The evaluation of the quality of fruits and vegetables not only influences customer choices but also plays a crucial role for companies to monitor the production and distribution processes. In recent years, there has been a growing interest on the adoption of non-destructive techniques to automatically assess the product quality along the agroalimentary supply chain, also to be adopted in live environments on devices with low computational resources (e.g., fridges). In this paper, we present a solution that leverages the color distribution to mitigate the sensitivity of machine learning models to light/color variations. Specifically, we extract the color histogram and create multiple color aggregations to reduce the impact of perturbations on the model output. Our experiments, conducted on two real-world datasets and across two distinct learning tasks, demonstrated the effectiveness of the proposed method, also compared to state-of-the-art approaches based on complex neural network architectures.
Improving the Robustness to Color Perturbations of Classification and Regression Models in the Visual Evaluation of Fruits and Vegetables
Stefano Polimena;Gianvito Pio
;Michelangelo Ceci
2024-01-01
Abstract
The evaluation of the quality of fruits and vegetables not only influences customer choices but also plays a crucial role for companies to monitor the production and distribution processes. In recent years, there has been a growing interest on the adoption of non-destructive techniques to automatically assess the product quality along the agroalimentary supply chain, also to be adopted in live environments on devices with low computational resources (e.g., fridges). In this paper, we present a solution that leverages the color distribution to mitigate the sensitivity of machine learning models to light/color variations. Specifically, we extract the color histogram and create multiple color aggregations to reduce the impact of perturbations on the model output. Our experiments, conducted on two real-world datasets and across two distinct learning tasks, demonstrated the effectiveness of the proposed method, also compared to state-of-the-art approaches based on complex neural network architectures.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.