The Improvement of Character Recognition on ANPR Algorithm using CNN Method with Efficient Grid Size Reduction

Abstract

Automatic Number Plate Recognition (ANPR) is mainly divided into three steps: plate localization, character segmentation, and character recognition. Among those steps, character recognition is the most significant influencer on ANPR accuracy. One of the popular methods that have impressive performance and commonly used recently is Convolution Neural Network (CNN). However, max-pooling layers within CNN architecture are prone to lose information during the downsampling of feature maps. Our proposed method is using efficient grid size reduction, replacing the max-pooling layer to overcome the problem. To evaluate our proposed method, a dataset that contains images of number plate characters divided into 36 classes, which represent letters A - Z and numbers 0 - 9. Each class consists of 100 images as a data test and 400 images as a data train. Experiments showed that our proposed method improved accuracy from 91.51% to 93.87%, which is 2.36% better.