Object recognition using binary image quantization and hough kernels

A system for training at least one general purpose computing device to recognize objects in an input image of a scene, comprising:

• at least one general purpose computing device;
• and
• a computer program comprising program modules executable by the at least one computing device, wherein the at least one computing device is directed by the program modules of the computer program to,
• generate training images depicting a surface of interest of an object it is desired to recognized in said input image of the scene, wherein said generating comprises,
• inputting a model image that depicts the surface of interest of the object,
• creating a base training image depicting the Portion of the model image containing said surface of interest, and
• synthesizing training images from the base image which depict the object's surface of interest in different orientations, wherein said synthesizing comprises,
• (a) synthetically pointing a normal vector of the object's surface of interest at one of a prescribed number of nodes of a tessellated hemisphere defined as overlying the surface of interest,
• (b) simulating a paraperspective projection representation of the surface of interest, and
• (c) repeating sub-module (a) and (b) for each of the remaining nodes of the tessellated hemisphere;
• create a set of Prototype edge features which collectively represent the edge pixel patterns encountered within a sub-window centered on each pixel depicting an edge of the object in the training images, and
• define a Hough kernel for each prototype edge feature, wherein a Hough kernel for a particular prototype edge feature is defined by a set of offset vectors representing the distance and direction, from each edge pixel having a sub-window associated therewith that has an edge pixel pattern best represented by the Prototype edge feature, to a prescribed reference point on the surface of interest of the object, wherein said offset vectors are represented in the Hough kernel as originating at a central point thereof, wherein
• said set of prototype edge features and said Hough kernels associated therewith are used to recognize said object in the input image and to identify the location of the object in the input image.

A system for recognizing objects in an input image of a scene, said system having a set of prototype edge features and a set of Hough kernels associated therewith each of which is comprised of a plurality of offset vectors originating from a central point of the associated Hough kernel, said system comprising:

• at least one general purpose computing device;
• and
• a computer program comprising program modules executable by the at least one computing device, wherein the at least one computing device is directed by the program modules of the computer program to,
• identify for each pixel in the input image depicting an edge, which prototype edge feature best represents the edge pixel pattern exhibited within a sub-window centered on the input image edge pixel under consideration,
• for each pixel location in the input image, identify how many offset vectors terminate at that location from Hough kernels centered on each edge pixel location of the input image, wherein the Hough kernel centered on each pixel location is the Hough kernel associated with the prototype edge feature best representing the edge pixel pattern exhibited within said sub-window centered on that input image edge pixel location, and
• declare the object to be present in the input image if any of the pixel locations in the input image have a quantity of offset vectors terminating thereat that exceed a detection threshold which is indicative of the presence of the surface of interest of the object.

A computer-readable memory for recognizing objects in an input image of a scene, comprising:

• a computer-readable storage medium;
• and
• a computer program comprising program modules stored in the storage medium, wherein the storage medium is so configured by the computer program that it causes a computer to,
• generate training images depicting an object it is desired to recognized in said input image of the scene,
• for each training image,
• detect edge pixels and producing an edge pixel training image representing the training image,
• define a raw edge feature associated with each edge pixel, said raw edge feature comprising a sub-window in the edge pixel training image of a prescribed size which is centered on an edge pixel, and
• assign an offset vector to each raw edge feature, said offset vector defining the distance and direction from the center edge pixel of a raw edge feature to a pixel location in the edge pixel training image corresponding to a prescribed reference point on the object in the training image,
• create a prescribed number of prototype edge features each of which represents a group of the raw edge features and assigning a unique index number to each prototype edge feature, wherein each raw edge feature in a group of raw edge features represented by a prototype edge feature is more similar to that prototype edge feature than any of the other prototype edge features,
• produce an indexed training image from each edge pixel training image by assigning an appropriate prototype index number to each edge pixel location in the edge pixel training image, wherein the appropriate index number is the index number assigned to the prototype edge feature which is most similar to the raw edge feature associated with the edge pixel at the edge pixel location under consideration,
• create a Hough kernel for each prototype edge feature, wherein a Hough kernel for a particular prototype edge feature comprises every offset vector associated with a raw edge feature in the group of raw edge features represented by the prototype edge feature, and wherein each of said offset vectors originates at a central point of the Hough kernel,
• produce an indexed input image by,
• detecting edge pixels and producing an edge pixel input image representing the input image,
• defining a raw edge feature associated with each edge pixel, and
• assigning an appropriate prototype index number to each edge pixel location in the edge pixel input image, wherein the appropriate index number is the index number assigned to the prototype edge feature which is most similar to the raw edge feature associated with the edge pixel at the edge pixel location under consideration,
• generate a voting image from the indexed input image by,
• for each pixel location in the indexed input image, identifying how many offset vectors terminate at that pixel location from Hough kernels centered on each pixel of the indexed input image having a prototype index number assigned thereto, wherein the Hough kernel centered on each pixel is the Hough kernel associated with the prototype edge feature to which the index number of that pixel has been assigned, and
• assigning a vote count to each pixel of the indexed input image indicative of how many of said offset vectors terminated at that pixel location,
• ascertain if any of the vote counts in the voting image exceed a prescribed detection threshold which is indicative of the presence of the object being recognized in the input image, and
• declare the object to be present in the input image if the detection threshold is exceeded by any of the vote counts of the voting image pixels and designate the pixel location associated with the pixel having the largest vote count exceeding the threshold as corresponding to the pixel location of said prescribed reference point of the object being recognized in the input image.

A computer-implemented process for computing the similarity between a pair of edge features, wherein an edge feature represents the edge pixel patterns encountered within a sub-window centered on a pixel of an image, comprising using a computer to perform the following process actions:

• computing a grassfire transform from each edge feature, such that each edge pixel in the edge features is assigned a “0” and each non-edge pixel is assigned an integer number that increases the farther away the non-edge pixel is from the nearest edge pixel;
• identifying all the pixel locations in the first edge feature that correspond to edge pixel locations in the second edge feature;
• summing the grassfire values assigned to the identified pixel locations in the first edge feature to create a first edge feature grassfire sum;
• identifying all the pixel locations in the second edge feature that correspond to edge pixel locations in the first edge feature;
• summing the grassfire values assigned to the identified pixel locations in the second edge feature to create a second edge feature grassfire sum;
• and
• adding the first edge feature grassfire sum to the second edge feature grassfire sum to create a combined grassfire sum, wherein the smaller the combined grassfire sum, the more similar the first edge feature is to the second edge feature.

A computer-implemented process for computing a composite edge feature from a plurality of edge features, wherein an edge feature represents the edge pixel patterns encountered within a sub-window centered on a pixel of an image, comprising using a computer to perform the following process actions:

• (a) computing an edge pixel mean by dividing the sum of the number of edge pixels in each of the edge features by the total number of edge features;
• (b) computing a grassfire transform from each edge feature, such that each edge pixel in the edge features is assigned a “0” and each non-edge pixel is assigned an integer number that increases the farther away the non-edge pixel is from the nearest edge pixel;
• (c) respectively summing the grassfire values in each corresponding pixel location of the edge features to create a summed grassfire-transformed edge feature;
• (d) identifying locations in the summed grassfire-transformed edge feature having the lowest grassfire value and recording the quantity of these locations;
• (e) identifying locations in the summed grassfire-transformed edge feature having the next higher grassfire value and recording the quantity of these locations;
• (f) adding the number of pixel locations having the lowest grass fire value to the number of pixel locations having the next higher grassfire value to produce a combined pixel location sum;
• (g) comparing the number of pixel locations having the lowest grassfire value, and the combined pixel location sum, to the edge pixel mean;
• (h) whenever the number of pixel locations having the lowest grassfire value is closer to the edge pixel mean than the combined pixel location sum, designate the pixel locations having the lowest grassfire value as edge pixels in the composite edge feature;
• (i) whenever the combined pixel location sum is closer to the edge pixel mean than the number of pixel locations having the lowest grassfire value, and the mean is less than the combined pixel location sum, designate the pixel locations associated with the combined pixel location sum as edge pixels in the composite edge feature;
• (j) whenever the combined pixel location sum is closer to the edge pixel mean than the number of pixel locations having the lowest grassfire value, and the mean is greater than the combined pixel location sum, identifying locations in the summed grassfire-transformed edge feature having the next higher, previously unconsidered grassfire value, and recording the quantity of these locations;
• (k) adding the number of pixel locations associated with said next higher, previously unconsidered grassfire value to the last previously computed combined pixel location sum to produce a current combined pixel location sum;
• (l) comparing the number of pixel locations associated with the last previously-computed combined pixel location sum, and the current combined pixel location sum, to the edge pixel mean;
• (m) whenever the last previously-computed combined pixel location sum is closer to the edge pixel mean than the current combined pixel location sum, designating the pixel locations associated with the last previously-computed combined pixel location sum as edge pixels in the composite edge feature;
• (n) whenever the current combined pixel location sum is closer to the edge pixel mean than the last previously-computed combined pixel location sum, and the mean is less than the current combined pixel location sum, designate the pixel locations associated with the current combined pixel location sum as edge pixels in the composite edge feature;
• (o) whenever the current combined pixel location sum is closer to the edge pixel mean than the last previously-computed combined pixel location sum, and the mean is greater than the current combined pixel location sum, identifying locations in the summed grassfire-transformed edge feature having the next higher, previously unconsidered grassfire value, recording the quantity of these locations, and repeat process actions (k) through (n).