A statistical-style contour-based feature extractor for digit recognition has been developed. The features are calculated from the chain code representation of the image and are used as input to a neural network classifier. The image must be binarized prior to generation of the chain code. Then the contour is smoothed by local averaging of each pixel's slope. The curvature is also calculated at every boundary pixel, and quantized into one of 5 values.
A 4x4 grid is used to partition the image into 16 equal-sized regions. The number of boundary pixels in each region with a particular slope and curvature is determined. Since there are 8 values for slope and 5 values for curvature, there are 40 possible categories per pixel. Finally, these counts of pixel categories are converted to percentages of total pixels per region. The result is a set of 640 features (40 per subregion) which describe the curvature profile of the i mage, and which constitute the input vector to a neural network classifier. The classifier is a 2-layer back propagation network with 640 input nodes, 100 nodes in the hidden layer, and 10 output nodes.
Another version of the algorithm uses the same feature extraction routine. However, instead of a complete 8x5 table to accumulate all possible categories of the pixels, the slopes and curvatures are tabulated in two separate arrays of length 8 and 5, respectively. These 208 features (13x4x4), together with 84 transition features and 4 hole features, constitute a set of 296 features. The transition features encode all possible transitions between adjacent subregions following the chaincode contour. Three hole features encode the proportion of the size of the hole with respect to the character bounding box centered at the top, center and bottom of the image. The number of holes is recorded in the last feature. This smaller set of features performed as well as the 640 features on digit recognition, and has replaced the original 640 features in almost all cases.