Recognizing handwritten single digits and digit strings using deep architecture of neural networks

Abstract

Automatic handwriting recognition of digits and digit strings, are of real interest commercially and as an academic research topic. Recent advances using neural networks and especially deep learning algorithms such as convolutional neural nets present impressive results for single digit recognition. Such results enable developing efficient tools for automatic mail sorting and reading amounts and dates on personal checks. Artificial- Neural-Networks is a powerful technology for classification of visual inputs in many fields due to their ability to approximate complex nonlinear mappings directly from input samples. In this paper we present an approach compromising between the full connectivity of traditional Multi Layer Neural Network trained by Back Propagation and deep architecture. This enables, reasonable training time using a four hidden layers Neural Network and keeps high recognition rates. Pre-trained layers using sparse auto encoders with predefined sequences of training process and rounds, are used to train the net to attain high recognition rates. We have extended the training set to include CVL, MNIST and manually crafted images of single digits from the ORAND-CAR and a private collection of bank checks. Sliding windows technique is used to handle digit strings recognition and obtain encouraging results on CVL and ORAND-CAR benchmarks and our private collection of local bank checks.