Image inspection/recognition method, method of generating reference data for use therein, and apparatuses therefor
First Claim
Patent Images
1. An image inspection/recognition method comprising the steps of:
- (1) obtaining feature images of predetermined plural features each of a plurality of training images;
(2) quantizing feature values at respective feature points of the feature images with a predetermined quantization step;
(3) obtaining a histogram of feature values of the quantized feature images of each of said plural features at each of the respective feature points;
(4) determining weights for respective feature points of each of said plural features based on the feature value histograms to define reference data of each of said plural features having the weights at the respective feature points;
(5) extracting an image feature of at least one of said plural features from an input image to form a feature image of the input image;
(6) quantizing feature values at respective feature points of said feature image of the input image with the predetermined quantization step to obtain a quantized feature image of the input image; and
(7) calculating the similarity or difference at respective feature points between said quantized feature image of the input image and the reference data of each of said plural features and combining the similarity or different for said plural features with one another with predetermined feature weights to produce weighted similarity or difference;
said step (7) comprising obtaining weighted difference by subjecting said quantized feature image of the input image to a weighted Hough transform operation by voting the weights of said feature points in said reference data to a parameter space of positions for each quantized level to obtain a Hough plane and calculating the difference between said Hough plane and a reference Hough plane obtained from said quantized feature images of said training images.
3 Assignments
0 Petitions
Accused Products
Abstract
An image of a part is used to judge whether it is good or no good. A predetermined intensity, edge or similar feature image is extracted from the inspected image and is quantized. The quantized feature image is obtained from a training image in advance, and the feature is subjected to a generalized Hough transform operation which refers to the weight of each feature point for each feature value in the image and vote the feature point weight. The similarity is obtained from the results of the transform operation, and a good/no good judgement is made depending upon whether the similarity is above or below a threshold value.
134 Citations
21 Claims
-
1. An image inspection/recognition method comprising the steps of:
-
(1) obtaining feature images of predetermined plural features each of a plurality of training images; (2) quantizing feature values at respective feature points of the feature images with a predetermined quantization step; (3) obtaining a histogram of feature values of the quantized feature images of each of said plural features at each of the respective feature points; (4) determining weights for respective feature points of each of said plural features based on the feature value histograms to define reference data of each of said plural features having the weights at the respective feature points; (5) extracting an image feature of at least one of said plural features from an input image to form a feature image of the input image; (6) quantizing feature values at respective feature points of said feature image of the input image with the predetermined quantization step to obtain a quantized feature image of the input image; and (7) calculating the similarity or difference at respective feature points between said quantized feature image of the input image and the reference data of each of said plural features and combining the similarity or different for said plural features with one another with predetermined feature weights to produce weighted similarity or difference; said step (7) comprising obtaining weighted difference by subjecting said quantized feature image of the input image to a weighted Hough transform operation by voting the weights of said feature points in said reference data to a parameter space of positions for each quantized level to obtain a Hough plane and calculating the difference between said Hough plane and a reference Hough plane obtained from said quantized feature images of said training images.
-
-
2. An image inspection/recognition method comprising the steps of:
-
(1) obtaining feature images of predetermined plural features each of a plurality of training images; (2) quantizing feature values at respective feature points of the feature images with a predetermined quantization step; (3) obtaining a histogram of feature values of the quantized feature images of each of said plural features at each of the respective feature points; (4) determining weights for respective feature points of each of said plural features based on the feature value histograms to define reference data of each of said plural features having the weights at the respective feature points; (5) extracting an image feature of at least one of said plural features from an input image to form a feature image of the input image; (6) quantizing feature values at respective feature points of said feature image of the input image with the predetermined quantization step to obtain a quantized feature image of the input image; and (7) calculating the similarity or difference at respective feature points between said quantized feature image of the input image and the reference data of each of said plural features and combining the similarity or different for said plural features with one another with predetermined feature weights to produce weighted similarity or difference; wherein in said step (4) the sum total of the weights of said feature points in each of said plural features is normalized with the number of feature points.
-
-
3. An image inspection/recognition method comprising the steps of:
-
(1) extracting an image feature from an input image and forming a feature image; (2) quantizing said feature image with a predetermined quantization step to obtain a quantized feature image; and (3) calculating the similarity between each feature point in said quantized feature image and reference data of the corresponding features obtained with training images of the same category; said step (3) comprising; (4) generating from said quantized feature image a multiple bit plane composed of bit planes corresponding to respective quantized levels of said quantized feature images; (5) subjecting predetermined ones of said bit planes to a generalized Hough transform operation through the use of said reference data; and (6) calculating said similarity from said Hough transformed bit planes.
-
-
4. An image inspection/recognition method comprising the steps of:
-
(1) obtaining feature images of predetermined plural features each of a plurality of training images; (2) quantizing feature values at respective feature points of the feature images with a predetermined quantization step; (3) obtaining a histogram of feature values of the quantized feature images of each of said plural features at each of the respective feature points; (4) determining weights for respective feature points of each of said plural features based on the feature value histograms to define reference data of each of said plural features having the weights at the respective feature points; (5) extracting an image feature of at least one of said plural features from an input image to form a feature image of the input image; (6) quantizing feature values at respective feature points of said feature image of the input image with the predetermined quantization step to obtain a quantized feature image of the input image; and (7) calculating the similarity or difference at respective feature points between said quantized feature image of the input image and the reference data of each of said plural features and combining the similarity or different for said plural features with one another with predetermined feature weights to produce weighted similarity or difference; wherein said step (1) further includes a step of deciding whether the contrast of said plurality of training images of the same category is higher than a predetermined value or not; and
wherein when the contrast is higher than said predetermined value, a region feature image is extracted as said feature image in said step (1) and a weighted normalized correlation operation is performed in said step (7) to obtain said weighted similarity, and wherein when said contrast is lower than said predetermined value, a contour feature image is extracted as said feature image in said step (1) and a weighted generalized Hough transform operation is performed in said step (7) to obtain said weighted similarity.
-
-
5. An image inspection/recognition method comprising the steps of:
-
(1) obtaining feature images of predetermined plural features each of a plurality of training images; (2) quantizing feature values at respective feature points of the feature images with a predetermined quantization step; (3) obtaining a histogram of feature values of the quantized feature images of each of said plural features at each of the respective feature points; (4) determining weights for respective feature points of each of said plural features based on the feature value histograms to define reference data of each of said plural features having the weights at the respective feature points; (5) extracting an image feature of at least one of said plural features from an input image to form a feature image of the input image; (6) quantizing feature values at respective feature points of said feature image of the input image with the predetermined quantization step to obtain a quantized feature image of the input image; and (7) calculating the similarity or difference at respective feature points between said quantized feature image of the input image and the reference data of each of said plural features and combining the similarity or different for said plural features with one another with predetermined feature weights to produce weighted similarity or difference; wherein said method further includes the steps of; performing a weighted generalized Hough transform operation on contour feature images of pluralities of different categories and measuring distance between intra-category similarity distribution and inter-category similarity distribution; performing a weighted normalized correlation operation on region feature images of said pluralities of different categories and measuring distance between intra-category similarity distribution and inter-category similarity distribution; selecting the one of said weighted generalized Hough transform operation and said weighted normalized correlation operation which provided a larger distance than the other; and said step (7) performs the selected operation to obtain said similarity or difference of feature points.
-
-
6. An image inspection/recognition method comprising the steps of:
-
(1) obtaining feature images of predetermined plural features each of a plurality of training images; (2) quantizing feature values at respective feature points of the feature images with a predetermined quantization step; (3) obtaining a histogram of feature values of the quantized feature images of each of said plural features at each of the respective feature points; (4) determining weights for respective feature points of each of said plural features based on the feature value histograms to define reference data of each of said plural features having the weights at the respective feature points; (5) extracting an image feature of at least one of said plural features from an input image to form a feature image of the input image; (6) quantizing feature values at respective feature points of said feature image of the input image with the predetermined quantization step to obtain a quantized feature image of the input image; and (7) calculating the similarity or difference at respective feature points between said quantized feature image of the input image and the reference data of each of said plural features and combining the similarity or different for said plural features with one another with predetermined feature weights to produce weighted similarity or difference; said steps (2) through (4) being repeated for each of different quantization steps for object training images to obtain distribution of maximum frequencies as similarity in the histograms of each of feature images of first and second categories of the training images, said method further including a step of determining one of the quantization steps which maximized the distance between the distribution of similarities in feature images of said first and second categories; and said step (7) quantizes said plurality of training images through the use of a quantization width which maximizes the distance between the distribution of similarity or difference of the first and second categories of feature images. - View Dependent Claims (8)
-
-
7. An image inspection/recognition method comprising the steps of:
-
(1) obtaining feature images of predetermined plural features each of a plurality of training images; (2) quantizing feature values at respective feature points of the feature images with a predetermined quantization step; (3) obtaining a histogram of feature values of the quantized feature images of each of said plural features at each of the respective feature points; (4) determining weights for respective feature points of each of said plural features based on the feature value histograms to define reference data of each of said plural features having the weights at the respective feature points; (5) extracting an image feature of at least one of said plural features from an input image to form a feature image of the input image; (6) quantizing feature values at respective feature points of said feature image of the input image with the predetermined quantization step to obtain a quantized feature image of the input image; and (7) calculating the similarity or difference at respective feature points between said quantized feature image of the input image and the reference data of each of said plural features and combining the similarity or different for said plural features with one another with predetermined feature weights to produce weighted similarity or difference; said steps (2) and (3) being repeated for each of a plurality of different quantization steps to obtain mean integrated squared errors of feature value histograms of feature points in said feature images of said plurality of training images of a predetermined category, and said step (6) quantizing said feature image through the use of a quantization width which minimizes the mean integrated squared errors. - View Dependent Claims (9, 10, 11)
-
-
12. An image inspection/recognition method comprising the steps of:
-
(1) obtaining feature images of predetermined plural features each of a plurality of training images; (2) quantizing feature values at respective feature points of the feature images with a predetermined quantization step; (3) obtaining a histogram of feature values of the quantized feature images of each of said plural features at each of the respective feature points; (4) determining weights for respective feature points of each of said plural features based on the feature value histograms to define reference data of each of said plural features having the weights at the respective feature points; (5) extracting an image feature of at least one of said plural features from an input image to form a feature image of the input image; (6) quantizing feature values at respective feature points of said feature image of the input image with the predetermined quantization step to obtain a quantized feature image of the input image; and (7) calculating the similarity or difference at respective feature points between said quantized feature image of the input image and the reference data of each of said plural features and combining the similarity or different for said plural features with one another with predetermined feature weights to produce weighted similarity or difference; said step (7) determining said plurality of different feature images by the steps of; (8) obtaining pluralities of quantized feature images from a plurality of training images of a plurality of categories; (9) obtaining the weighted similarity or difference of training images through the use of said reference data obtained in said step (4); (10) obtaining the similarity or difference of each of said feature images based on the similarity or difference of feature points in each feature image obtained in said step (9); (11) determining weight of each feature from said similarities or differences of the feature images of the same feature in different categories obtained in said step (10); (12) combining plural ones of the features for each of the different categories to obtain similarity or difference of each category; (13) determining the distance in similarity or difference between said categories; and (14) changing said combination of features until said distance satisfies a predetermined condition. - View Dependent Claims (13)
-
-
14. A method of generating reference data for image inspection/recognition, comprising:
-
(1) extracting a plurality of image features from each of a plurality of training images of a same first category and forming first feature images; (2) quantizing said first feature images with a predetermined quantization step to obtain first quantized feature images; (3) obtaining a distribution of feature values at each feature point in said first quantized feature images of each feature; (4) determining the weight of said each feature point on the basis of said distribution at said each feature point in each of said first feature images to obtain a reference feature image having weighted feature values at respective feature points; (5) generating quantized second and third feature images from a second training image of said first category and a third training image of a second category for each of the features; (6) obtaining the similarity or difference in each feature between said reference feature image and each of said second and third feature images by calculating a weighted correlation in feature value between said reference feature image and each of said second and third feature images; (7) determining a weight of each of said features on the basis of said similarities or differences in said each feature corresponding to said second and third feature images of said first and second categories, respectively; (8) selecting desired ones of said features, combining the reference images of the selected features through the corresponding weights and obtaining the distance between said similarities or differences corresponding to said first and second categories; and (9) changing the selection of said features, repeating said step (8) until said distance becomes sufficiently large and determining the feature points and the corresponding weights as reference data of a final combination. - View Dependent Claims (15, 16, 17)
-
-
18. An apparatus for generating reference data for image inspection/reception, comprising:
-
an image memory for storing training images of a first category; means for extracting images of different features from each of said training images and forming feature images; means for quantizing each of said feature images with a predetermined quantization step; feature image memory means for storing said quantized feature images; means for obtaining from said quantized feature images of each feature a distribution of feature values of the quantized feature images at each corresponding feature point; feature point value distribution memory means for storing said distributions of feature values for each of said features in each of said first and second categories; means for determining a weight for each of the feature points from corresponding ones of said feature point value distributions for each of said first and second categories; reference table memory means for storing said feature points and corresponding weights as a reference image for each feature image; weight similarity and difference calculating means for performing a correlation operation between each of training feature images in each of said first and second categories and said reference image stored in said reference table memory means to obtain a weighted similarity or difference of each training feature image; similarity and difference distribution memory means for storing the distribution of said weighted similarities or differences of the training feature images for each of said first and second categories; means for obtaining the feature weight from said similarity or difference distributions of the corresponding features in said first and second categories; means for determining the combination of different features in both said first and second categories and for obtaining the distance between the weighted similarity or difference in said first category and the corresponding weighted similarity or difference in said second category; and means for changing said feature combination until said distance satisfies a predetermined condition and for determining combined weights at respective feature points corresponding to a final feature combination as reference dam.
-
-
19. An image inspection/recognition method comprising the steps of:
-
(1) extracting an image feature from an input image and forming a feature image; (2) quantizing said feature image with a predetermined quantization step to obtain a quantized feature image; and (3) calculating the similarity between each feature point in said quantized feature image and reference data of the corresponding features obtained with training images of the same category; said step (3) comprising; (4) generating from said quantized feature image a multiple bit plane composed of bit planes corresponding to respective quantized levels of said quantized feature images; (5) subjecting each bit plane of said multiple bit plane to a Hough transform operation to form a multiple Hough plane; and (6) obtaining similarity by calculating the difference between said multiple Hough plane and said reference data used as a reference multiple Hough plane. - View Dependent Claims (20, 21)
-
Specification