Pincode detection using deep CNN for postal automation
International Conference Image and Vision Computing New Zealand
Postal automation has been a topic of research over a decade. The challenges and complexity involved in developing a postal automation system for a multi-lingual and multi-script country like India are many-fold. The characteristics of Indian postal documents include: multi-lingual behaviour, unconstrained handwritten addresses, structured/unstructured envelopes and postcards, being among the most challenging aspects. This paper examines the state-of-the-art Deep CNN architectures for detecting pin-code in both structured and unstructured postal envelopes and documents. Region-based Convolutional Neural Networks (RCNN) are used for detecting the various significant regions, namely Pin-code blocks/regions, destination address block, seal and stamp in a postal document. Three network architectures, namely Zeiler and Fergus (ZF), Visual Geometry Group (VGG16), and VGG M were considered for analysis and identifying their potential. A dataset consisting of 2300 multilingual Indian postal documents of three different categories was developed and used for experiments. The VGG-M architecture with Faster-RCNN performed better than others and promising results were obtained.
Sharma, Nabin; Sengupta, Abira; Sharma, Rabi; Pal, Umapada; and Blumenstein, Michael, "Pincode detection using deep CNN for postal automation" (2018). Conference Articles. 71.