Ordinal-Meta Learning for Fine-Grained Fruit Quality Prediction

Abstract

To mitigate the wastage of perishable food items such as fruits, it is crucial to predict their shelf-life, preferably by using non-invasive image-based techniques. Due to the cost associated with the expert fine-grained fruit quality annotations, there is a need to develop deep learning techniques for fruit quality prediction yielding acceptable performance with the availability of scarce annotated data. To address this need, in this paper, we enhance the performance of existing approaches which meta-learn the visual cues of fruit quality degradation common across multiple fruits, by leveraging the ordinal nature of the fruit quality labels. We consider a realistic industrial supply chain setting with fixed camera placement in warehouses and/or retail stores, and thus exposed to consistent background and lighting conditions while capturing fruit images. We formulate a dataset to simulate this scenario by sampling train and test data from the same set of time-lapse videos ensuring same data distribution across splits. We also augment the dataset by web-curating images, which is a representative of data that an end-user would capture from a hand-held device such as a smartphone. The results indicate that our ordinal regression-based meta-learning approach provides a high accuracy in few-shot learning regime (85.12%) as compared to only meta-learning (77.23%) and few-shot transfer learning (21.49%) approaches. The average accuracy further increases to 93.23% on time-lapse images of climacteric fruits, making it acceptable for usage in realistic fruit supply chain settings.