WO2021208424A1

WO2021208424A1 - Laplacian feature mapping learning-based machine testing method for graphics card interface

Info

Publication number: WO2021208424A1
Application number: PCT/CN2020/129654
Authority: WO
Inventors: 陈博
Original assignee: 艾瑞思检测技术(苏州)有限公司
Priority date: 2020-04-14
Filing date: 2020-11-18
Publication date: 2021-10-21
Also published as: CN111507393A; CN111507393B

Abstract

A Laplacian feature mapping learning-based machine testing method for a graphics card interface, which is carried out according to the following steps: step 1: a product to be tested is connected to a video collector by using a corresponding type of wire; step 2: a test video is played back, and the video collector collects output information of a display interface and uploads same to an industrial host computer; step 3: the industrial host computer decodes graphics card interface information and saves a frame image; step 4: Laplacian feature mapping and dimensionality reduction preprocessing is performed on rows and columns of the frame image separately; and step 5: an analyzed image is used as an input for a machine learning classifier so as to obtain a detection result. The method is compatible with various types of interfaces without needing to be equipped with a display corresponding to the interface, and uses a video collector and an industrial host computer to complete the collection and decoding of interface information, which reduces the amount of material resources and costs.

Description

A machine testing method for graphics card interface based on Laplacian feature mapping learning

Technical field

The invention belongs to the field of graphics card interface testing, and in particular relates to an automatic testing method of a graphics card interface.

Background technique

In the field of graphics card interface testing, it is usually necessary to connect the product to be tested to a monitor with a corresponding type of interface, and then play the test video, and manually check whether the video display quality is qualified. Not only need to be equipped with displays of various interface types on the equipment, but manual inspections also affect production efficiency due to manual errors and long inspection time. Especially in recent years, the market's requirements for image quality have continued to increase, requiring higher-performance displays, and manual visual inspection is also more difficult to distinguish small defects, so a lot of manpower and material resources are required. For this reason, the present invention uses an industrial host computer equipped with a composite interface video collector to replace the display, collects the output signal of the graphics card interface through the video collector, and decodes it in the host computer, and realizes the automatic classification of video quality in combination with a machine learning method.

Summary of the invention

In order to overcome the above-mentioned difficulties, the present invention provides an automated test method for a graphics card interface, which includes the following steps:

1. A test method for graphics card interface machines based on Laplace feature mapping learning, according to the following steps:

Step 1: Connect the product to be tested to the video capture device with the corresponding type of wire;

Step 2: Play the test video, the video collector collects the output information of the display interface, and uploads it to the industrial host computer;

Step 3: The industrial host computer decodes the graphics card interface information and saves the frame image;

Step 4: Perform Laplacian feature mapping dimensionality reduction preprocessing on the rows and columns of the frame image respectively, the purpose of which is to reduce the data dimension and increase the detection rate;

First of all, each frame of image can be digitized into an m×n matrix A. Each element in matrix A represents a pixel in the picture. The frame image has a total of mn pixels. The size of m and n is determined by the test video. Determined by the resolution; in order to achieve dimensionality reduction of Laplacian feature mapping, the image A is rearranged and combined into the following vector form

z=[A _v1 A _v2 … A _vn ] ^T

Among them, A _v1 A _v2 … A _vn represents the column vector of image A; then z is put into the historical data set for Laplacian feature map dimensionality reduction processing;

Step 5: Use the analyzed image as the input of the machine learning classifier to obtain the detection result; select the three-layer neural network as the training model of the machine learning classifier, and the training data

It is obtained from the marked historical test data, where y _l =1 means that the _{quality of z kl} is unqualified, and y _l =0 means that the _{quality of z kl} is qualified;

The input layer of the neural network is composed of k neurons, which corresponds to the k elements of image A after dimensionality reduction. The hidden layer is composed of p neurons, and its output is

_h [omega] corresponding to the offset of each hidden layer neuron, ω _hi corresponds to the neuron input u _i t _h, _weightings, σ is the activation function

Finally, the output layer is composed of 2 neurons, which represent qualified and unqualified respectively; its expression is

Where v _j is the offset of the corresponding output, and v _jh is the corresponding input t _h to the output

the weight of.

Laplacian Eigenmaps (LE) is a non-linear dimensionality reduction method that uses a local perspective to construct the relationship between data, which can reflect the inherent manifold structure of the data. Its intuitive idea is to hope that the points that are related to each other are as close as possible in the space after dimensionality reduction. The preferred step 4 is

The objective function for minimizing the Laplacian feature map is

in

Is ^{the point after z (i) is} reduced to k dimension, w _ij is the weight of the connection between the measurement samples z ⁽ⁱ⁾ and z ^(j) _{; w ij} is determined according to whether the two sample points are close, first use KNN The method determines whether to set an edge connection between the ^{sample z (i)} and z ^(j) ^{. If z (i) is} among the k nearest neighbors of ^{z (j)} ^{, then z (i)} and z ^{(j) are} connected, k is a preset value, or set an appropriate ε, will

The nodes are connected; then the weight is determined, and the Heat kernel function is used to set the weight of the connected nodes to

Here t is a preset value, or let t=∞, set the weights of all connected nodes to w _ij =1, and all other unconnected nodes are 0; finally, a symmetric adjacency matrix W can be obtained.

By minimizing the objective function formula (1), it is ensured that the two points u ⁽ⁱ⁾ and u ^(j) can still remain close ^{after the similar z (i)} and z ^{(j) are mapped.} The preferred objective function can be expressed as the following quadratic form after sorting

Where u = (u ⁽¹⁾ , u ⁽²⁾ ,..., u ⁽ⁿ⁾ ) ^T , L = DW is the Laplacian matrix, D is a diagonal matrix, satisfying D _ii ＝∑ _j w _ij , W Is a symmetric adjacency matrix, and the Laplacian matrix L is positive semi-definite;

The final solution is the following minimization problem

Among them, the constraint u ^T Du = 1 avoids the impact of scaling, and the vector u that minimizes the objective function is given by the minimum eigenvalue solution of the generalized eigenvalue problem:

Lu = λDu (4)

The eigenvector corresponding to the non-zero eigenvalue obtained by the solution is the output after dimensionality reduction.

Preferably, step 5: Finally, train to obtain the weights ω _hi , v _jh and the bias ω _h , v _{j by solving the following optimization problem}

Where y _lj is the label of the sample A _kl , and N is the number of samples of historical data; finally, the output image of the graphics card interface of the product to be tested is used as input to obtain the final detection result.

The advantages of the present invention are mainly embodied in that it can be compatible with various types of interfaces, does not need to be equipped with a display corresponding to the interface, uses a video collector and an industrial upper computer to complete the collection and decoding of interface information, and reduces the material cost. In terms of detection, it combines principal component analysis and machine learning methods to achieve rapid and efficient automatic video quality detection, reducing labor costs and improving efficiency.

Description of the drawings

Fig. 1 is a flowchart of an automated testing method for a graphics card interface of the present invention.

Figure 2 is a framework diagram of an automated testing method for a graphics card interface of the present invention.

Detailed ways

In the following, the present invention will be further described with reference to Figures 1 and 2.

Step 1: Connect the product to be tested to the industrial host computer with the video capture card installed with the corresponding type of wire;

Step 2: Play the test video, and the video capture card collects the graphics interface information of the product to be tested and sends it to the industrial host computer

Step 3: The industrial host computer decodes the graphics card interface information to generate and save a frame image;

Step 4: Perform Laplacian feature mapping dimensionality reduction preprocessing on the rows and columns of the frame image respectively, the purpose of which is to reduce the data dimension and increase the detection rate. First of all, each frame of image can be digitized into an m×n matrix A. Each element in matrix A represents a pixel in the picture. The frame image has a total of mn pixels. The size of m and n is determined by the test video. The resolution is determined. In order to achieve dimensionality reduction of Laplacian feature mapping, image A is rearranged and combined into the following vector form

z=[A _v1 A _v2 … A _vn ] ^T

Among them, A _v1 A _v2 … A _vn represents the column vector of image A. Then put z into the historical data set for Laplacian feature map dimensionality reduction processing.

First, construct the adjacency matrix W:

Then, according to the adjacency matrix W, calculate the diagonal matrix D

D _ii =∑ _j w _ij

Finally, the Laplace matrix L=D-W is obtained, and the following minimization problem is finally solved

Lu=λDu

Step 5: Use the image after the dimensionality reduction process as the input of the machine learning classifier to obtain the detection result. Among them, the three-layer neural network is selected as the training model of the machine learning classifier, and the training data

It is obtained from the marked historical test data, where y _l =1 indicates that the _{quality of z kl} is unqualified, and y _l =0 indicates that the _{quality of z kl} is qualified.

_h [omega] corresponding to the offset of each hidden layer neuron, ω _hi corresponds to the neuron input u _i t _h, weightings, σ is the activation function

Finally, the output layer is composed of 2 neurons, which represent qualified and unqualified respectively. Its expression is

the weight of. _{Finally, the weights ω hi} , v _jh and bias ω _h , v _{j are} obtained by training by solving the following optimization problem

Where y _lj is the label of the sample A _kl.

Finally, the output image of the graphics card interface of the product to be tested is used as input, and the machine learning classifier outputs qualified or unqualified classification results.

As described above, the present invention can be implemented well. The above-mentioned embodiments are only typical embodiments of the present invention, and are not used to limit the scope of implementation of the present invention. The scope of protection claimed by the claims of the present invention is covered.

The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited to this. Any person skilled in the art can easily think of changes or changes within the technical scope disclosed in the present invention. All replacements should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

A graphics interface machine testing method based on Laplacian feature mapping learning, which is characterized by:

Step 1: Connect the product to be tested to the video capture device with the corresponding type of wire;

Step 2: Play the test video, the video collector collects the output information of the display interface, and uploads it to the industrial host computer;

Step 3: The industrial host computer decodes the graphics card interface information and saves the frame image;

Step 4: Perform Laplacian feature map dimensionality reduction preprocessing on the rows and columns of the frame image respectively;

First of all, each frame of image can be digitized into an m×n matrix A. Each element in matrix A represents a pixel in the picture. The frame image has a total of mn pixels. The size of m and n is determined by the test video. Determined by the resolution; in order to achieve dimensionality reduction of Laplacian feature mapping, the image A is rearranged and combined into the following vector form

z=[A v1 A v2 … A vn ] T

Among them, A v1 A v2 … A vn represents the column vector of image A; then z is put into the historical data set for Laplacian feature map dimensionality reduction processing;

Step 5: Use the analyzed image as the input of the machine learning classifier to obtain the detection result; select the three-layer neural network as the training model of the machine learning classifier, and the training data S = (z kl , y l ) is marked Good historical test data is obtained, where y l =1 means that the quality of z kl is unqualified, and y l =0 means that the quality of z kl is qualified;

The input layer of the neural network is composed of k neurons, which corresponds to the k elements of image A after dimensionality reduction. The hidden layer is composed of p neurons, and its output is

h [omega] corresponding to the offset of each hidden layer neuron, ω hi corresponds to the neuron input u i t h, weightings, σ is the activation function
Finally, the output layer is composed of 2 neurons, which represent qualified and unqualified respectively; its expression is

Where v j is the offset of the corresponding output, and v jh is the corresponding input t h to the output
the weight of.
The graphics card interface machine testing method based on Laplacian feature mapping learning according to claim 1, characterized in that: in step 4

The objective function for minimizing the Laplacian feature map is

in
Is the point after z (i) is reduced to k dimension, w ij is the weight of the connection between the measurement samples z (i) and z (j) ; w ij is determined according to whether the two sample points are close, first use KNN The method determines whether to set an edge connection between the sample z (i) and z (j) . If z (i) is among the k nearest neighbors of z (j) , then z (i) and z (j) are connected, k is a preset value, or set an appropriate ε, connect the nodes with ||z (i) -z (j) || 2 ≤ ε; then determine the weight and use the Heat kernel function to connect The weight of the node is set to
Here t is a preset value, or let t=∞, set the weights of all connected nodes as w ij =1, and other unconnected ones are 0; finally, a symmetric adjacency matrix W can be obtained.
The method for testing a graphics card interface machine based on Laplacian feature mapping learning according to claim 2, characterized in that: the objective function can be expressed as the following quadratic form after sorting

Where u = (u (1) , u (2) ,..., u (n) ) T , L = DW is the Laplacian matrix, D is a diagonal matrix, satisfying D ii ＝Σ j w ij , W Is a symmetric adjacency matrix, and the Laplacian matrix L is positive semi-definite;

The final solution is the following minimization problem

u T Du = 1

Among them, the constraint u T Du = 1 avoids the impact of scaling, and the vector u that minimizes the objective function is given by the minimum eigenvalue solution of the generalized eigenvalue problem:

Lu = λDu (4)

The eigenvector corresponding to the non-zero eigenvalue obtained by the solution is the output after dimensionality reduction.
The graphics card interface machine testing method based on Laplacian feature mapping learning according to claim 1, characterized in that step 5: Finally, the weights ω hi , v jh and bias ω h are obtained by training by solving the following optimization problem, v j

Where y lj is the label of the sample A kl , and N is the number of samples of historical data; finally, the output image of the graphics card interface of the product to be tested is used as input to obtain the final detection result.