Automatic Framework for Vegetable Classification using Transfer-Learning

Abstract

Globally, fresh vegetables are a crucial part of our lives and they provide most of the vitamins, minerals, and proteins, in short, every nutrition that a growing body need. They vary in colors like; red, green, and yellow but as our ancestors say that green vegetables are a must for every age. To identify the fresh vegetable that makes our body healthy and notion positi ve the proposed automatic multi-class vegetable classifier is used. In this paper, a framework based on a deep learning approach has been proposed for multi-class vegetable classification from scratch. The accuracy of the proposed model is further increased using the transfer-learning concept (DenseNet201). The whole process is divided into four modules; data collection and pre-processing, data splitting, CNN model training, and testing, and performance improvement using a pre-trained DenseNet201 network. Data augmentation and data shuffling are used to free from lack of data availability during the training phase of the model. The proposed framework is more efficient and can predict the type of vegetables comparatively in less computational time (2 to 3 minutes) with an ‘Accuracy’ of 98.58%, ‘Sensitivity’ of 98.23%, and ‘Specificity’ of 94.25%.