关键词: 特征提取, 距离映射, 相对链码, 形状上下文, 支持向量机

Abstract: Due to the mutability of unstrained or handwritten digits, most algorithms in previous study either forfeited easy implementation for high accuracy, or vice versa. This paper proposed a new feature descriptor named Distance Distribution Histogram (DDH) and adapted Shape Accumulate Histogram (SAH) feature descriptor based on shape context which was not only easy to implement, but also was robust to noise and distortion. To make hybrid features more comprehensive, some other adapted topological features were combined. The new congregated features were complementary as they were formed from different original feature sets extracted by different means. What's more, they were not complicate. Meanwhile, three Support Vector Machine (SVM) with different feature vector were used as classifier and their results were integrated to get the final classification. The average accurate rate of several experiments based on self-established data sets, MNIST and USPS is as high as 99.21%, which demonstrates that the proposed algorithm is robust and effective.

Key words: feature extraction, distance map, relative chain code, shape context, Support Vector Machine (SVM)