Dimensionality Reduction in Handwritten Digit Recognition

For visualization, the concept of dimension is normally enclosed to 2–3 degrees in individuals. A computing node can extend it significantly. However, any increase in the number of dimensions usually introduces an extra computational burden, and it becomes more challenging to extract the exact information. Therefore, dimensionality reduction methods are an increasingly important area of study to help identify methods to mitigate challenges associated with high-dimensional feature sets. Handwritten digit recognition is one of the most relevant fields of study due to the variety of issues faced such as the age of texts, the professional context and norms in which the text is written, and individual differences in writing styles. Research on handwritten digit recognition using various algorithms has been conducted in a variety of languages. In the Bangla character set, there are ten digits. Due to geometry, complicated forms, and similarities between the individual numerals, individual characters are difficult to identify. Also, there are limited open datasets available to researchers to conduct Bangla digit recognition upon. This work discusses dimensionality reduction techniques used in the Bangla Handwritten Digit dataset NumtaDB. Principal Component Analysis (PCA), Neighborhood Component Analysis (NCA), and Linear Discriminant Analysis (LDA) algorithms are examined as feature extraction techniques. CNN is a deep learning technique to classify the input automatically. Over the years, CNN has found a good grip over classifying images for computer visions and now it is being used in other domains too. The numeric digits are then classified using CNN utilizing the lower dimension vectors acquired. These models can recognize most of the digits successfully with a satisfactory level of performance for identifying different digits.