A Generic Few-Shot Solution for Food Shelf-Life Prediction using Meta-Learning

Abstract

Checking the quality of agricultural produce at every step of its supply chain is the need of the hour to reduce food wastage. Manual checking of food quality at every step can be inconsistent and time consuming. Automation of food quality detection, using non-invasive imagery based techniques, needs availability of ample amount of annotated data to train models. Collecting such data in large quantity in a controlled lab setting is an expensive affair. More-over, providing a point solution for every individual food item by training food item specific models is an impractical solution. Thus, there is a need for a mechanism which would capture the common meta-level visual degradation properties across a set of food items belonging to a specific category and use this meta-knowledge to predict the quality of a new food item belonging to that category with a paucity of training data. To address this challenge, as a part of the preliminary work, we conduct an initial set of experiments to demonstrate the applicability of existing Model Agnostic Meta-Learning (MAML) algorithm for fruit freshness detection task. The results indicate that for such a task, meta-learning can serve to be a more generic and efficient solution than using few-shot transfer-learning technique and traditional ML based approaches requiring explicit feature engineering.