A Novel Multi-Input Hybrid Deep Convolution Neural Network Methodology for Offline Writer Recognition

Abstract

Identifying a writer from a small sample of handwriting is a challenging task. In addition, it’s a key domain of research for the field of forensic investigation of documents. A novel multi-input hybrid convolution neural network (CNN) is developed to address the offline writer recognition problem involving monolingual and bilingual handwritten scripts. Model generates global features for classification by combining local CNN features with traditional handcrafted features. HOG (Histogram of Oriented Gradients), a standard hand-crafted feature descriptor, is employed. Two distinct CNN paths are utilized to extract distinct feature maps. English script benchmark dataset CVL and an in-house bilingual dataset is utilized in experiments. The bilingual dataset consists mixture of handwritten English and Hindi text. Model achieved a 98.23% accuracy rate with CVL and a 96.93% accuracy rate with an in-house bilingual dataset.