Defect detection method for display panel based on histogram of oriented gradient
View Patent ↗In order to detect defects in display panels by an automatic way accurately and quickly, the present invention proposes a method combining image feature extraction and classifier model. It calculates the histograms of oriented gradient (HOG) of images of the display panel collected by an industrial camera of a detection apparatus as feature vectors. Then use them as input samples to train the classifier model to recognize the defects of the display panel.
1. A defect detection method for a display panel based on histogram of oriented gradient, said method comprising:
capturing an image of a substrate of the display panel by an industrial camera;
performing detection on a region of interest (ROI) on the image, where the ROI of the image is taken as a to-be-detected image;
extracting a feature descriptor from a histogram of oriented gradient of the to-be-detected image;
inputting the feature descriptor into a trained classifier model for proceeding with recognition and classification;
outputting a detection result on a basis of a determination made by the trained classifier model,
collecting positive image samples with defects and negative image samples without defects;
performing feature extraction on histograms of oriented gradient of the positive image samples and the negative image samples; and
training a classifier with inputted features of the histograms of oriented gradient of the positive image samples and the negative image samples to obtain the trained classifier model.
2. The method according to claim 1 , before the step of performing detection on the ROI on the image, said method further comprising:
performing image preprocessing on the image of the substrate of the display panel captured by the industrial camera.
3. The method according to claim 2 , wherein the image preprocessing comprises noise reduction.
4. The method according to claim 1 , wherein the classifier comprises an extreme learning machine (ELM) classifier.
5. The method according to claim 1 , wherein the step of extracting the feature descriptor from the histogram of oriented gradient of the to-be-detected image comprises:
computing gradient for each pixel of the to-be-detected image;
establishing the histogram of oriented gradient of the to-be-detected image in unit of a cell, where a cell is consisted of a plurality of pixels;
normalizing the histogram of oriented gradient in unit of a block, where a block is consisted of a plurality of cells; and
collecting all of the blocks of the to-be-detected image as the feature descriptor of the histogram of oriented gradient of the to-be-detected image.
6. The method according to claim 5 , wherein the step of computing gradient for each pixel of the to-be-detected image performs the gradient computation by adopting a [−1, 0, 1] differential model with symmetry of first derivatives.
7. The method according to claim 5 , wherein in the step of establishing the histogram of oriented gradient of the to-be-detected image in unit of a cell, 9 intervals of gradient directions are adopted to classify gradient directions of the pixels.
8. The method according to claim 5 , wherein a normalization approach adopted in the step of normalizing the histogram of oriented gradient in unit of a block is L2-norm, which is represented by:
v
->
v
/
v
2
2
+
ɛ
2
,
where v is a vector of a non-normalized descriptor, ∥V∥k is its k-norm (k=1, 2), and ε is a sufficiently small constant which makes the denominator of the above equation not become a zero.
9. A defect detection method for a display panel based on histogram of oriented gradient, said method comprising:
collecting positive image samples and negative image samples;
performing feature extraction on histograms of oriented gradient of the positive image samples and the negative image samples;
training a classifier with inputted features of the histograms of oriented gradient of the positive image samples and the negative image samples to obtain a trained classifier model;
thoroughly scanning an image captured by a detection apparatus by adopting a detection window with a predetermined size;
extracting a feature descriptor from a histogram of oriented gradient of an image defined in the detection window, and performing the same operation for all the images defined with the detection window;
inputting each of the feature descriptors of the images defined with the detection window into the trained classifier model for proceeding with recognition and classification; and
outputting a defect detection result for each image defined in the detection window on a basis of a determination made by the trained classifier model.
10. The method according to claim 9 , wherein the classifier comprises an extreme learning machine (ELM) classifier.
11. The method according to claim 9 , wherein the step of extracting the feature descriptor from the histogram of oriented gradient of the image defined in the detection window comprises:
computing gradient for each pixel of the image defined in the detection window;
establishing the histogram of oriented gradient of the image defined in the detection window in unit of a cell, where a cell is consisted of a plurality of pixels;
normalizing the histogram of oriented gradient in unit of a block, where a block is consisted of a plurality of cells; and
collecting all of the blocks of the image defined in the detection window as the feature descriptor of the histogram of oriented gradient of the image defined in the detection window.
12. The method according to claim 11 , wherein the step of computing gradient for each pixel of the image defined in the detection window performs the gradient computation by adopting a [−1, 0, 1] differential model with symmetry of first derivatives.
13. The method according to claim 11 , wherein in the step of establishing the histogram of oriented gradient of the image defined in the detection window in unit of a cell, 9 intervals of gradient directions are adopted to classify gradient directions of the pixels.
14. The method according to claim 11 , wherein a normalization approach adopted in the step of normalizing the histogram of oriented gradient in unit of a block is L2-norm, which is represented by:
v
->
v
/
v
2
2
+
ɛ
2
,
where v is a vector of a non-normalized descriptor, ∥V∥k is its k-norm (k=1, 2), and ε is a sufficiently small constant which makes the denominator of the above equation not become a zero.
15. A defect detection method for a display panel based on histogram of oriented gradient, said method comprising:
thoroughly scanning an image captured by a detection apparatus by adopting a detection window with a predetermined size;
extracting a feature descriptor from a histogram of oriented gradient of an image defined in the detection window, and performing the same operation for all the images defined with the detection window;
inputting each of the feature descriptors of the images defined with the detection window into a trained classifier model for proceeding with recognition and classification; and
outputting a defect detection result for each image defined in the detection window on a basis of a determination made by the trained classifier model.
16. The method according to claim 15 , wherein the step of extracting the feature descriptor from the histogram of oriented gradient of the image defined in the detection window comprises:
computing gradient for each pixel of the image defined in the detection window;
establishing the histogram of oriented gradient of the image defined in the detection window in unit of a cell, where a cell is consisted of a plurality of pixels;
normalizing the histogram of oriented gradient in unit of a block, where a block is consisted of a plurality of cells; and
collecting all of the blocks of the image defined in the detection window as the feature descriptor of the histogram of oriented gradient of the image defined in the detection window.
17. The method according to claim 16 , wherein the step of computing gradient for each pixel of the image defined in the detection window performs the gradient computation by adopting a [−1, 0, 1] differential model with symmetry of first derivatives.
18. The method according to claim 16 , wherein in the step of establishing the histogram of oriented gradient of the image defined in the detection window in unit of a cell, 9 intervals of gradient directions are adopted to classify gradient directions of the pixels.
19. The method according to claim 16 , wherein a normalization approach adopted in the step of normalizing the histogram of oriented gradient in unit of a block is L2-norm, which is represented by:
v
->
v
/
v
2
2
+
ɛ
2
,
where v is a vector of a non-normalized descriptor, ∥V∥k is its k-norm (k=1, 2), and ε is a sufficiently small constant which makes the denominator of the above equation not become a zero.