Pure-CNN: A Framework for Fruit Images Classification

Abstract

Automation in fields like robot harvesting, farming, health and education require object classification using machine learning and computer vision techniques. Among these fruit classification is a challenging task because of its several varieties and similarity in color, shape, size and texture features. In order to recognize multiple fruits more accurately, we proposed a Pure Convolutional Neural Network (PCNN) with minimum number of parameters. The PCNN consists of 7 convolutional layers, some of them followed with stride. Additionally, to reduce overfitting and taking average of whole feature maps we employed recently developed Global Average Pooling (GAP) layer that verified to be very effective. We demonstrate our classification performance using PCNN on recently introduced fruit-360 dataset. The experimental results of the 55244 color fruit images from the 81 categories, show that the PCNN achieve a classification accuracy of 98.88%.