Enhancement of Deep Architecture using Dropout/ DropConnect Techniques Applied for AHR System
Mohamed Elleuch, Adel M. Alimi, Monji Kherallah · 2018
Remarkable performance on computer vision, and especially on pattern recognition field has been known for a long time to be produced by Deep learning algorithms. It is clear that amongst the successful applications in the pattern recognition domain, Arabic handwriting recognition (AHR) is a must. In this survey, we use two deep networks: Deep Belief Network (DBN) and Convolutional Neural Networks (CNN), for Arabic handwritten script (AHS) recognition. Despite the triumph of DBN and CNN methods, over-fitting is able to take place on these networks thanks to the massive number of parameters. In order to fight over-fitting, we have deeply inquired two regularization techniques called Dropout and DropConnect. While training with the two regularization methods, a randomly chosen subsets of activations/weights are dropped. Consequently, the assessment on the HACDB database to treat character level proves shows an improvement of classification error rate once adding Dropout and DropConnect techniques.