MVCNN-CASHNET: Multi-View Convolution Neural Network for classifying WW, SW, Split Cashews

Abstract

Recently multi-view image CNN has grabbed more attention in image classification applications. In particular the computer vision has gained a lot of importance for its abundant prospective applications in food quality control. Cashew nut is an important crop in India among all the available dry fruits. Specifically, high quality cashew nuts are having high popularity in the international market. Though there are many existing methods to automatically classify cashew nuts, most of them are concentrating only on a single view of the cashew nut. The major difficulty in classifying whole cashew and split down cashew using existing methods is, a single view image of cashew nut can’t cover the entire cashew nut and thereby resulting in lack of classification accuracy. This paper proposes a novel framework for classifying three grades of cashew nut by using Multi-view CNN. Here the images of the sample cashew nuts are captured from three different angles (Top-view, Left-view, and Right-view) and given as input to our modified CNN architecture. The modified CNN extracts various features from these three images and combines for further classification. The cashew samples are collected from the local cashew industries and the images are captured from these samples for training and testing process. The proposed work provides an overall accuracy of 98.87% and 4.49% error rate.

Keywords- Multi-view, Convolutional neural network, image classification, feature extraction.