Recently, the triangle features have been applied in digit recognition by adopting the angle as a part of the features. Most of the studies in digit recognition area which applied these features have given impressive result. However, the issue of big gap values that occurred between angle, ratio and gradient has given big impact to the accuracy of result. Therefore, we introduce our proposed method which is data normalization that has adopted the nature of triangle geometry in order to resolve the issue. Besides, we have applied other techniques such as Z-score, Minimax and LibSVM function in the experiment. There are four digit datasets used which are HODA, MNIST, IFCHDB and BANGLA. The result of classification have shown our proposed method have given better result compared to other technique as aforementioned.
Key words: Triangle Features, Triangle Geometry, Feature Extraction, Feature Normalization, Feature Scaling
|
