Long Short Term Memory Recurrent Neural Network Application To Handwritten Recognition

Handwriting is a skill that is personal to individuals. Handwriting consists of artificialgraphical marks on a surface; its purpose is to communicate something, achieved by virtue ofthe mark’s conventional relation to language. Handwriting recognition is the task oftransforming a language represented in the spatial form of graphical marks into its symbolicrepresentation. Despite more than30years of handwriting recognition research, Recognizingthe unconstrained sequence is still a challenge task. The difficulty of segmenting cursivescript has led to the low recognition rate. Therefore, this thesis will focus on how to classifythe online unconstrained English handwriting words without segmentation.The main work and contribution of this thesis includes:1.implement the pre-process procedure and feature extraction of the onlineunconstrained English handwriting. The pre-process includes incline regularization, sizenormalization, re-samplize, smoothing, and virtual stroke generation.2.we will apply a novel type of recurrent neural network, termed as Bidirectional LongShort-Term M emory (BLSTM) architecture. BLSTM is equipped with multi-active neurons,including input gate, output gate, forget gate and instant connection. This has enabled BLSTMdeal with invariant sequence and maintain long memory.3.a sequence concatenating technique called Connecionist Temporal Classification (CTC)is applied. Besides, Three extended decoding algorithm: Levenshtein Distance(LD), fullpath(FD), max path(M D) are proposed insighted by HMM to have a lexicon-basedclassification.4. Three dataset are used for experiment: IRONOFF, Unipen-CDROW, Unipen-ICROW.Besides, in terms of practical words dataset, we apply augmentation skill to enlarge the size ofthe dataset, enhance the variety of the samples, which has boosted the performance of theneural network.Finally, the model BLSTM-CTC-FP has proved to be robust to lexicon-based recognitionand reduce50%error compared to the existing best model.