Start-End Writing Integration with Convolutional Neural Network for Bengali Handwritten Numeral Recognition

Convolutional neural network (CNN) based methods have been very successful for handwritten numeral recognition (HNR) applications. However, CNN seems to misclassify similar shaped numerals (i.e., the silhouette of the numerals that look same). This paper presents an enhanced HNR system to improve the classification accuracy of the similar shaped handwritten numerals incorporating the terminals points (i.e., the start and end positions) with CNN’s recognition. Start-End Writing Measure (SEWM) and its integration with CNN is the main contribution of this research. Traditionally, CNN’s classification (i.e., $\mathrm{C}\mathrm{L}_{\mathrm{C}\mathrm{N}\mathrm{N}}\in\{0,1, \ldots,9\})$ for highest probability is the outcome of CNN-based system. In the proposed system, along with classification in CLCNN, its probability value (say CNN’s confidence level 0CNN) is also used as a regulating element. Parallel to CNN’s classification operation, SEWM measures start-end points of the numeral image, suggests CLSEWM numeral category for which measured start-end points are found closed to reference start-end points of the numeral class. Finally, output label or system’s classification $(\mathrm{C}\mathrm{L}\mathrm{s}_{\mathrm{y}\mathrm{s}})$ of the given numeral image is provided comparing 0CNN with predefined threshold value (o0): CNN’s outcome is considered as system’s outcome $(i.e., \mathrm{C}\mathrm{L}\mathrm{s}_{\mathrm{y}\mathrm{s}}=CLcNN)$ if $0CNN = \gt \sigma 0$; otherwise, SEWM rectifies CNN’s decision (i.e., $\mathrm{C}\mathrm{L}\mathrm{s}_{\mathrm{y}\mathrm{s}}=$CLsEWM) as CNN’s confidence is low. The proposed method is tested on rich benchmark numeral datasets of Bengali and revealed itself as a suitable HNR method.