Information processing device, information processing method, and program

An information processing device comprising:

• a correction unit configured to correct parameters of an ensemble learner using correction information for correcting the parameters of the ensemble learner having a plurality of weak learners that receive characteristic amounts extracted from an input image and output weak hypotheses indicating whether or not a predetermined target is reflected in the input image, the correction information being calculated based on a light source illuminating the target;
• a discrimination unit configured to integrate the weak hypotheses output from the weak learners with respect to the characteristic amounts using the ensemble learner of which the parameters are corrected and discriminate whether or not the target is reflected in the input image based on an integrated value of the weak hypotheses,
• wherein the characteristic amounts are inter-pixel difference characteristic amounts that are differences between two pixel locations of the input image,
• wherein, in ensemble learning for obtaining the ensemble learner, a pixel location pair that has two pixel locations used to obtain the characteristic amount, a pixel location pair having a minimum error ratio of the weak learner as a threshold value for comparison with the characteristic amount by the weak learner, and a threshold value are obtained in each of the weak learners constituting the ensemble learner as the parameters of the ensemble learner, and the correction unit corrects the threshold value,
• and said information processing device further comprising:
• a light source estimation unit configured to estimate the light source illuminating the target from the input image;
• a shaded region recognition unit configured to recognize a shaded region on which shading occurs due to the light source in the input image based on an estimated result of the light source;
• and
• a correction information calculation unit,
• wherein the correction information calculation unit obtains a correction coefficient correcting the threshold of the weak learner of which at least one pixel location of the pixel location pair used to obtain the inter-pixel difference characteristic amount is within the shaded region among the plural weak learners constituting the ensemble learner, and outputs the correction coefficient and the shaded region as the correction information,
• wherein the correction unit corrects the threshold of the weak learner of which at least one pixel location of the pixel location pair used to obtain the inter-pixel difference characteristic amount is within the shaded region among the plural weak learners constituting the ensemble learner by virtue of the correction coefficient, and
• wherein the ensemble learning is carried out using a plurality of learning images, the learning images include positive images in which the target is reflected and negative images in which the target is not reflected, and the correction unit corrects the threshold value of the weak learner of which one pixel location is within the shaded region and the other pixel location is outside the shaded region in the pixel location pair used to obtain the inter-pixel difference characteristic amount by a multiplication value obtained by multiplying a subtraction value by a pixel value of the one pixel location, the subtraction value being obtained by subtracting the correction coefficient, which is an integer equal to or smaller than 1, from 1, in an average image of which the positive images included in the learning image are averaged.

An information processing method comprising:

• estimating a light source illuminating a predetermined target from an input image;
• recognizing a shaded region on which shading occurs due to the light source in the input image based on an estimated result of the light source;
• obtaining a correction coefficient correcting a threshold of the weak learner and providing correction information;
• correcting parameters of an ensemble learner using the correction information for correcting the parameters of the ensemble learner having a plurality of weak learners that receive characteristic amounts extracted from the input image and output weak hypotheses indicating whether or not the predetermined target is reflected in the input image;
• and
• integrating the weak hypotheses output from the weak learners with respect to the characteristic amounts using the ensemble learner of which the parameters are corrected and discriminating whether or not the target is reflected in the input image based on an integrated value of the weak hypotheses,
• wherein the ensemble learning is carried out using a plurality of learning images which include positive images in which the target is reflected and negative images in which the target is not reflected, and
• wherein the correcting corrects the threshold value of the weak learner of which one pixel location is within the shaded region and the other pixel location is outside the shaded region in a pixel location pair by a multiplication value obtained by multiplying a subtraction value by a pixel value of the one pixel location, the subtraction value being obtained by subtracting the correction coefficient, which is equal to or smaller than 1, from 1, in an average image of which the positive images included in the learning image are averaged.

A non-transitory computer readable medium having stored thereon a program for causing a computer to act as units comprising:

• a light source estimation unit configured to estimate a light source illuminating a predetermined target from an input image;
• a shaded region recognition unit configured to recognize a shaded region on which shading occurs due to the light source in the input image based on an estimated result of the light source;
• a correction information calculation unit configured to obtain a correction coefficient correcting a threshold of the weak learner and to provide correction information;
• a correction unit configured to correct parameters of an ensemble learner using the correction information for correcting the parameters of the ensemble learner having a plurality of weak learners that receive characteristic amounts extracted from the input image and output weak hypotheses indicating whether or not the predetermined target is reflected in the input image, the correction information being calculated based on a light source illuminating the target;
• and
• a discrimination unit configured to integrate the weak hypotheses output from the weak learners with respect to the characteristic amounts using the ensemble learner of which the parameters are corrected and discriminate whether or not the target is reflected in the input image based on an integrated value of the weak hypotheses,
• wherein the ensemble learning is carried out using a plurality of learning images which include positive images in which the target is reflected and negative images in which the target is not reflected, and
• wherein the correction unit corrects the threshold value of the weak learner of which one pixel location is within the shaded region and the other pixel location is outside the shaded region in a pixel location pair by a multiplication value obtained by multiplying a subtraction value by a pixel value of the one pixel location, the subtraction value being obtained by subtracting the correction coefficient, which is equal to or smaller than 1, from 1, in an average image of which the positive images included in the learning image are averaged.