CN113723500B - Image data expansion method based on combination of feature similarity and linear smoothing - Google Patents
Image data expansion method based on combination of feature similarity and linear smoothing Download PDFInfo
- Publication number
- CN113723500B CN113723500B CN202110994684.1A CN202110994684A CN113723500B CN 113723500 B CN113723500 B CN 113723500B CN 202110994684 A CN202110994684 A CN 202110994684A CN 113723500 B CN113723500 B CN 113723500B
- Authority
- CN
- China
- Prior art keywords
- image
- target
- pixel
- point
- pixel point
- 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.)
- Active
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention discloses an image data expansion method based on combination of feature similarity and linear smoothness, which comprises the steps of determining expansion requirements according to existing images, performing feature analysis, and collecting targets and images to be combined; fusion of the image and the target is accomplished by linear smoothing: will annotate the target G i The center point of the (B) is coincident with a point of a target appearance area in the scene image, and the pixel value of the boundary of the annular smooth area and the innermost layer is set as A 1 Pixel value a up to the outermost boundary n Pixel point P connected with outer boundary j And an inner pixel point P i Forming a smooth line, setting the distance from each pixel point to the line segment vertical projection point as A 1 ~A n Calculating a pixel value A of each pixel point i The method comprises the steps of carrying out a first treatment on the surface of the And outputting the combined image and forming a new data set with the original small batch of data. According to the invention, the target and the image are combined into the needed data in a linear smoothing mode, enough data samples can be obtained only by paying a small amount of cost, and the training effect of the deep learning model is improved.
Description
Technical Field
The invention relates to the technical field of deep learning and image processing, in particular to an image data expansion method based on combination of feature similarity and linear smoothing.
Background
With the rise of artificial intelligence, the application field of deep learning is becoming wider and wider. A major difficulty of deep learning applications is that the processing of the data source and the raw data requires the preparation of large volumes of data for deep learning training, and the data is preferably relatively comprehensive, i.e. the data set needs to contain as much as possible of all the situations where the detected object is present, in order to enrich the learning process. However, due to the limitations of many external factors, some target scenes such as defect detection are high in construction cost, damage to objects is caused or the objects are not easy to realize, so that the amount of data which can be finally obtained is far insufficient to support deep learning model training, and engineers often face the problems of insufficient data amount and repeated data processing machines.
Disclosure of Invention
The invention aims to provide an image data expansion method based on combination of feature similarity and linear smoothing, which is used for solving the problems that in the prior art, deep learning model training cannot be supported due to insufficient data sets and data processing machinery is repeated.
The invention solves the problems by the following technical proposal:
an image data expansion method based on a combination of feature similarity and linear smoothing, comprising:
step S100, determining expansion requirements according to the existing images, performing feature analysis, and collecting targets and images to be combined;
step S200, a region substitution rule is adopted to realize the combination of the target and the image, and the fusion of the image and the target is completed through linear smoothing, and the method specifically comprises the following steps:
step S210, randomly selecting one point in a possible area where the target appears in the scene image as the center, and labeling one of the targets G in the target set i The center point of the (B) is overlapped with the point, the marked edge L is taken as a central line, and the annular areas with the width of H/2 are respectively extended inwards and outwards simultaneously to be smooth areas; let the pixel value of the innermost border be A 1 The pixel values which are sequentially increased to the outermost boundary from outside are A n The formula for calculating the distance from any pixel point on the outermost boundary to the pixel point of the innermost boundary having the smallest pixel distance from the pixel point is as follows:
wherein P is j P is any pixel point of the outermost boundary i Is equal to the P j An inner pixel point having a minimum pixel distance; m is the maximum value of i;
the distance from all the pixel points on the outermost boundary to the nearest innermost boundary pixel point can be obtained by the above formula;
step S220, in the local small region, each pixel point can be regarded as a region with a unique small block value, and is connected with P j (x j ,y j )、P i (x i ,y i ) A line segment, wherein each pixel point passing through the annular smoothing region forms a smoothing line, and the sequence of each pixel point is from each pixel point to a vertical projection point P of a straight line where the line segment is located 0 (x 0 ,y 0 ) To P j Distance determination of (2):
step S230, setting the distance to be A from far to near 1 ~A n Pixel value a of each pixel point i The calculation formula is as follows:
step S240, because the number of the pixels of the outer boundary is greater than that of the pixels of the inner boundary, the final pixel value of each pixel is the average value of the mapping calculation values;
step S300: and comparing pixel data according to the calculated average value, expanding the image, and outputting the combined image to form a new data set with the original small batch data.
The expanding requirement includes combining the objects according to an existing object context and combining the objects according to an existing context, wherein,
the feature analysis corresponding to the existing target combination background is to carry out target labeling on the existing image, and a corresponding target area and a label file are generated; simultaneously analyzing background scene characteristics, obtaining images with similarity exceeding a first threshold, and outputting labels and images;
feature analysis of existing background combined targets defines background areas for determining possible areas of targets in existing images; and meanwhile, analyzing target characteristics, obtaining targets with feature similarity exceeding a second threshold, and outputting a background and the targets.
And (3) performing one or more of scale transformation, rotation, translation, filtering or noise reduction on the new data set obtained in the step (S300) to perform secondary expansion of the image data.
Compared with the prior art, the invention has the following advantages:
(1) According to the method, the existing small-batch image data are subjected to labeling analysis, the label frame target image and the image basic characteristics are disassembled, then the class target or class scene image with higher similarity is searched according to the basic characteristics, and the target area and the class image are combined into new image data in a linear smooth combination mode, so that the image sample size is enriched, the training effect of the deep learning model is enhanced, and the scene construction cost is not increased additionally.
(2) The invention increases the sample size of the deep learning data set, realizes the labeling of the targets and the selection of the class targets and class scene images with high similarity based on small batches of existing data, combines the targets and the images into the data needed by us in a linear smoothing mode, can obtain enough data samples only by paying a small amount of cost, and improves the training effect of the deep learning model.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a flow chart of a method for determining a combination of an object and an image according to the present invention, wherein (a) is a feature analysis process of an existing object combination background; (b) a characteristic analysis flow of the existing background combined target;
FIG. 3 is a schematic view of a linear smoothing combined region of an object and an image in the present invention.
Detailed Description
The present invention will be described in further detail with reference to examples, but embodiments of the present invention are not limited thereto.
Examples:
referring to fig. 1-3, an image data expansion method based on combination of feature similarity and linear smoothing can be applied to deep learning model training data set expansion, and comprises the following steps:
step S01: the existing image data is acquired, and basic features of targets in the image are analyzed based on human vision and logic discriminant. Specifically, the existing image data is obtained and includes self image data or video samples converted into frame image data; the basic features comprise basic forms of the targets, categories to which the targets belong and logical scenes in which the targets may exist;
step S02: the combination type is selected, and the combination type specifically comprises two modes of an existing target combination background and an existing background combination target.
Existing target combination background: labeling (i.e. labeling) the existing image to generate a corresponding target area and a label file; while analyzing the background situation that the target may appear, the background image and the target area output are collected, as shown in fig. 2 (a).
Existing background combined targets: analyzing the existing image, determining a possible target existence area in the image, and defining a background area; and meanwhile, analyzing the characteristics of the target, and acquiring the target with high characteristic similarity and high coincidence degree, and outputting the target and the background together, as shown in (b) of fig. 2.
Step S03: according to the target data and the target scene data collected in batches in the step S02, specifically, the target scene data comprises images of possible scenes of the target and the existing background similar scenes, and the sizes of the images are unified;
step S04: one-to-one linear smoothing of the selected image and the target is combined as shown in fig. 3:
assume that the set of labeled targets is { G i Randomly selecting one point in a possible area where a target appears in a scene image as a center, and labeling the target G i The center point of (2) coincides with the point, the annular region with the width H/2 is extended inwards and outwards by taking the marked edge L as a central line and is taken as a smooth region, and the pixel value of the innermost boundary is taken as A 1 Sequentially increasing to the outermost boundary A n For each pixel point P of the outer boundary j Calculating to obtain an inner pixel point P with the minimum pixel distance i ,P j 、P i The following should be satisfied:
in the local small area, each pixel point can be regarded as a small block value unique area and is connected with P j (x j ,y j )、P i (x i ,y i ) The line segment, the pixel point passing through in the annular smooth area is the smooth line, and the points are sequentially projected from each pixel point to the vertical projection point P of the straight line 0 (x 0 ,y 0 ) To P j Distance determination of (2):
respectively set as A according to distance from far to near 1 ~A n Pixel value A of each pixel point i The calculation formula is as follows:
because the number of the outer side pixel points is larger than that of the inner side pixel points, a single pixel point is selected when a plurality of straight lines are mapped correspondingly, and an optional solution is that the pixel value of the final point is determined by the average value of the values calculated by each mapping;
step S05: outputting the combined image, forming a new data set with the original small batch data, and performing traditional image enhancement operations such as traditional scale conversion, rotation, translation, filtering, noise reduction and the like after the optional new image is generated
Although the invention has been described herein with reference to the above-described illustrative embodiments thereof, the above-described embodiments are merely preferred embodiments of the present invention, and the embodiments of the present invention are not limited by the above-described embodiments, it should be understood that numerous other modifications and embodiments can be devised by those skilled in the art that will fall within the scope and spirit of the principles of this disclosure.
Claims (3)
1. An image data expansion method based on a combination of feature similarity and linear smoothing, comprising:
step S100, determining expansion requirements according to the existing images, performing feature analysis, and collecting targets and images to be combined;
step S200, a region substitution rule is adopted to realize the combination of the target and the image, and the fusion of the image and the target is completed through linear smoothing, and the method specifically comprises the following steps:
step S210, randomly selecting one point in a possible area where the target appears in the scene image as the center, and labeling one of the targets G in the target set i The center point of the (B) is overlapped with the point, the marked edge L is taken as a central line, and the annular areas with the width of H/2 are respectively extended inwards and outwards simultaneously to be smooth areas; let the pixel value of the innermost border be A 1 The pixel values which are sequentially increased to the outermost boundary from outside are A n For each pixel point P of the outermost boundary j An inner pixel point P having a minimum pixel distance from the pixel point P i The method comprises the following steps:
wherein m is the maximum value of i;
step S220, in the local small area, each pixel point is regarded as a small block value unique area, and P is connected j (x j ,y j )、P i (x i ,y i ) Line segment, each pixel point group penetrated in annular smooth areaForming a smooth line, and sequentially projecting each pixel point to the vertical projection point P of the straight line where the line segment is positioned from each pixel point 0 (x 0 ,y 0 ) To P j Distance determination of (2):
step S230, setting the distance to be A from far to near 1 ~A n Pixel value a of each pixel point i The calculation formula is as follows:
step S240, because the number of the pixels of the outer boundary is greater than that of the pixels of the inner boundary, the final pixel value of each pixel is the average value of the calculated values;
step S300: and comparing pixel data according to the calculated average value, expanding the image, and outputting the combined image to form a new data set with the original small batch data.
2. The method for expanding image data based on combination of feature similarity and linear smoothing as claimed in claim 1, wherein the expanding requirement includes combining the background according to an existing object and combining the object according to an existing background, wherein,
the feature analysis corresponding to the existing target combination background is to carry out target labeling on the existing image, and a corresponding target area and a label file are generated; simultaneously analyzing background scene characteristics, obtaining images with similarity exceeding a first threshold, and outputting labels and images;
feature analysis of existing background combined targets defines background areas for determining possible areas of targets in existing images; and meanwhile, analyzing target characteristics, obtaining targets with feature similarity exceeding a second threshold, and outputting a background and the targets.
3. The image data expansion method based on the combination of feature similarity and linear smoothing as claimed in claim 1, wherein the new data set obtained in step S300 is subjected to one or more of scaling, rotation, translation, filtering or noise reduction to perform secondary expansion of the image data.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110994684.1A CN113723500B (en) | 2021-08-27 | 2021-08-27 | Image data expansion method based on combination of feature similarity and linear smoothing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110994684.1A CN113723500B (en) | 2021-08-27 | 2021-08-27 | Image data expansion method based on combination of feature similarity and linear smoothing |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113723500A CN113723500A (en) | 2021-11-30 |
CN113723500B true CN113723500B (en) | 2023-06-16 |
Family
ID=78678410
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110994684.1A Active CN113723500B (en) | 2021-08-27 | 2021-08-27 | Image data expansion method based on combination of feature similarity and linear smoothing |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113723500B (en) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102314599A (en) * | 2011-10-11 | 2012-01-11 | 东华大学 | Identification and deviation-detection method for lane |
CN102831472A (en) * | 2012-08-03 | 2012-12-19 | 无锡慧眼电子科技有限公司 | People counting method based on video flowing image processing |
CN104217208A (en) * | 2013-06-03 | 2014-12-17 | 株式会社理光 | Target detection method and device |
CN106600592A (en) * | 2016-12-14 | 2017-04-26 | 中南大学 | Track long chord measurement method based on the splicing of continuous frame images |
CN106910159A (en) * | 2016-10-12 | 2017-06-30 | 王玲 | Video-splicing method and device |
CN107622504A (en) * | 2017-09-30 | 2018-01-23 | 北京百度网讯科技有限公司 | Method and apparatus for handling picture |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8422826B2 (en) * | 2009-06-05 | 2013-04-16 | Varian Medical Systems, Inc. | Method and apparatus to facilitate using fused images to identify materials |
CN104869389B (en) * | 2015-05-15 | 2016-10-05 | 北京邮电大学 | Off-axis formula virtual video camera parameter determination method and system |
US10186023B2 (en) * | 2016-01-25 | 2019-01-22 | Qualcomm Incorporated | Unified multi-image fusion approach |
-
2021
- 2021-08-27 CN CN202110994684.1A patent/CN113723500B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102314599A (en) * | 2011-10-11 | 2012-01-11 | 东华大学 | Identification and deviation-detection method for lane |
CN102831472A (en) * | 2012-08-03 | 2012-12-19 | 无锡慧眼电子科技有限公司 | People counting method based on video flowing image processing |
CN104217208A (en) * | 2013-06-03 | 2014-12-17 | 株式会社理光 | Target detection method and device |
CN106910159A (en) * | 2016-10-12 | 2017-06-30 | 王玲 | Video-splicing method and device |
CN106600592A (en) * | 2016-12-14 | 2017-04-26 | 中南大学 | Track long chord measurement method based on the splicing of continuous frame images |
CN107622504A (en) * | 2017-09-30 | 2018-01-23 | 北京百度网讯科技有限公司 | Method and apparatus for handling picture |
Non-Patent Citations (3)
Title |
---|
"基于FPGA的小型化图像采集和预处理系统设计";王子;《中国优秀硕士学位论文全文数据库 信息科技辑》(第2021 年 第05期期);第I135-396页 * |
P. R. Mirajkar."Acute ischemic stroke detection using wavelet based fusion of CT and MRI images".《2015 International Conference on Advances in Computing, Communications and Informatics (ICACCI)》.2015,第1123-1130页. * |
郭盼等."基于导向滤波器的医学图像融合方法".《液晶与显示 》.2019,第34卷(第6期),第605-612页. * |
Also Published As
Publication number | Publication date |
---|---|
CN113723500A (en) | 2021-11-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110428432B (en) | Deep neural network algorithm for automatically segmenting colon gland image | |
CN108550133B (en) | Cancer cell detection method based on fast R-CNN | |
CN110751655B (en) | Automatic cutout method based on semantic segmentation and significance analysis | |
CN109410168B (en) | Modeling method of convolutional neural network for determining sub-tile classes in an image | |
Guan et al. | Lip image segmentation based on a fuzzy convolutional neural network | |
CN106529573A (en) | Real-time object detection method based on combination of three-dimensional point cloud segmentation and local feature matching | |
CN107369158B (en) | Indoor scene layout estimation and target area extraction method based on RGB-D image | |
JP2013536960A (en) | System and method for synthesizing portrait sketches from photographs | |
Gupta et al. | Automatic trimap generation for image matting | |
WO2019197021A1 (en) | Device and method for instance-level segmentation of an image | |
CN104021566A (en) | GrabCut algorithm-based automatic segmentation method of tongue diagnosis image | |
CN112330561B (en) | Medical image segmentation method based on interactive foreground extraction and information entropy watershed | |
CN110827304A (en) | Traditional Chinese medicine tongue image positioning method and system based on deep convolutional network and level set method | |
CN111009005A (en) | Scene classification point cloud rough registration method combining geometric information and photometric information | |
CN114663371A (en) | Image salient target detection method based on modal unique and common feature extraction | |
CN111161213B (en) | Industrial product defect image classification method based on knowledge graph | |
Yuan et al. | Superpixels with content-adaptive criteria | |
JP2012043436A (en) | Apparatus and method for object detection, and apparatus and method for creating sorter | |
CN107644427B (en) | Multi-modal image segmentation method based on MSEW-CA combined gray histogram | |
CN113723500B (en) | Image data expansion method based on combination of feature similarity and linear smoothing | |
CN111932645B (en) | Method for automatically generating ink and wash painting based on generation countermeasure network GAN | |
JP4285640B2 (en) | Object identification method, apparatus and program | |
CN109377455B (en) | Improved multi-sequence magnetic resonance image registration method based on self-similarity | |
JP2009123234A (en) | Object identification method, apparatus and program | |
Huang et al. | Segdiscover: Visual concept discovery via unsupervised semantic segmentation |
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 |