Chili fruits maturity estimation using various convolutional neural network architecture

Abstract

Agricultural robots recently become popular by helping the farmer to conduct their daily chores. A slow process of picking and grading will leads to an inaccurate result thus increasing the production cost. This study represents an innovative and economical alternative for farmers who require to undergone the process of estimating their maturity categories. A total of 1,200 chili images with 256×256 pixel are used, where 840 is used for training and the remaining 360 being served for testing. The maturity is determined by measuring the length of chili structure between the calyx and apex. Various convolutional neural network (CNN) architectures are applied to learn and recognize the chili fruits into three maturity categories; immature, moderately mature, and mature. ADAM and stochastic gradient descent with momentum (SGDM) optimizers with multiple CNN architectures is capable in recognising and classifying chilli fruits with an accuracy of above 85%.