Bayesian-Driven Deep Learning and Machine Learning Approaches for Handwritten Word Recognition

Identifying handwritten words poses a complicated problem owing to the variances in handwriting styles and the possible noise and distortions present in the data. The performance of used deep learning and traditional machine learning handwriting recognition methods for recognizing handwritten words is studied in this paper and the advantages of using deep learning are discussed. In particular, neural network models such as CNN and 2DLSTM networks had a great work in this area. In contrast, we detect whole handwritten words without the need of segments into standalone characters. The CNNs are used for feature extraction, the Bidirectional LSTMs for sequence prediction and output is decoded via a CTC layer. The model was evaluated over the IAM words dataset from the IAM handwriting database, achieving 87% accuracy and a character error rate of 7.77%. In addition, we experimented with ML techniques, including SVM, RF has got an accuracy of 72% and 80% accuracies, and a Bayesian network of 61% accuracy, to provide a comparative analysis.