Non-destructive and contactless quality evaluation of table grapes by a computer vision system

Abstract

Quality rating is currently accomplished by non-destructive and subjective sensory evaluation or by objective and destructive analytical techniques. There is a strong need of an objective non-destructive contactless quality evaluation system to monitor fruit and vegetable along the whole supply chain. This paper proposes a Computer vision system to satisfy this request. Image processing and machine learning techniques have been combined to develop a Computer vision system whose configuration and tuning has been strongly simplified: that makes easier its deployment in real applications. The system has been verified on two white table grape cultivars (Italia and Victoria) against three different classification tasks. The first considered five quality levels (5, 4, 3, 2, 1); the second separated the higher fully marketable quality levels (5 and 4) from the boundary (3) and the waste (2 and 1); the third separated the higher fully marketable quality levels (5 and 4) from the other three (3, 2 and 1). The system achieved a cross-validation classification accuracy up to 92% on the cultivar Victoria and up to 100% on the cultivar Italia for binary or binomial classification between fully marketable and residual quality levels. The obtained results support its capability of powerfully, flexibly and continuously monitoring the quality of the complete production along the whole supply chain.