CN111783885A - Millimeter wave image quality classification model construction method based on local enhancement - Google Patents
Millimeter wave image quality classification model construction method based on local enhancement Download PDFInfo
- Publication number
- CN111783885A CN111783885A CN202010624153.9A CN202010624153A CN111783885A CN 111783885 A CN111783885 A CN 111783885A CN 202010624153 A CN202010624153 A CN 202010624153A CN 111783885 A CN111783885 A CN 111783885A
- Authority
- CN
- China
- Prior art keywords
- image
- millimeter wave
- gray value
- gradient
- classification model
- 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
- 238000013145 classification model Methods 0.000 title claims abstract description 26
- 238000010276 construction Methods 0.000 title claims abstract description 8
- 238000000034 method Methods 0.000 claims abstract description 55
- 239000013598 vector Substances 0.000 claims abstract description 37
- 239000011159 matrix material Substances 0.000 claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000012545 processing Methods 0.000 claims description 29
- 238000010606 normalization Methods 0.000 claims description 11
- 238000013507 mapping Methods 0.000 claims description 8
- 230000008859 change Effects 0.000 claims description 6
- 238000003672 processing method Methods 0.000 claims description 4
- 238000012937 correction Methods 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 3
- 230000008569 process Effects 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 21
- 238000012706 support-vector machine Methods 0.000 description 12
- 238000000605 extraction Methods 0.000 description 10
- 238000012360 testing method Methods 0.000 description 9
- 238000013135 deep learning Methods 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 4
- 238000001514 detection method Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000003384 imaging method Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000007635 classification algorithm Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/40—Image enhancement or restoration by the use of histogram techniques
-
- G06T5/90—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/40—Analysis of texture
- G06T7/41—Analysis of texture based on statistical description of texture
- G06T7/45—Analysis of texture based on statistical description of texture using co-occurrence matrix computation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30168—Image quality inspection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a millimeter wave image quality classification model construction method based on local enhancement, which can improve the definition of the whole image after the local enhancement of an input image, and has larger distinction degree with a fuzzy image, and the definition of the fuzzy image after the enhancement can not be obviously changed; and constructing and extracting effective characteristic vectors aiming at the enhanced images, and sending the effective characteristic vectors to a classifier for training. The invention has the advantages that: the contrast ratio of a clear image and a blurred image is improved through local enhancement, the difference between the two images is increased, the feature vector is constructed through the histogram feature of the directional gradient and the gray level co-occurrence matrix feature, the information of the millimeter wave image can be better described, therefore, after the images are classified, the difference of the images of different classes is as large as possible, the kernel function of the SVM classifier is constructed based on the kernel function form of the radial basis function, the relation between the feature vector and the class label is more flexible to process, and the application is wide.
Description
Technical Field
The invention relates to the technical field of millimeter wave imaging, in particular to a quality classification model construction method of a millimeter wave image based on local enhancement.
Background
The definition of the millimeter wave image is an important index for measuring the quality of the millimeter wave image, and can be used for guiding the construction and adjustment of a millimeter wave image processing system, optimizing algorithm design and parameter selection. Millimeter wave imaging technology is an important research direction for dangerous goods detection under hidden clothes, wherein the quality of millimeter wave images affects whether dangerous goods targets in the images can be successfully detected. Therefore, filtering out images with poor image quality and obtaining clear images for detection and recognition have attracted attention of researchers in recent years. The basic idea of image quality classification is to firstly perform feature description on training images, apply a relevant method of machine learning, combine prior knowledge (such as image category labels), construct a classification model of learning training, and finally classify test images. The following image quality classification methods are mainly included at present: one is the traditional image classification based on feature extraction, and the method is used for separately carrying out feature extraction and classifier design; the other type is image classification based on deep learning, and the method is an end-to-end learning method, namely, feature extraction and classifier design are learned together.
(1) Image classification based on feature extraction
The traditional image classification method comprises two aspects of feature extraction and classifier design, wherein the feature extraction is the main research direction. Because only by using proper feature vectors, the classifier can achieve better classification effect on the classification. The specific flow chart is shown in the attached figure 1, namely after an image is input, the characteristics of the image are extracted by adopting a designed algorithm, the characteristic vector is used as the input of a classifier, and then the prediction analysis is carried out on the picture to be detected.
Common image classification algorithms comprise a nearest neighbor algorithm (KNN), a Support Vector Machine (SVM) and the like, and can achieve high recognition accuracy in practical application. However, with the image classification algorithm based on the SVM, it is necessary to extract more effective features and construct feature descriptors. Although the descriptors can well describe the image information, the single descriptors belong to local descriptors, so that the descriptors are easily influenced by noise and are difficult to be directly applied to an image classification task.
(2) Image classification based on deep learning
Deep Learning (Deep Learning) is an important branch of machine Learning, i.e., an algorithm that tries to construct a complex structure to perform high-level abstraction on data. In image classification, a Convolutional Neural Network (CNN) method is often used. The CNN-based image classification method mainly consists in how to train the network model. And after the trained model is obtained, directly inputting the test picture, and obtaining a label with the maximum response, namely the label of the prediction picture.
Although the image classification technology based on deep learning is endless, there are some major problems: the amount of training data is relatively lacking. Aiming at the classification problem of millimeter wave images, the number of labeled images is limited, and too much effective data cannot be provided for training, so that the training data is deficient, and the model obtained by training cannot meet the use requirement under the condition of insufficient samples; meanwhile, the processing efficiency based on deep learning is low, and the network layer number and the like have high requirements on the configuration of computer hardware.
For example, chinese patent application CN108364017A discloses an image quality classification method suitable for millimeter wave images, which extracts three-dimensional feature vectors of images, inputs the vectors into a convolutional neural network for training, and performs quality classification on the images according to a trained neural network model and a preset omission factor and false detection factor.
Disclosure of Invention
The invention aims to provide a method for constructing a millimeter wave image quality classification model for accurately identifying the quality of a millimeter wave image based on local enhancement processing.
The invention solves the technical problems through the following technical scheme:
after the input image is locally enhanced, the definition of the whole image can be improved, the input image has higher discrimination with a blurred image, and the definition of the blurred image cannot be obviously changed after the input image is enhanced; and aiming at the enhanced image, effective characteristic vectors are constructed and extracted and sent to a classifier for training, and the method has a better distinguishing effect on the distinguishing of the good and bad image quality.
The invention provides a millimeter wave image quality classification model construction method based on local enhancement, which comprises the following steps:
step A: acquiring a training image, and manually calibrating the image quality to be clear or fuzzy;
and B: making a square template, performing smooth movement on the template on an image line by line, performing histogram equalization processing on the image in a template region, and replacing an original central point pixel value with an image region central point pixel value corresponding to the equalized template;
and C: extracting the directional gradient histogram characteristic and the gray level co-occurrence matrix characteristic of the image;
step D: and inputting the extracted image features into an SVM classifier to solve an optimal hyperplane so as to obtain a millimeter wave image quality classification model, wherein the kernel function of the SVM classifier adopts a kernel function form of a radial basis function.
According to the invention, the contrast ratio of a clear image and a blurred image is improved through local enhancement, the difference between the two types of images is increased, the information of the millimeter wave image can be better described by constructing the feature vector through the directional gradient histogram feature and the gray level co-occurrence matrix feature, so that after the images are classified, the difference of different types of images is as large as possible, the kernel function of the SVM classifier is constructed based on the kernel function form of the radial basis function, the relation between the feature vector and the class label is more flexible to process, and the application is wide.
Preferably, in the step a, the image is calibrated to be clear or fuzzy according to the fact that whether the head, the arms and the legs of the person in the image are ghosted or not.
Preferably, the template pixel size in step B is 3 × 3 or 5 × 5, and the histogram equalization relationship in the template region is g (x, y) ═ T (f (x, y))
The histogram equalization relation of the center point of the template is
g(x0,y0)＝T(f(x0,y0))
Wherein f (x, y) represents the original gray value of the pixel point with the coordinate (x, y), g (x, y) represents the gray value of the pixel point with the coordinate (x, y) after histogram equalization processing, and the operator T (-) represents the histogram equalization operation.
Preferably, the processing method of the gray scale value after histogram equalization of the edge position of the original millimeter wave image is as follows:
calculating pixel points with gray values of 0 from the edge to the outside; or pixel points with the same gray value as the changed gray value are added from the edge to the outside, and the gray value at the corner is the average value of the adjacent pixel points; or preserving the original grey value; or discarding the pixel point.
Preferably, the method for extracting histogram feature of oriented gradient described in step C includes the following steps:
step i: carrying out normalization processing on the millimeter wave image subjected to the local enhancement processing;
step ii: performing convolution operation on the normalized millimeter wave image to obtain gradient components of each pixel point in the horizontal direction and the vertical direction;
step iii: dividing an image into a plurality of units, equally dividing the gradient direction into 9 intervals within 0-180 degrees, and performing weighted projection on each pixel in each unit in the gradient direction of each interval to obtain a gradient direction histogram and a 9-dimensional feature vector of each unit;
step iv: constructing a block by 2 multiplied by 2 units, adopting a sliding window form, and scanning an image by using the block by taking one unit as a step length;
step v: and carrying out normalization processing on the feature vectors in the blocks to obtain block feature strings, namely, the feature vectors of the directional gradient histograms.
Preferably, the method for performing normalization processing on the millimeter wave image in step i includes: the image gray scale is normalized by Gamma correction,
f(I)＝Iγ
wherein, I represents the gray value of the pixel point, when gamma is less than 1, the dynamic range change is large in the low gray value area, and the image contrast is enhanced; in the high gray value area, the dynamic range is reduced, the image contrast is reduced, and the integral gray value of the image is increased; when gamma is larger than 1, the dynamic range of the low gray value area is reduced, the dynamic range of the high gray value area is enlarged, the image contrast of the low gray value area is reduced, the image contrast of the high gray value area is improved, and the integral gray value of the image is reduced.
Preferably, a one-dimensional discrete differential template [ -1,0,1 ] is used in step ii]And its transpose [ -1,0,1 [ ]]TPerforming convolution operation, namely obtaining the gradient components of the image in the horizontal direction and the vertical direction:
Gx(x,y)＝H(x+1,y)-H(x-1,y)
Gy(x,y)＝H(x,y+1)-H(x,y-1)
wherein G isx(x,y)、Gy(x, y) and G (x, y) are respectively the horizontal gradient, the vertical gradient and the pixel value of the current pixel point, and respectively represent the gradient amplitude and the gradient direction of the current pixel point.
Preferably, each unit in step iii is 8 × 8 pixel points, and for each pixel point in the unit, projection is performed for each direction interval according to the gradient amplitude of the pixel point.
Preferably, each block in step v is a 36-dimensional feature vector of 2 × 2 × 9, the 36-dimensional feature vector in each block is normalized,
among them, a constant having an extremely small value.
Preferably, the method for acquiring the gray level co-occurrence matrix characteristic in step C includes: for a pixel point (x, y) with a gray value of i in the image, counting the number p (i, j, d, theta) of simultaneous occurrence of a pixel (x + Dx, y + Dy) with a distance d and a gray value of j from the pixel point, namely
p(i,j,d,θ)＝[(x,y),(x+Dx,y+Dx)|f(x,y)＝i,f(x+Dx,y+Dy)＝j]
Where i, j is 0,1,2, …, L-1 is a gray scale value, Dx,DyIs the position offset, d is the compensation for generating the gray level co-occurrence matrix, and theta is the direction of the generated co-occurrence matrix; and carrying out normalization processing on the gray level co-occurrence matrix to obtain:
and selecting the mean values in 4 directions as texture characteristic parameters and as the characteristics of the gray level co-occurrence matrix.
Preferably, the input data input into the SVM classifier in the step D is a feature vector obtained by combining the histogram feature of the direction gradient and the gray level co-occurrence matrix feature of each picture;
mapping the input vector to a high-dimensional feature space by nonlinear mapping, the mapping equation being
Wherein the content of the first and second substances,
constructing a kernel function k (x) based on kernel function form of radial basis function for transformed space inner producti,xj) Is made equal to the inner product operation of the transformed space
Namely, it is
Obtaining an optimal solution by means of operational solution
I.e. the optimal hyperplane.
The method for constructing the quality classification model based on the locally enhanced millimeter wave image has the advantages that: the contrast ratio of a clear image and a blurred image is improved through local enhancement, the difference between the two images is increased, the feature vector is constructed through the histogram feature of the directional gradient and the gray level co-occurrence matrix feature, the information of the millimeter wave image can be better described, therefore, after the images are classified, the difference of the images of different classes is as large as possible, the kernel function of the SVM classifier is constructed based on the kernel function form of the radial basis function, the relation between the feature vector and the class label is more flexible to process, and the application is wide.
The method comprises the steps of firstly cleaning millimeter wave image data according to the definition degree of an image, namely whether the head shakes or not and whether the hands and legs are clear or not, and judging whether the image is a clean clear image or a fuzzy image.
Before the image features are extracted, local image enhancement operation is firstly carried out on a training image and a test image, namely self-adaptive histogram equalization is used, so that a clear image is clearer and contains rich texture features and detail information, the image quality of a blurred image cannot be remarkably improved, and the difference of the quality of the two types of images is further enhanced.
According to the method, for the enhanced millimeter wave image, the features of the image are extracted by using a HOG and GLCM combined method, the feature vector is constructed, and the information of the millimeter wave image is better described, so that the image difference of different classes is as large as possible after the image is classified.
Drawings
FIG. 1 is a flow chart of a classification method based on feature extraction in the background art of the present invention;
FIG. 2 is a flowchart of a method for constructing a quality classification model based on locally enhanced millimeter wave images according to an embodiment of the present invention;
FIG. 3 is a comparison diagram of a locally enhanced sharp image and a blurred image provided by an embodiment of the present invention;
fig. 4 is a nonlinear separable binary diagram of the method for constructing the quality classification model based on the locally enhanced millimeter wave image according to the embodiment of the present invention.
Detailed Description
In order that the objects, technical solutions and advantages of the present invention will become more apparent, the present invention will be further described in detail with reference to the accompanying drawings in conjunction with the following specific embodiments.
As shown in fig. 2, the embodiment provides a quality classification model construction method based on a locally enhanced millimeter wave image, including the following steps:
step A: acquiring a training image, and manually calibrating the image quality to be clear or fuzzy;
with reference to fig. 3, the present embodiment marks the image as sharp or blurred based on whether there are ghosts on the head, arms and legs of the person in the image, where fig. 3(a) is an original sharp image and fig. 3(c) is an original blurred image.
And B: making a square template, performing smooth movement on the template on an image line by line, performing histogram equalization processing on the image in a template region, and replacing an original central point pixel value with an image region central point pixel value corresponding to the equalized template;
the pixel size of the template is 3X 3 or 5X 5, the histogram equalization relation in the template area is g (x, y) ═ T (f (x, y))
The histogram equalization relation of the center point of the template is
g(x0,y0)＝T(f(x0,y0))
Wherein f (x, y) represents the original gray value of the pixel point with the coordinate (x, y), g (x, y) represents the gray value of the pixel point with the coordinate (x, y) after histogram equalization processing, and the operator T (-) represents the histogram equalization operation.
The processing method of the gray value after the histogram equalization of the edge position of the original millimeter wave image comprises the following steps:
calculating pixel points with gray values of 0 from the edge to the outside; or pixel points with the same gray value as the changed gray value are added from the edge to the outside, and the gray value at the corner is the average value of the adjacent pixel points; when the template size is 3 x 3, the outer lane mends a circle can, and when the template size is 5 x 5, the outer lane need mend two circles, and the corresponding pixel point number of turns of confirming concrete replenishment when selecting for use the template of other sizes can.
Or directly keeping the original gray value for the edge pixel point; or discarding the pixel point; after the processing of the steps, the image is optimally enhanced, fig. 3(b) is a processed clear image, compared with fig. 3(a), the texture and the details are richer, and fig. 3(d) is a locally enhanced blurred image, compared with fig. 3(c), the quality of the clear image can be improved and the blurred image is not obviously affected after the processing of the local enhancement provided by the implementation; therefore, the discrimination of the clear image and the fuzzy image is improved, and the training of the model is facilitated. And after local histogram equalization, the brightness is redistributed to improve the contrast of the image, further improve the imaging quality of the image at the chest and the arms, conveniently distinguish a clear image from a blurred image, and improve the classification precision.
And C: the characteristics of the image are extracted,
the process of feature selection of the image after the local enhancement processing comprises the following steps: (1) the method comprises the steps of (1) manually selecting, analyzing the shape, the texture and the like of a clear image and a blurred image, selecting features which can be separated into two types, (2) obtaining through experiments, extracting image features, training by using a classifier, selecting or optimizing better features according to the quality of a test result, and selecting two features of a direction gradient Histogram (HOG) and a gray level co-occurrence matrix (GLCM) for feature extraction.
The directional gradient histogram feature extraction packet block comprises the following steps:
step i: carrying out normalization processing on the millimeter wave image subjected to the local enhancement processing; in particular to the method for normalizing the image gray scale by using Gamma correction,
f(I)＝Iγ
wherein, I represents the gray value of the pixel point, when gamma is less than 1, the dynamic range change is large in the low gray value area, and the image contrast is enhanced; in the high gray value area, the dynamic range is reduced, the image contrast is reduced, and the integral gray value of the image is increased; when gamma is larger than 1, the dynamic range of the low gray value area is reduced, the dynamic range of the high gray value area is enlarged, the image contrast of the low gray value area is reduced, the image contrast of the high gray value area is improved, and the integral gray value of the image is reduced. By the method, local shadow and illumination change of the millimeter wave image can be reduced, so that the directional gradient histogram feature has better robustness to the illumination change.
Step ii: performing convolution operation on the normalized millimeter wave image to obtain gradient components of each pixel point in the horizontal direction and the vertical direction; the specific method comprises the following steps: using a one-dimensional discrete differential template [ -1,0,1 [ -1]And its transpose [ -1,0,1 [ ]]TPerforming convolution operation, namely obtaining the gradient components of the image in the horizontal direction and the vertical direction:
Gx(x,y)＝H(x+1,y)-H(x-1,y)
Gy(x,y)＝H(x,y+1)-H(x,y-1)
wherein G isx(x,y)、Gy(x, y) and G (x, y) are respectively the horizontal gradient, the vertical gradient and the pixel value of the current pixel point, and respectively represent the gradient amplitude and the gradient direction of the current pixel point.
Step iii: dividing an image into a plurality of units, wherein each unit (cell) is 8 × 8 pixel points, dividing the gradient direction into 9 intervals in 0-180 degrees, each interval is 20 degrees, performing weighted projection on each pixel in each unit in the gradient direction of each interval, namely projecting each pixel in the unit in each direction according to the gradient amplitude to obtain a 9-dimensional feature vector of a gradient direction histogram of each unit;
step iv: the segmentation operation is performed on 16 × 16 image blocks selected from the whole image, and for each 16 × 16 image block, 4 cells can be formed by taking 8 × 8 cells as units. Scanning an image block by using blocks with one cell as a step length in a sliding window mode, wherein each obtained block is a 36-dimensional feature vector v of 2 x 9;
step v: normalizing the feature vector v in the block to obtain a block feature string which is the feature vector of the directional gradient histogram; the normalization processing method comprises the following steps:
wherein, the method is a constant term with extremely small numerical value and is used for preventing the denominator from being 0.
The method for acquiring the gray level co-occurrence matrix characteristics comprises the following steps: for a pixel point (x, y) with a gray value of i in the image, counting the number p (i, j, d, theta) of simultaneous occurrence of a pixel (x + Dx, y + Dy) with a distance d and a gray value of j from the pixel point, namely
p(i,j,d,θ)＝[(x,y),(x+Dx,y+Dx)|f(x,y)＝i,f(x+Dx,y+Dy)＝j]
Where i, j is 0,1,2, …, L-1 is a gray scale value, Dx,DyIs the position offset, d is the compensation for generating the gray level co-occurrence matrix, and theta is the direction of the generated co-occurrence matrix; the gray level co-occurrence matrix can derive a plurality of texture characteristics, and normalization processing is carried out on the gray level co-occurrence matrix to obtain the following characteristics:
wherein Con represents contrast, reflecting the image definition and the groove depth of the texture; asm represents energy, and reflects the uniformity degree and the texture thickness of the gray level distribution of the image; ent represents entropy, which is a measure of randomness of the amount of information contained in an image; h represents an inverse variance, and reflects the size of local change of the image texture; corr represents the correlation and is used to measure how similar the gray levels of an image are in the row or column direction.
In the embodiment, gray level co-occurrence matrixes of (0 degrees, 45 degrees, 90 degrees and 135 degrees) are calculated, then the gray level co-occurrence matrixes in the directions are circularly calculated, normalization processing is performed, and contrast Con, energy Asm, entropy Ent, inverse variance H and correlation Corr are selected as characteristics finally extracted in a test.
After the processing, combining the histogram feature of the direction gradient and the gray level co-occurrence matrix feature of each image to obtain the feature vector of the image.
Step D: and inputting the extracted image features into an SVM classifier to solve an optimal hyperplane so as to obtain a millimeter wave image quality classification model, wherein the kernel function of the SVM classifier adopts a kernel function form of a radial basis function.
The SVM is a popular method for processing machine learning based on an optimization theory, and is mainly used for solving the problem of two classifications. The classification problem mainly includes two aspects: linear separable and non-linear separable. The objective is to solve the classification plane, i.e. the optimal hyperplane, that maximizes the distance between the two classes.
The present invention considers the case of non-linear divisible, i.e. the input vector is mapped to a high-dimensional feature space by a pre-selected non-linear mapping (kernel function), and referring to fig. 4, the SVM linear inseparable mapping equation is:
wherein the content of the first and second substances,
constructing a kernel function k (x) based on kernel function form of radial basis function for transformed space inner producti,xj) Is made equal to the inner product operation of the transformed space
Namely, it is
Because the kernel function is used, the millimeter wave image of the sample can be mapped to a space with higher dimensionality for processing the nonlinear relation between the characteristic and the class label, and the method has higher flexibility and wide application. Solving to obtain an optimal solution
A separating hyperplane is obtained.
And (3) performing quality classification on the test picture based on the hyperplane function, wherein the histogram equalization processing and the extraction of the direction gradient histogram and the gray level co-occurrence matrix in the steps B-C still need to be performed on the test picture at the moment, then the processed picture is input into a classifier, and the result that the picture is a clear picture or a fuzzy picture is output.
The embodiment further evaluates the quality of the classification result through the confusion matrix, namely, the classification accuracy is obtained by calculating the ratio of the value on the diagonal line of the confusion matrix to the total number of each row. The confusion matrix M is:
wherein m isijAnd the data which represents the ith type pattern judged as the jth type pattern by the classifier accounts for the percentage of the total number of the ith type pattern samples. In the experiment, a plurality of clear images and fuzzy images are selected and divided into a training picture set and a test set according to the proportion of 7: 3. The experimental effect of classification is measured by accuracy, namely:
table 1: verification of test results
The final experimental results are shown in table 1, where all the blurred pictures are classified and detected, and the clear pictures have only a few classification errors. The quality classification model constructed by the method provided by the embodiment can effectively classify and identify the millimeter wave image.
The technical idea of this embodiment is to first apply Adaptive Histogram Equalization (AHE) to the millimeter wave images for training. The common histogram equalization algorithm performs the same histogram transformation on all pixel points in the whole image, but if the image contains locally too bright or locally too dark portions, the contrast of the portions cannot be effectively enhanced. The adaptive histogram equalization technology can better improve the contrast of the millimeter wave image through the histogram equalization which responds to the local area. It can be simply understood that for each pixel point in the image, all pixels in a rectangular range around the pixel point are used for histogram equalization.
And secondly, extracting a feature vector from the enhanced image. The vector is composed of two parts, one is Histogram of Oriented Gradients (HOG), in which the directional distribution of Gradients is used as a feature, because edges and corners in the millimetric wave image contain much shape information, and the Gradients in the X-axis and Y-axis directions of a picture are useful feature information. The other is to use a Gray-level Co-occurrence Matrix (GLCM), which is defined by the joint probability density of two pixels, and it can reflect not only the distribution characteristics of the brightness, but also the position distribution characteristics between pixels with the same brightness or close to the brightness, which is a second-order statistical characteristic related to the brightness variation of the image. Which is the basis for defining a set of texture features. Since the texture is formed by repeatedly generating the gray scale at a spatial position, a certain gray scale relationship, that is, a spatial correlation characteristic of the gray scale in the image, exists between two pixels spaced apart from each other in the image space. The purpose of using the gray level co-occurrence matrix is to describe the texture of the image by studying the spatial correlation characteristics of the gray levels, and 14 statistics can be calculated using the gray level co-occurrence matrix, namely: energy, entropy, contrast, homogeneity, correlation, variance, sum-mean, sum-variance, sum-entropy, difference variance, difference-mean, difference entropy, measure of correlation information, and maximum correlation coefficient. With regard to the millimeter wave image related to the present application, a sharp image and a blurred image have a great difference at a texture level, and therefore, the texture characteristics of the millimeter wave image can be well described using GLCM. Therefore, the method for calculating the gray level co-occurrence matrix of the millimeter wave image is adopted as a part of characteristics in the application. And finally, combining the HOG and GLCM characteristics to obtain a finally extracted characteristic vector.
Considering that the classification and judgment of the image quality are not a linear problem and do not have linear separability, a nonlinear classifier model is adopted. The extracted features are sent to a classifier for training to obtain a classifier model; and finally, carrying out prediction classification on the pictures to be predicted to obtain two types of images with good image quality and bad image quality.
Claims (11)
1. A millimeter wave image quality classification model construction method based on local enhancement is characterized in that: the method comprises the following steps:
step A: acquiring a training image, and manually calibrating the image quality to be clear or fuzzy;
and B: making a square template, performing smooth movement on the template on an image line by line, performing histogram equalization processing on the image in a template region, and replacing an original central point pixel value with an image region central point pixel value corresponding to the equalized template;
and C: extracting the directional gradient histogram characteristic and the gray level co-occurrence matrix characteristic of the image;
step D: and inputting the extracted image features into an SVM classifier to solve an optimal hyperplane so as to obtain a millimeter wave image quality classification model, wherein the kernel function of the SVM classifier adopts a kernel function form of a radial basis function.
2. The method for constructing the millimeter wave image quality classification model based on local enhancement according to claim 1, wherein the method comprises the following steps: and B, calibrating the image to be clear or fuzzy by taking whether the head, the arms and the legs of the person in the image have double images as a standard in the step A.
3. The method for constructing the millimeter wave image quality classification model based on local enhancement according to claim 1, wherein the method comprises the following steps: the size of the template pixel in the step B is 3X 3 or 5X 5, and the histogram equalization relation in the template area is
g(x,y)＝T(f(x,y))
The histogram equalization relation of the center point of the template is
g(x0,y0)＝T(f(x0,y0))
Wherein f (x, y) represents the original gray value of the pixel point with the coordinate (x, y), g (x, y) represents the gray value of the pixel point with the coordinate (x, y) after histogram equalization processing, and the operator T (-) represents the histogram equalization operation.
4. The method for constructing the millimeter wave image quality classification model based on local enhancement according to claim 3, wherein the method comprises the following steps: the processing method of the gray value after the histogram equalization of the edge position of the original millimeter wave image comprises the following steps:
calculating pixel points with gray values of 0 from the edge to the outside; or pixel points with the same gray value as the changed gray value are added from the edge to the outside, and the gray value at the corner is the average value of the adjacent pixel points; or preserving the original grey value; or discarding the pixel point.
5. The method for constructing the quality classification model based on the locally enhanced millimeter wave image according to claim 3, wherein the method comprises the following steps: the method for extracting the histogram feature of the directional gradient in the step C comprises the following steps:
step i: carrying out normalization processing on the millimeter wave image subjected to the local enhancement processing;
step ii: performing convolution operation on the normalized millimeter wave image to obtain gradient components of each pixel point in the horizontal direction and the vertical direction;
step iii: dividing an image into a plurality of units, equally dividing the gradient direction into 9 intervals within 0-180 degrees, and performing weighted projection on each pixel in each unit in the gradient direction of each interval to obtain a gradient direction histogram and a 9-dimensional feature vector of each unit;
step iv: constructing a block by 2 multiplied by 2 units, adopting a sliding window form, and scanning an image by using the block by taking one unit as a step length;
step v: and carrying out normalization processing on the feature vectors in the blocks to obtain block feature strings, namely, the feature vectors of the directional gradient histograms.
6. The method for constructing the quality classification model based on the locally enhanced millimeter wave image according to claim 5, wherein the method comprises the following steps: the method for normalizing the millimeter wave image in the step i comprises the following steps: the image gray scale is normalized by Gamma correction,
f(I)＝Iγ
wherein, I represents the gray value of the pixel point, when gamma is less than 1, the dynamic range change is large in the low gray value area, and the image contrast is enhanced; in the high gray value area, the dynamic range is reduced, the image contrast is reduced, and the integral gray value of the image is increased; when gamma is larger than 1, the dynamic range of the low gray value area is reduced, the dynamic range of the high gray value area is enlarged, the image contrast of the low gray value area is reduced, the image contrast of the high gray value area is improved, and the integral gray value of the image is reduced.
7. The method for constructing the quality classification model based on the locally enhanced millimeter wave image according to claim 5, wherein: step (ii) ofii using a one-dimensional discrete differential template [ -1,0,1 [ -1 [ ]]And its transpose [ -1,0,1 [ ]]TPerforming convolution operation, namely obtaining the gradient components of the image in the horizontal direction and the vertical direction:
Gx(x,y)＝H(x+1,y)-H(x-1,y)
Gy(x,y)＝H(x,y+1)-H(x,y-1)
wherein G isx(x,y)、Gy(x, y) and G (x, y) are respectively the horizontal gradient, the vertical gradient and the pixel value of the current pixel point, and respectively represent the gradient amplitude and the gradient direction of the current pixel point.
8. The method for constructing the quality classification model based on the locally enhanced millimeter wave image according to claim 7, wherein: and in the step iii, each unit is 8 multiplied by 8 pixel points, and for each pixel point in the unit, projection is carried out for each direction interval according to the gradient amplitude of the pixel point.
9. The method for constructing the quality classification model based on the locally enhanced millimeter wave image according to claim 5, wherein the method comprises the following steps: in the step v, each block is a 36-dimensional feature vector of 2 multiplied by 9, the 36-dimensional feature vector in each block is normalized,
among them, a constant having an extremely small value.
10. The method for constructing the quality classification model based on the locally enhanced millimeter wave image according to claim 9, wherein: the method for acquiring the gray level co-occurrence matrix characteristics in the step C comprises the following steps: for a pixel point (x, y) with a gray value of i in the image, counting the number p (i, j, d, theta) of simultaneous occurrence of a pixel (x + Dx, y + Dy) with a distance d and a gray value of j from the pixel point, namely
p(i,j,d,θ)＝[(x,y),(x+Dx,y+Dx)|f(x,y)＝i,f(x+Dx,y+Dy)＝j]
Where i, j is 0,1,2, …, L-1 is a gray scale value, Dx,DyIs the position offset, d is the compensation for generating the gray level co-occurrence matrix, and theta is the direction of the generated co-occurrence matrix; and carrying out normalization processing on the gray level co-occurrence matrix to obtain:
and selecting the mean values in 4 directions as texture characteristic parameters and as the characteristics of the gray level co-occurrence matrix.
11. The method for constructing the millimeter wave image quality classification model based on local enhancement according to claim 10, wherein: d, inputting the input data of the SVM classifier into a feature vector obtained by combining the histogram feature of the direction gradient and the gray level co-occurrence matrix feature of each picture;
mapping the input vector to a high-dimensional feature space by nonlinear mapping, the mapping equation being
Wherein the content of the first and second substances,
constructing a kernel function k (x) based on kernel function form of radial basis function for transformed space inner producti,xj) Is made equal to the inner product operation of the transformed space
Namely, it is
Obtaining an optimal solution by means of operational solution
I.e. the optimal hyperplane.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010624153.9A CN111783885A (en) | 2020-07-01 | 2020-07-01 | Millimeter wave image quality classification model construction method based on local enhancement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010624153.9A CN111783885A (en) | 2020-07-01 | 2020-07-01 | Millimeter wave image quality classification model construction method based on local enhancement |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111783885A true CN111783885A (en) | 2020-10-16 |
Family
ID=72760133
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010624153.9A Pending CN111783885A (en) | 2020-07-01 | 2020-07-01 | Millimeter wave image quality classification model construction method based on local enhancement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111783885A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112308793A (en) * | 2020-10-21 | 2021-02-02 | 淮阴工学院 | Novel method for enhancing contrast and detail of non-uniform illumination image |
| CN113291055A (en) * | 2021-04-14 | 2021-08-24 | 西安理工大学 | Artificial intelligent flexographic printing pressure prediction method |
| CN113408651A (en) * | 2021-07-12 | 2021-09-17 | 厦门大学 | Unsupervised three-dimensional object classification method based on local discriminability enhancement |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103116744A (en) * | 2013-02-05 | 2013-05-22 | 浙江工业大学 | Fake fingerprint detection method based on markov random field (MRF) and support vector machine-k nearest neighbor (SVM-KNN) classification |
| CN103426156A (en) * | 2012-05-15 | 2013-12-04 | 中国科学院声学研究所 | SAS image segmentation method and system based on SVM classifier |
| CN110781913A (en) * | 2019-09-11 | 2020-02-11 | 西安电子科技大学 | Zipper cloth belt defect detection method |
-
2020
- 2020-07-01 CN CN202010624153.9A patent/CN111783885A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103426156A (en) * | 2012-05-15 | 2013-12-04 | 中国科学院声学研究所 | SAS image segmentation method and system based on SVM classifier |
| CN103116744A (en) * | 2013-02-05 | 2013-05-22 | 浙江工业大学 | Fake fingerprint detection method based on markov random field (MRF) and support vector machine-k nearest neighbor (SVM-KNN) classification |
| CN110781913A (en) * | 2019-09-11 | 2020-02-11 | 西安电子科技大学 | Zipper cloth belt defect detection method |
Non-Patent Citations (3)
| Title |
|---|
| 储理才: "基于最优超平面方法的乳房肿瘤分类模型" * |
| 祝志慧;汤勇;洪琪;黄飘;王巧华;马美湖;: "基于种蛋图像血线特征和深度置信网络的早期鸡胚雌雄识别" * |
| 陈夏玲: "胆囊B超图像自动识别的研究" * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112308793A (en) * | 2020-10-21 | 2021-02-02 | 淮阴工学院 | Novel method for enhancing contrast and detail of non-uniform illumination image |
| CN113291055A (en) * | 2021-04-14 | 2021-08-24 | 西安理工大学 | Artificial intelligent flexographic printing pressure prediction method |
| CN113408651A (en) * | 2021-07-12 | 2021-09-17 | 厦门大学 | Unsupervised three-dimensional object classification method based on local discriminability enhancement |
| CN113408651B (en) * | 2021-07-12 | 2024-01-23 | 厦门大学 | Unsupervised three-dimensional object classification method based on local discriminant enhancement |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110334706B (en) | Image target identification method and device | |
| CN108765412B (en) | Strip steel surface defect classification method | |
| CN113592845A (en) | Defect detection method and device for battery coating and storage medium | |
| Moallem et al. | Optimal threshold computing in automatic image thresholding using adaptive particle swarm optimization | |
| CN111783885A (en) | Millimeter wave image quality classification model construction method based on local enhancement | |
| CN107038416B (en) | Pedestrian detection method based on binary image improved HOG characteristics | |
| CN116664559B (en) | Machine vision-based memory bank damage rapid detection method | |
| CN111340824B (en) | Image feature segmentation method based on data mining | |
| CN111753828A (en) | Natural scene horizontal character detection method based on deep convolutional neural network | |
| CN109801305B (en) | SAR image change detection method based on deep capsule network | |
| CN115375690B (en) | Classification and identification method for greasy tongue coating | |
| CN108596195B (en) | Scene recognition method based on sparse coding feature extraction | |
| CN109948566A (en) | A kind of anti-fraud detection method of double-current face based on weight fusion and feature selecting | |
| CN107209942A (en) | Method for checking object and image indexing system | |
| CN108647706A (en) | Article identification classification based on machine vision and flaw detection method | |
| CN108073940A (en) | A kind of method of 3D object instance object detections in unstructured moving grids | |
| CN107610136B (en) | Salient object detection method based on convex hull structure center query point sorting | |
| CN111597875A (en) | Traffic sign identification method, device, equipment and storage medium | |
| CN107423771B (en) | Two-time-phase remote sensing image change detection method | |
| CN105844299B (en) | A kind of image classification method based on bag of words | |
| CN111127407B (en) | Fourier transform-based style migration forged image detection device and method | |
| CN103902965B (en) | Spatial domain symbiosis image representing method and its application in image classification, identification | |
| CN105512682B (en) | A kind of security level identification recognition methods based on Krawtchouk square and KNN-SMO classifier | |
| CN112101283A (en) | Intelligent identification method and system for traffic signs | |
| CN109829511B (en) | Texture classification-based method for detecting cloud layer area in downward-looking infrared 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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20201016 |
|
| RJ01 | Rejection of invention patent application after publication |