Evaluation of Pooling Layers in Convolutional Neural Network for Script Recognition

Abstract

This paper investigates the suitable position and number of pooling layers in Convolutional Neural Network (CNN) for script recognition from scene images. A common practice of CNN for object recognition is to position a convolve layer alternately with a pooling layer followed by a few layers of fully connected layers. We re-evaluate this basic principle by examining the position of pooling layer after every convolve layer, reducing and increasing its numbers. Experimental results on MLe2e dataset for script recognition show that a CNN with less number of pooling layers and non-overlapping pooling stride can reach excellent percentage of accuracy compared to alternating convolve layer with pooling layer.