST_PATH_IMAGE008
The maximum value of the overall feature matrix is,
Figure 224776DEST_PATH_IMAGE009
includes all that
Figure 438720DEST_PATH_IMAGE026
。
Normalization has the function of eliminating the influence of the abnormal value and improves the accuracy of feature extraction.
Further, the feature matrix in step S4
Figure 672255DEST_PATH_IMAGE027
The obtaining method is as follows:
(1) feature MSCN factor and feature image entropy single-feature salient object extraction
Setting a threshold value to extract a single-feature significant target:
Figure 96283DEST_PATH_IMAGE028
when is coming into contact with
Figure 807887DEST_PATH_IMAGE029
When it is, then
Figure 938654DEST_PATH_IMAGE016
The corresponding position of the point in the feature layer is a single-feature salient target; the value of the single-feature significant target is recorded as 1, and the value of the non-single-feature significant target is recorded as 0;
wherein
Figure 280774DEST_PATH_IMAGE030
，
Figure 129781DEST_PATH_IMAGE031
Represents a threshold value;
(2) single-feature significant target extraction of H channel and S channel of feature dark channel and feature HSV channel
Extracting the significant target by adopting a statistical histogram: for normalized down-sampling feature matrix
Figure 379497DEST_PATH_IMAGE032
Performing histogram statistics to obtain histogram distribution as follows:
Figure 692667DEST_PATH_IMAGE033
(ii) a For normalized down-sampling feature matrix
Figure 268005DEST_PATH_IMAGE034
Making histogram statistics on the peripheral edge to obtain the histogram distribution as follows:
Figure 338729DEST_PATH_IMAGE035
(ii) a Extracting the single-feature significant target according to the following formula:
Figure 595398DEST_PATH_IMAGE036
；
Figure 700757DEST_PATH_IMAGE037
；
Figure 181417DEST_PATH_IMAGE038
representing a whole/four-week processing of the histogram, the selected value being satisfied in the feature layer
Figure 801754DEST_PATH_IMAGE039
Or
Figure 924431DEST_PATH_IMAGE040
Is/are as follows
Figure 618717DEST_PATH_IMAGE041
The position corresponding to the point is a single-feature salient object, and the value of the single-feature salient object is recorded as
Figure 473541DEST_PATH_IMAGE026
The value of the non-single-feature salient object is recorded as 0;
Figure 784437DEST_PATH_IMAGE042
indicating that a value not present in the local histogram appears in the global statistical histogram,
Figure 179646DEST_PATH_IMAGE040
a segment of values representing an abnormal increase in the overall statistical histogram;
wherein:
Figure 56335DEST_PATH_IMAGE043
，
Figure 409956DEST_PATH_IMAGE044
a threshold value is indicated which is indicative of,
Figure 880251DEST_PATH_IMAGE045
are the coordinates of the down-sampled matrix and,
Figure 813572DEST_PATH_IMAGE046
=1、2、3、、、50，
Figure 748030DEST_PATH_IMAGE047
a representation value segment;
Figure 334870DEST_PATH_IMAGE048
includes all that
Figure 823620DEST_PATH_IMAGE049
。
Further, T1=0.4。
Further, the air conditioner is provided with a fan,
Figure 91790DEST_PATH_IMAGE050
；
wherein:
Figure 818438DEST_PATH_IMAGE051
to represent
Figure 982703DEST_PATH_IMAGE052
Is determined by the average value of (a) of (b),
Figure 489907DEST_PATH_IMAGE053
to represent
Figure 358506DEST_PATH_IMAGE052
Standard deviation of (2).
Further, step S5 feature matrix
Figure 736398DEST_PATH_IMAGE054
The obtaining method is as follows:
Figure 602723DEST_PATH_IMAGE055
；
wherein the content of the first and second substances,
Figure 269327DEST_PATH_IMAGE056
the weight is represented by a weight that is,
Figure 613721DEST_PATH_IMAGE054
includes all that
Figure 377278DEST_PATH_IMAGE026
；
Characteristic MSCN factor weight
Figure 211242DEST_PATH_IMAGE057
=1 weight of feature image entropy
Figure 427459DEST_PATH_IMAGE058
=1, weight of characteristic dark channel is
Figure 575544DEST_PATH_IMAGE059
=1.5, weight of H channel of characteristic HSV channel
Figure 131290DEST_PATH_IMAGE060
=1.5, S channel weight of characteristic HSV channel
Figure 808259DEST_PATH_IMAGE061
= 1;
feature matrix
Figure 511773DEST_PATH_IMAGE054
Includes all that
Figure 525865DEST_PATH_IMAGE026
，
Figure 732856DEST_PATH_IMAGE026
The specific calculation of (A) is as follows:
Figure 111884DEST_PATH_IMAGE062
；
when in use
Figure 240377DEST_PATH_IMAGE063
When it is, then
Figure 730265DEST_PATH_IMAGE016
The position of the salient object is a salient object I on the original picture;
different features are used for fusion, different weights are given to the different features, and fusion is performed from multi-source complementary information of different levels such as features and decisions, so that the detection accuracy is improved.
Further, the first salient target and the pixel block obtained by the super-pixel segmentation are fused to obtain a second salient target.
Further, the fusion method comprises the following steps:
Figure 57341DEST_PATH_IMAGE064
；
when in use
Figure 669588DEST_PATH_IMAGE065
When the value of the super-pixel is more than 40%, the total ratio of the pixel point of the first significant target in the single pixel block obtained by super-pixel segmentation to the pixel point in the single pixel block exceeds 40%, and the super-pixel block is a second significant target;
wherein the content of the first and second substances,
Figure 82115DEST_PATH_IMAGE066
a single block of pixels representing a super-pixel partition,
Figure 110113DEST_PATH_IMAGE067
representing the total number of pixels of the single pixel block,
Figure 494958DEST_PATH_IMAGE068
representing the total number of pixels in the single pixel block that contain the salient object one.
By adopting the technical scheme, the invention has the following advantages:
1. the method and the device have the advantages that different features are fused, different weights are given to the different features, and the accuracy of extracting the obvious target is improved by fusing multi-source complementary information of different levels such as features and decisions.
2. The method does not depend on single feature for feature extraction, depends on a plurality of features to jointly determine feature extraction, and has wide application range and high accuracy in the extraction of the significant target area.
3. The method can effectively extract the significant target under the condition that the color of the significant target area is similar to that of the background.
4. The invention has low operation complexity and can obtain results with less resources.
5. The invention utilizes the characteristic information to simulate the obvious target perceived by human eyes; meanwhile, the position information is utilized to simulate the obvious target perceived by human eyes; the characteristic information and the position information are combined to jointly determine a significant target, so that the method can adapt to more picture conditions; and has better accuracy and robustness.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments of the present invention or the prior art will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is an original picture;
FIG. 2 is a graph of a downsampled matrix;
FIG. 3 is a single-feature significant feature diagram obtained by performing single-feature extraction using a feature MSCN factor as an example;
FIG. 4 is an overall statistical histogram of an embodiment;
FIG. 5 is a statistical histogram of the peripheral edges in the example;
FIG. 6 is a graph showing the increase in the overall ratio to the peripheral edge in the example;
FIG. 7 is a single-feature saliency map taken for example for the S channel of a feature HSV channel;
FIG. 8 is a salient object map after fusion by weight;
FIG. 9 is a super-pixel segmented picture of an original picture;
FIG. 10 is a map of the salient objects after fusion;
fig. 11 is a diagram of a downsampling process.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments.
The embodiment provides a salient object extraction method based on feature perception, which comprises the following steps:
s1, processing the original picture through the characteristic MSCN factor, the characteristic image entropy, the characteristic dark channel, the H channel of the characteristic HSV channel and the S channel of the characteristic HSV channel to obtain 5 characteristic layers and obtain a characteristic matrix
Figure 215790DEST_PATH_IMAGE069
(ii) a In the processing process, as shown in fig. 1, after the original picture is imported, 5 feature layers with the same size as the original picture are generated, wherein the MSCN factor represents the texture feature, and the feature layers obtain a feature matrix
Figure 115613DEST_PATH_IMAGE070
The image entropy expresses detail features, and the feature layer obtains a feature matrix
Figure 9619DEST_PATH_IMAGE003
The dark channel feature layer obtains a feature matrix
Figure 311288DEST_PATH_IMAGE004
The dark channel can act when the color of the foreground and the background of the original picture are similar, the H channel of the HSV channel represents the color degree characteristic, and the characteristic layer obtains a characteristic matrix
Figure 937441DEST_PATH_IMAGE005
The S channel of the HSV channel represents the color tone characteristic, and the characteristic layer obtains a characteristic matrix
Figure 262243DEST_PATH_IMAGE071
(ii) a Feature matrix
Figure 897624DEST_PATH_IMAGE072
Including all positions
Figure 53799DEST_PATH_IMAGE073
Characteristic values of (1), in the same way
Figure 178749DEST_PATH_IMAGE003
、
Figure 53165DEST_PATH_IMAGE004
、
Figure 961078DEST_PATH_IMAGE005
、
Figure 175021DEST_PATH_IMAGE074
Also includes all positions
Figure 408557DEST_PATH_IMAGE075
The characteristic value of (2).
S2: respectively combining the feature matrices
Figure 504689DEST_PATH_IMAGE070
、
Figure 544189DEST_PATH_IMAGE003
、
Figure 940535DEST_PATH_IMAGE004
、
Figure 282655DEST_PATH_IMAGE005
、
Figure 600504DEST_PATH_IMAGE076
Carrying out 4-time down-sampling to obtain a down-sampling feature matrix
Figure 381378DEST_PATH_IMAGE077
、
Figure 366651DEST_PATH_IMAGE078
、
Figure 4306DEST_PATH_IMAGE079
、
Figure 75030DEST_PATH_IMAGE080
、
Figure 331699DEST_PATH_IMAGE081
；
The downsampling matrix is obtained by the following method:
Figure 171479DEST_PATH_IMAGE082
；
as shown in fig. 2, is a feature matrix
Figure 917718DEST_PATH_IMAGE083
As shown in fig. 11, first, an original picture with a size of M × N is input to obtain an MSCN feature matrix with a size of M × N
Figure 803635DEST_PATH_IMAGE084
Then is aligned with
Figure 395153DEST_PATH_IMAGE085
Performing a 4-fold downsampling to obtain a downsampled matrix of size M1N 1
Figure 620598DEST_PATH_IMAGE077
Wherein
Figure 475422DEST_PATH_IMAGE086
，
Figure 255159DEST_PATH_IMAGE087
。
The detailed process of down-sampling is to divide the feature matrix with the size of M × N into non-repetitive 4X4 blocks, and then represent the whole block by the mean value of the blocks, so as to obtain a down-sampling feature matrix with the size of M1 × N1. As shown in fig. 2, is any 4X4 block
Figure 915947DEST_PATH_IMAGE077
The obtaining method is as follows:
Figure 58216DEST_PATH_IMAGE088
=（
Figure 880678DEST_PATH_IMAGE089
+
Figure 413291DEST_PATH_IMAGE090
+
Figure 815453DEST_PATH_IMAGE091
+
Figure 484332DEST_PATH_IMAGE092
+
Figure 743275DEST_PATH_IMAGE093
+
Figure 825500DEST_PATH_IMAGE094
+
Figure 828091DEST_PATH_IMAGE095
+
Figure 617056DEST_PATH_IMAGE096
+
Figure 984583DEST_PATH_IMAGE097
+
Figure 226209DEST_PATH_IMAGE098
+
Figure 32491DEST_PATH_IMAGE099
+
Figure 738279DEST_PATH_IMAGE100
+
Figure 339024DEST_PATH_IMAGE101
+
Figure 67946DEST_PATH_IMAGE102
+
Figure 615602DEST_PATH_IMAGE103
+
Figure 848000DEST_PATH_IMAGE104
）/16；
Figure 885226DEST_PATH_IMAGE077
includes all that
Figure 163761DEST_PATH_IMAGE105
The same principle can be obtained
Figure 46266DEST_PATH_IMAGE078
、
Figure 664329DEST_PATH_IMAGE079
、
Figure 810140DEST_PATH_IMAGE080
、
Figure 248074DEST_PATH_IMAGE081
。
The feature matrix can be reduced by downsampling, and the function of improving the operation speed is achieved. The method is simple in obtaining mode and high in operation efficiency.
S3: respectively down-sampling feature matrices
Figure 199850DEST_PATH_IMAGE077
、
Figure 734736DEST_PATH_IMAGE078
、
Figure 848186DEST_PATH_IMAGE079
、
Figure 38996DEST_PATH_IMAGE080
、
Figure 466566DEST_PATH_IMAGE081
Normalization processing is carried out to obtain a normalized down-sampling feature matrix
Figure 793642DEST_PATH_IMAGE106
、
Figure 77993DEST_PATH_IMAGE107
、
Figure 818416DEST_PATH_IMAGE108
、
Figure 846415DEST_PATH_IMAGE109
、
Figure 293577DEST_PATH_IMAGE110
；
Downsampling feature matrices
Figure 952091DEST_PATH_IMAGE106
The calculation method is as follows:
Figure 851914DEST_PATH_IMAGE111
；
wherein:
Figure 418025DEST_PATH_IMAGE112
to represent
Figure 47589DEST_PATH_IMAGE105
The minimum value of (a) is determined,
Figure 939322DEST_PATH_IMAGE113
to represent
Figure 60861DEST_PATH_IMAGE105
The maximum value of (a) is,
Figure 633925DEST_PATH_IMAGE041
is the coordinates of the downsampled matrix;
feature matrix
Figure 790100DEST_PATH_IMAGE106
Includes all that
Figure 852734DEST_PATH_IMAGE114
The same can be obtained for the values of
Figure 789466DEST_PATH_IMAGE107
、
Figure 962958DEST_PATH_IMAGE108
、
Figure 973640DEST_PATH_IMAGE109
、
Figure 144858DEST_PATH_IMAGE110
。
Normalization has the function of eliminating the influence of the abnormal value and improves the accuracy of feature extraction.
S4, the normalized down-sampling feature matrix
Figure 506569DEST_PATH_IMAGE106
、
Figure 218173DEST_PATH_IMAGE107
、
Figure 411257DEST_PATH_IMAGE108
、
Figure 815694DEST_PATH_IMAGE109
、
Figure 399122DEST_PATH_IMAGE110
Extracting the single-feature significant target to obtain a feature matrix
Figure 117679DEST_PATH_IMAGE115
、
Figure 102953DEST_PATH_IMAGE116
、
Figure 412711DEST_PATH_IMAGE117
、
Figure 811332DEST_PATH_IMAGE118
、
Figure 130317DEST_PATH_IMAGE119
；
(1) Extracting the single-feature significant target of the feature MSCN factor to obtain a feature matrix
Figure 970097DEST_PATH_IMAGE115
As shown in fig. 3, the single-feature significant feature is obtained by performing single-feature extraction with the feature MSCN factor as an example;
setting a threshold value to extract a single-feature significant target:
Figure 654020DEST_PATH_IMAGE120
when is coming into contact with
Figure 477619DEST_PATH_IMAGE121
When, T1=0.4, then
Figure 69138DEST_PATH_IMAGE041
The corresponding position of the point in the feature layer is a single-feature salient target; the value of the single-feature significant target is recorded as 1, the rest of the single-feature significant targets are non-single-feature significant targets, and the value of the non-single-feature significant target is recorded as 0;
Figure 91320DEST_PATH_IMAGE115
the feature matrix for extracting the single-feature salient object of the feature image entropy can be obtained by the same method including all the values marked as 0 and 1
Figure 274040DEST_PATH_IMAGE116
；
(2) Single-feature significant target extraction of H channel and S channel of feature dark channel and feature HSV channel
Single feature saliency as shown in FIG. 7 is exemplified by the S channel of a feature HSV channel
Features, FIG. 4 is a feature matrix
Figure 53777DEST_PATH_IMAGE110
Value of (A)
Figure 652249DEST_PATH_IMAGE122
Histogram of ensemble statistics
Figure 732200DEST_PATH_IMAGE123
As shown in fig. 5, the feature matrix
Figure 554663DEST_PATH_IMAGE110
Value of (A)
Figure 149592DEST_PATH_IMAGE122
Histogram of the peripheral edge statistics
Figure 614071DEST_PATH_IMAGE123
(ii) a The value statistics are counted in value segments, because the normalized values range between 0 and 1]Inner, so according to the value segment [0,0.02]、[0.02、0.04]、[0.04、0.06]、[0.06、0.08]、[0.08、0.1]Make statistics, i.e.
Figure 282950DEST_PATH_IMAGE124
Representative value segment [0,0.02]，
Figure 479576DEST_PATH_IMAGE125
Representative value segments [0.02, 0.04 ]]，
Figure 499485DEST_PATH_IMAGE126
Representative value segment [0.04, 0.06 ]](ii) a Specifying values at the endpoints to be uniformly sorted into upper or lower values, e.g. 0.02 into value segments [0,0.02]Put 0.04 into the value section [0.02, 0.04 ]]The 0.06 is classified into value segments [0.04, 0.06 ]]Put 0.08 into the value section [0.08, 0.1 ]](ii) a Fig. 6 is a growth multiple obtained by performing the whole/peripheral edge processing on each value segment, that is, the number of values of each value segment in the whole statistical histogram is a multiple of the number of values of each value segment in the peripheral edge statistical histogram.
Extracting the S channel significant target of the single-feature HSV channel according to the following formula:
Figure 236497DEST_PATH_IMAGE127
；
Figure 87778DEST_PATH_IMAGE128
；
wherein:
Figure 783202DEST_PATH_IMAGE129
；
Figure 24827DEST_PATH_IMAGE130
to represent
Figure 503213DEST_PATH_IMAGE131
Is determined by the average value of (a) of (b),
Figure 412263DEST_PATH_IMAGE132
to represent
Figure 747429DEST_PATH_IMAGE131
Standard deviation of (2).
Figure 804247DEST_PATH_IMAGE133
Representing a whole/four-week processing of the histogram, the selected value being satisfied in the feature layer
Figure 148641DEST_PATH_IMAGE042
Or
Figure 646618DEST_PATH_IMAGE134
Is/are as follows
Figure 621528DEST_PATH_IMAGE041
The position corresponding to the point is a single-feature salient object,
Figure 837745DEST_PATH_IMAGE042
indicating that a value not present in the local histogram appears in the global statistical histogram,
Figure 720251DEST_PATH_IMAGE135
values representing anomalous increases in global statistical histogramsA segment; the value of the single-feature salient object is recorded as
Figure 400631DEST_PATH_IMAGE026
The value of the non-single feature salient object is noted as 0.
By the same token can obtain
Figure 608758DEST_PATH_IMAGE117
、
Figure 46693DEST_PATH_IMAGE118
。
S5: carrying out weight fusion on the single-feature significant target to obtain a feature matrix
Figure 670572DEST_PATH_IMAGE136
Feature matrix
Figure 408721DEST_PATH_IMAGE137
The position of the original image is a first obvious target; feature matrix
Figure 850067DEST_PATH_IMAGE138
The calculation formula is as follows:
Figure 775297DEST_PATH_IMAGE139
；
the characteristic MSCN factor is weighted by 1, i.e.
Figure 265184DEST_PATH_IMAGE057
=1, weight of feature image entropy is 1, i.e.
Figure 326681DEST_PATH_IMAGE058
=1, the weight of the characteristic dark channel is 1.5, i.e.
Figure 814294DEST_PATH_IMAGE059
=1.5, the weight of the H channel of the characteristic HSV channel is 1.5,
Figure 492400DEST_PATH_IMAGE060
=1.5, feature HThe S channel weight of the SV channel is 1, i.e.
Figure 520399DEST_PATH_IMAGE061
=1.5。
Feature matrix
Figure 41597DEST_PATH_IMAGE140
Includes all that
Figure 496849DEST_PATH_IMAGE141
The value of (a) is,
Figure 599934DEST_PATH_IMAGE141
the specific calculation of (A) is as follows:
Figure 166045DEST_PATH_IMAGE142
；
when in use
Figure 733292DEST_PATH_IMAGE143
When it is, then
Figure 93866DEST_PATH_IMAGE041
The position of the salient object is a salient object I on the original picture; fig. 8 is obtained by performing weight fusion on 5 features in this embodiment.
Different features are used for fusion, different weights are given to the different features, and fusion is performed from multi-source complementary information of different levels such as features and decisions, so that the detection accuracy is improved.
Further, fusing the salient object one shown in fig. 8 with the pixel block obtained by the super-pixel division shown in fig. 9 results in the salient object two shown in fig. 10.
Further, the fusion method comprises the following steps:
Figure 808882DEST_PATH_IMAGE144
；
when in use
Figure 444262DEST_PATH_IMAGE065
Is large in valueAnd when the sum of the pixel point of the first significant target in the single pixel block obtained by super-pixel segmentation and the pixel point in the single pixel block exceeds 40%, the single super-pixel block is the second significant target.
Wherein the content of the first and second substances,
Figure 538120DEST_PATH_IMAGE145
a single block of pixels representing a super-pixel partition,
Figure 335175DEST_PATH_IMAGE146
representing the total number of pixels of the single pixel block,
Figure 209590DEST_PATH_IMAGE147
representing the total number of pixel points of a first salient target contained in the single pixel block; note that a 4 × 4 block in the downsampled matrix represents a pixel, and as shown in fig. 9, the pixel block is obtained by superpixel division, and the pixel block calculates the pixel according to the same rule.
It should be noted that C in the present invention represents a matrix, i.e. a set of specific values,
Figure 710979DEST_PATH_IMAGE148
all represent specific numerical values; c comprises
Figure 721660DEST_PATH_IMAGE007
、
Figure 955195DEST_PATH_IMAGE008
、
Figure 254589DEST_PATH_IMAGE009
、
Figure 966193DEST_PATH_IMAGE027
、
Figure 96960DEST_PATH_IMAGE054
And the like,
Figure 563714DEST_PATH_IMAGE148
to representThe specific numerical values are, for example,
Figure 147142DEST_PATH_IMAGE148
Included
Figure 662437DEST_PATH_IMAGE149
、
Figure 850973DEST_PATH_IMAGE150
、
Figure 160731DEST_PATH_IMAGE151
、
Figure 497035DEST_PATH_IMAGE152
、
Figure 878338DEST_PATH_IMAGE141
and the like.
The first significant target and the second significant target are both significant targets extracted by a significant target extraction method based on feature perception, and the second significant target is obtained by fusing superpixel segmentation, so that the first significant target has higher accuracy than the first significant target.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principles of the present invention are intended to be included within the scope of the present invention.

Claims (9)

1. The method for extracting the salient object based on the feature perception is characterized by comprising the following steps of:
s1, processing the original picture through the characteristic MSCN factor, the characteristic image entropy, the characteristic dark channel, the H channel of the characteristic HSV channel and the S channel of the characteristic HSV channel to obtain 5 characteristic layers and obtain a characteristic matrix
Figure 821254DEST_PATH_IMAGE001
；
S2: feature matrix
Figure 215326DEST_PATH_IMAGE001
Carrying out 4-time down-sampling to obtain a down-sampling feature matrix
Figure 908476DEST_PATH_IMAGE002
；
S3: down-sampling feature matrix
Figure 704394DEST_PATH_IMAGE002
Normalization processing is carried out to obtain a normalized down-sampling feature matrix
Figure 926427DEST_PATH_IMAGE003
；
S4, the normalized down-sampling feature matrix
Figure 804866DEST_PATH_IMAGE003
Extracting the single-feature significant target to obtain a feature matrix
Figure 301707DEST_PATH_IMAGE004
；
S5: carrying out weight fusion on the single-feature significant target to obtain a feature matrix
Figure 952131DEST_PATH_IMAGE005
Feature matrix
Figure 876224DEST_PATH_IMAGE005
The position of the original image is a first obvious target;
wherein:
Figure 713730DEST_PATH_IMAGE006
=1, 2, 3, 4, 5,1 denotes the characteristic MSCN factor, 2 denotes the characteristic image entropy, 3 denotes the characteristic dark channel, 4 denotes the H channel of the characteristic HSV channel, 5 denotes the S channel of the characteristic HSV channel.
2. Feature perception based saliency as claimed in claim 1The target extraction method is characterized by comprising the following steps: feature matrix in step S2
Figure 748682DEST_PATH_IMAGE002
The obtaining method is as follows:
Figure 519192DEST_PATH_IMAGE007
；
wherein:
Figure 614187DEST_PATH_IMAGE008
the 4X4 block in the feature layer corresponds to the coordinates of the down-sampled point,
Figure 204569DEST_PATH_IMAGE009
representing the coordinates of a point in a 4X4 block in the downsampled feature layer, p being the number of rows in the 4X4 block, q being the number of columns in the 4X4 block,
Figure 777632DEST_PATH_IMAGE010
representing coordinates
Figure 402649DEST_PATH_IMAGE011
The characteristic value of the corresponding point is calculated,
Figure 668545DEST_PATH_IMAGE002
includes all that
Figure 746223DEST_PATH_IMAGE012
。
3. The feature perception-based salient object extraction method according to claim 2, wherein: feature matrix in step S3
Figure 122977DEST_PATH_IMAGE003
The obtaining method is as follows:
Figure 602500DEST_PATH_IMAGE013
；
wherein:
Figure 770789DEST_PATH_IMAGE014
to represent
Figure 335762DEST_PATH_IMAGE002
The minimum value of the overall feature matrix is,
Figure 516208DEST_PATH_IMAGE015
to represent
Figure 850237DEST_PATH_IMAGE002
The maximum value of the overall feature matrix is,
Figure 723515DEST_PATH_IMAGE005
includes all that
Figure 775785DEST_PATH_IMAGE016
。
4. The feature perception-based salient object extraction method according to claim 3, wherein: further, the feature matrix in step S4
Figure 759922DEST_PATH_IMAGE004
The obtaining method is as follows:
(1) extracting a single-feature salient object by using a feature MSCN factor and a feature image entropy:
setting a threshold value to extract a single-feature significant target:
Figure 948457DEST_PATH_IMAGE017
when is coming into contact with
Figure 461478DEST_PATH_IMAGE018
When it is, then
Figure 1044DEST_PATH_IMAGE008
The corresponding position of the point in the feature layer is a single-feature salient target;
wherein
Figure 523292DEST_PATH_IMAGE019
，
Figure 831914DEST_PATH_IMAGE020
Represents a threshold value;
(2) extracting single-feature significant targets of a feature dark channel, an H channel and a feature S channel of the feature HSV channel:
extracting the significant target by adopting a statistical histogram: for normalized down-sampling feature matrix
Figure 781415DEST_PATH_IMAGE003
And carrying out integral statistics to obtain a histogram as follows:
Figure 808277DEST_PATH_IMAGE021
(ii) a For normalized down-sampling feature matrix
Figure 868637DEST_PATH_IMAGE003
Making histogram statistics on the peripheral edge to obtain a histogram as follows:
Figure 28835DEST_PATH_IMAGE022
(ii) a Extracting the single-feature significant target according to the following formula:
Figure 414817DEST_PATH_IMAGE023
；
Figure 663396DEST_PATH_IMAGE024
；
wherein:
Figure 527447DEST_PATH_IMAGE025
，
Figure 810661DEST_PATH_IMAGE026
a threshold value is indicated which is indicative of,
Figure 101965DEST_PATH_IMAGE027
are the coordinates of the down-sampled matrix and,
Figure 837840DEST_PATH_IMAGE028
=1、2、3、、、50，
Figure 240002DEST_PATH_IMAGE029
the value section is represented by a value section,
Figure 377722DEST_PATH_IMAGE004
includes all that
Figure 839928DEST_PATH_IMAGE030
。
5. The feature perception-based salient object extraction method according to claim 4, wherein: t is1=0.4。
6. The feature perception-based salient object extraction method according to claim 4, wherein:
Figure 63099DEST_PATH_IMAGE031
；
wherein:
Figure 268952DEST_PATH_IMAGE032
to represent
Figure 261179DEST_PATH_IMAGE033
Is determined by the average value of (a) of (b),
Figure 159865DEST_PATH_IMAGE034
to represent
Figure 870332DEST_PATH_IMAGE033
Standard deviation of (2).
7. The feature perception-based salient object extraction method according to claim 4, wherein: step S5 feature matrix
Figure 614297DEST_PATH_IMAGE005
The obtaining method comprises the following steps:
Figure 723680DEST_PATH_IMAGE035
；
when in use
Figure 527688DEST_PATH_IMAGE036
When it is, then
Figure 725451DEST_PATH_IMAGE008
The position of the salient object is a salient object I on the original picture;
wherein the content of the first and second substances,
Figure 273107DEST_PATH_IMAGE037
the weight is represented by a weight that is,
Figure 239926DEST_PATH_IMAGE005
includes all that
Figure 886939DEST_PATH_IMAGE038
。
8. The feature perception based salient object extraction method of any one of claims 1-7, wherein: and fusing the first significant target and the pixel block obtained by the super-pixel segmentation to obtain a second significant target.
9. Feature-based awareness as claimed in claim 8The significant object extraction method is characterized in that: the fusion method comprises the following steps:
Figure 306419DEST_PATH_IMAGE039
；
when in use
Figure 392187DEST_PATH_IMAGE040
Is greater than 40%, the pixel block is a significant target two;
wherein the content of the first and second substances,
Figure 213512DEST_PATH_IMAGE041
a single block of pixels representing a super-pixel partition,
Figure 359323DEST_PATH_IMAGE042
representing the total number of pixels of the single pixel block,
Figure 266099DEST_PATH_IMAGE043
representing the total number of pixels in the single pixel block that contain the salient object one.
CN202210015109.7A 2022-01-07 2022-01-07 Salient object extraction method based on feature perception Active CN114022747B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210015109.7A CN114022747B (en) 2022-01-07 2022-01-07 Salient object extraction method based on feature perception

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210015109.7A CN114022747B (en) 2022-01-07 2022-01-07 Salient object extraction method based on feature perception

Publications (2)

Publication Number Publication Date
CN114022747A true CN114022747A (en) 2022-02-08
CN114022747B CN114022747B (en) 2022-03-15

Family

ID=80069837

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210015109.7A Active CN114022747B (en) 2022-01-07 2022-01-07 Salient object extraction method based on feature perception

Country Status (1)

Country Link
CN (1) CN114022747B (en)

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101303733A (en) * 2008-05-26 2008-11-12 东华大学 Method for viewing natural color at night with sense of space adopting pattern database
CN105678735A (en) * 2015-10-13 2016-06-15 中国人民解放军陆军军官学院 Target salience detection method for fog images
CN105825238A (en) * 2016-03-30 2016-08-03 江苏大学 Visual saliency object detection method
CN106127756A (en) * 2016-06-21 2016-11-16 西安工程大学 A kind of insulator recognition detection method based on multicharacteristic information integration technology
CN107730472A (en) * 2017-11-03 2018-02-23 昆明理工大学 A kind of image defogging optimized algorithm based on dark primary priori
CN109410171A (en) * 2018-09-14 2019-03-01 安徽三联学院 A kind of target conspicuousness detection method for rainy day image
CN110008969A (en) * 2019-04-15 2019-07-12 京东方科技集团股份有限公司 The detection method and device in saliency region
CN110675351A (en) * 2019-09-30 2020-01-10 集美大学 Marine image processing method based on global brightness adaptive equalization
CN111091129A (en) * 2019-12-24 2020-05-01 沈阳建筑大学 Image salient region extraction method based on multi-color characteristic manifold sorting
CN111310774A (en) * 2020-04-01 2020-06-19 江苏商贸职业学院 PM2.5 concentration measurement method based on image quality
CN111915592A (en) * 2020-08-04 2020-11-10 西安电子科技大学 Remote sensing image cloud detection method based on deep learning

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101303733A (en) * 2008-05-26 2008-11-12 东华大学 Method for viewing natural color at night with sense of space adopting pattern database
CN105678735A (en) * 2015-10-13 2016-06-15 中国人民解放军陆军军官学院 Target salience detection method for fog images
CN105825238A (en) * 2016-03-30 2016-08-03 江苏大学 Visual saliency object detection method
CN106127756A (en) * 2016-06-21 2016-11-16 西安工程大学 A kind of insulator recognition detection method based on multicharacteristic information integration technology
CN107730472A (en) * 2017-11-03 2018-02-23 昆明理工大学 A kind of image defogging optimized algorithm based on dark primary priori
CN109410171A (en) * 2018-09-14 2019-03-01 安徽三联学院 A kind of target conspicuousness detection method for rainy day image
CN110008969A (en) * 2019-04-15 2019-07-12 京东方科技集团股份有限公司 The detection method and device in saliency region
CN110675351A (en) * 2019-09-30 2020-01-10 集美大学 Marine image processing method based on global brightness adaptive equalization
CN111091129A (en) * 2019-12-24 2020-05-01 沈阳建筑大学 Image salient region extraction method based on multi-color characteristic manifold sorting
CN111310774A (en) * 2020-04-01 2020-06-19 江苏商贸职业学院 PM2.5 concentration measurement method based on image quality
CN111915592A (en) * 2020-08-04 2020-11-10 西安电子科技大学 Remote sensing image cloud detection method based on deep learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LING Z等: "Optimal transmission estimation via fog density perception for efficient single image defogging", 《IEEE TRANSACTIONS ON MULTIMEDIA》 *
刘坤等: "融合深度信息的雾霾情况下显著性目标提取", 《河北工业大学学报》 *

Also Published As

Publication number Publication date
CN114022747B (en) 2022-03-15

Similar Documents

Publication Publication Date Title
CN107622258B (en) Rapid pedestrian detection method combining static underlying characteristics and motion information
US20200410212A1 (en) Fast side-face interference resistant face detection method
WO2021254205A1 (en) Target detection method and apparatus
CN107133943A (en) A kind of visible detection method of stockbridge damper defects detection
CN108921820B (en) Saliency target detection method based on color features and clustering algorithm
CN111062278B (en) Abnormal behavior identification method based on improved residual error network
CN111524145A (en) Intelligent picture clipping method and system, computer equipment and storage medium
CN109544564A (en) A kind of medical image segmentation method
CN112529090B (en) Small target detection method based on improved YOLOv3
CN111626342B (en) Image sample processing method, device and storage medium
CN110852327A (en) Image processing method, image processing device, electronic equipment and storage medium
CN108776823A (en) Cervical carcinoma lesion analysis method based on cell image recognition
CN113850136A (en) Yolov5 and BCNN-based vehicle orientation identification method and system
CN111062854A (en) Method, device, terminal and storage medium for detecting watermark
CN114220126A (en) Target detection system and acquisition method
CN111553337A (en) Hyperspectral multi-target detection method based on improved anchor frame
CN112926667B (en) Method and device for detecting saliency target of depth fusion edge and high-level feature
Zeeshan et al. A newly developed ground truth dataset for visual saliency in videos
CN113870196A (en) Image processing method, device, equipment and medium based on anchor point cutting graph
CN111597845A (en) Two-dimensional code detection method, device and equipment and readable storage medium
CN114022747B (en) Salient object extraction method based on feature perception
JP4967045B2 (en) Background discriminating apparatus, method and program
CN107368847A (en) A kind of crop leaf diseases recognition methods and system
CN115775226B (en) Medical image classification method based on transducer
CN113705640B (en) Method for quickly constructing airplane detection data set based on remote sensing image

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