Descrizione: “Fresh fruits and vegetables are very perishable, with the main factors implying loss of consumer acceptability being discoloration or browning, dryness, and texture loss. These parameters determine the visual appearance, which influences consumer perception and therefore the level of purchase acceptability. This is because consumers associate desirable internal characteristics (i.e., nutraceutical and organoleptic properties) with the external appearance. In this context, quality control processes are usually applied, that include: a) the evaluation of their global visual quality; b) the estimation of some internal characteristics, that determine nutraceutical and organoleptic properties; c) the verification of the sustainability of their cultivation strategies in terms of the use of critical resources (water, fertilizers, etc.). These steps are usually performed manually. However, manual inspection requires people to carry out the assessment of the entity and make a judgment on it according to some prior training or knowledge, which may introduce inconsistencies and subjectivity in the quality evaluation. In this context, this research project concerns the design and application of machine learning and data mining techniques for non-destructive contactless quality control in the agroalimentary supply chain. The use of these techniques would make it possible to reduce labour costs, improve process efficiency and quality control reliability.
In the literature, we can find methods that rely on and assume optimal conditions that are hardly achievable in real contexts. These include heavy manual feature engineering phases; the use of images with a black background and the adoption of colour charts for calibration purposes, or lamps to reduce highlights. Moreover, they usually solve a classification task to estimate the quality level, without taking into account the ordering of the classes and, therefore, different degrees of severity in making a mistake in the estimation.
Therefore, the objectives of this research project can be summarized as follows:
- Design and development of specific regression or cost-based classification methods to evaluating the fruit’s and vegetable’s visual quality and relevant internal characteristics
- Use images that are independent from the background and from calibration patches;
- Automatic identification of the most relevant features
- Explainability: understand and emphasize which parts of the images have the greatest impact on the regression/classification output.”
Università: Università degli Studi Aldo Moro di Bari
Dottorando: Stefano Polimena
Responsabile STIIMA: Attolico Giovanni
Durata: 2021-2024