| 摘要 |
A system extracts a set of size invariant, rotation invariant features from pixel data of a character and sends the set of features to a statistical decision tree to effectuate automatic recognition of a character. The set of extracted features includes a first group obtained by generating arrays representable as dimensions and other geometric qualities of six minimum bounding rectangles rotated about the character, including the distances between the center points of the rotated bounding rectangles. A second group of features is extracted by generating a group of arrays representing the perimeter pixel locations, radii lengths, and direction codes. The arrays of that group are "resampled" to generate corresponding 64 point arrays. The resampled arrays of direction codes is smoothed and incremental direction codes are computed and compared to preselected thresholds to obtain features that indicate prominent convex and concave portions of the character. A third group of features is extracted by dividing a circle bounding the character into eight rings and 24 slices and counting the number of character pixels contained in each ring and in each slice. Various Fourier transforms, autocorrelations, moment calculations, and sorting operations are performed on many of the aforementioned arrays to yield further size and orientation invariant features. The distance between the centroid of the largest hole of the character and the center of the character is computed. The ratios of the perimeters of first and second holes of the character to the outer perimeter are computed. |
