Zero-shot transfer learned generic AI models for prediction of optimally ripe climacteric fruits.

Ideally, ripe fruits offer appropriate nutritional content and best quality in terms of taste and flavour. Prediction of ripe climacteric fruits acts as the main marketing indicator for quality from the consumer perspective and thus renders it a genuine industrial concern for all the stakeholders of the fruit supply chain. However, the building of fruit-specific individual model for the prediction of ripeness level remains an existing challenge due to the scarcity of sufficient labeled experimental data for each fruit. This paper describes the development of generic AI models based on the similarity in physico-chemical degradation phenomena of climacteric fruits for prediction of 'unripe' and 'ripe' levels using 'zero-shot' transfer learning techniques. Experiments were performed on a variety of climacteric and non-climacteric fruits, and it was observed that transfer learning works better for fruits within a cluster (climacteric fruits) as compared to across clusters (climacteric to non-climacteric fruits). The main contributions of this work are two-fold (i) Using domain knowledge of food chemistry to label the data in terms of age of the fruit, (ii) We hypothesize and prove that the zero-shot transfer learning works better within a set of fruits, sharing similar degradation chemistry depicted by their visual properties like black spot formations, wrinkles, discoloration, etc. The best models trained on banana, papaya and mango dataset resulted in s zero-shot transfer learned accuracies in the range of 70 to 82 for unknown climacteric fruits. To the best of our knowledge, this is the first study to demonstrate the same.