Machine Learning and Deep Learning Networks for The Classification of Rice Grain Images from Visual Testing

Abstract

Quality assessment of rice is mandatory during the export of rice. It necessitates a Non Destructive Testing (NDT) Technique for acquiring rice grain images, feature extraction and classification techniques.  In this work, visual images of rice grain images are collected from internet. Features are extracted using Discrete Wavelet Transform (Reverse Biorthogonal wavelet, Biorthogonal wavelet) and texture features. Generalized Recurrent Neural Network (GRNN) is used for the classification of rice grains from the above features. In order to further improve the performance of the classifier, squeezenet is used. The average sensitivity is 93.75%, the average specificity is 88.75% and the average accuracy is 90.625%. The performance of this SqueezeNet in deep learning is very high compared to the machine learning techniques.