Investigation on inter- and intra-class ambiguity between handwritten case-sensitive characters using customized MobileNetv2

The unconstrained writing styles and pattern variety in handwriting cause confusion among characters of the same class (intra-class) and characters of different classes (inter-class). The developers designed handwritten character recognition (HCR) model, which is specific to language script and dictionary. The ambiguity of alphabets has received very little attention, despite evidence that certain characters are more easily misinterpreted, and only a few people have explored the open-source deep architectures for HCR using transfer learning. Previous works focused on the accuracy of the HCR model, not explaining the reasons for the reduction in performance, and letters, digits recognized separately. Experiments on two widely used datasets, EMNIST and NIST SD19, demonstrate that the original database size information is required to distinguish case-sensitive (C-c, F-f, I-I, J-j, K-k, M-m, O-o, P-p, S-s, U-u, V-v, W-w, X-x, Y-y, Z-z) characters having similar pattern. The performance of MobileNetv2 on (i) EMNIST (By-class: accuracy (AR) – 87.88%), NIST SD19 (By-class: AR – 90.65%) and (ii) customized Mo-bileNetv2 EMNIST (By-class: accuracy (AR) – 87.12%), NIST SD19 (By-class: AR – 89.55 %). The results of the modified MobileNetv2 with NIST SD19 are likewise comparable to state-of-the-art and earlier results.