Determination of Fruit Freshness Using Near-Infrared Spectroscopy and Machine Learning Techniques

Abstract

AbstractThis paper addresses the problem of fruit freshness categorization in the context of fruit quality assessment during short storage periods. As it is hard to handle by using computer vision technology, we propose a novel method by using absorbance near-infrared spectroscopy combined with machine learning (ML) techniques. We collected several samples of five popular fruits with various properties and classified them into three degrees of freshness based on the storage duration. We then examined multiple combinations of feature extraction and machine learning techniques. Experimental results show that the proposed Convolutional Neural Network (CNN) architecture were superior to other traditional ML models regardless of the selected feature vector. In particular, the proposed CNN when trained on the concatenated first and second derivatives of the pre-processed absorbance spectrum achieved the highest accuracy of 80.0%. The obtained classification performance was evaluated on a variety of fruits, which shows the potential of our approach.KeywordsNear-infrared spectroscopyAbsorbance spectrumFeature extractionMachine learningFruit freshness classification