An Efficient Handwritten Digit Classification Framework Based on Dynamic Attention Pyramid Head, Coordinate Attention, and Adaptive EfficientNetB7 with RP-DOA Optimization

Abstract

In recent times, handwritten numeral classification has attracted enormous interest in computer vision, because recognizing the digits is critical due to a wide range of individual handwriting styles. In multiple language countries like India, there is confined research on identifying handwritten characters and numerical by using deep learning, particularly when comparing with other additional vernacular scripts. The variations in writing shape, size, style, and similarities of different characters show accurate classification, adding even more challenges and complexity. Deep structured neural networks have appeared as an advanced technology for automatic object image classification and character patterns. Although deep neural networks have provided better performance on large datasets comprising millions of raw images, using deep network models on smaller datasets remains a challenging task. Hence, it is crucial to tackle difficulties that arise in the traditional digit classification model. In order to rectify a few issues presented in the previous research works, a novel deep structured learning-based digit classification framework is developed in this work. In the initial phase, essential images required for the evaluation are gathered from standard resources. Next, the collected input images are fed into the digit classification phase. Here, Dynamic Attention Pyramid Head with Coordinate Attention Parallel Fusion-based Adaptive EfficientnetB7 (DAPH-CAPF-AENetB7) is employed to carry out the classification procedure. The parameters of DAPH-CAPF-AENetB7 are tuned through the Renovated Position-based Dollmaker Optimization Algorithm (RP-DOA), which supports achieving more accurate digit classification results. Finally, several experiments are performed in the developed framework over classical techniques to check the overall efficiency of the developed framework in different classes.