A Novel Siamese Network for Few/Zero-Shot Handwritten Character Recognition Tasks

Deep metric learning is one of the recommended methods for challenge supporting few/zero-shot by deep networks. It depends on building a Siamese architecture two homogeneous Convolutional Neural Networks (CNNs) distance function that can map input data from space to feature space. Instead determining class each sample, deals with existence few training samples deciding if share same identity or not. The traditional structure was built forming CNNs scratch randomly initialized weights and trained binary cross-entropy loss. Building trial error time-consuming phase. In addition, loss sometimes leads poor margins. this paper, novel network proposed applied Handwritten Character Recognition (HCR) tasks. novelties are in. 1) Utilizing transfer using pre-trained AlexNet as extractor in architecture. Fine-tuning typically faster easier than scratch. 2) Training contrastive instead cross-entropy. Contrastive helps learn nonlinear mapping enables it extracted features vector an optimal way. evaluated challenging Chars74K datasets conducting experiments. One testing few-shot while other zero-shot learning. recognition accuracy reaches 85.6% 82% few- respectively. comparison between performance conducted. results show achieves higher less time. reduces time days hours both