Transfer Learning with Pairwise Comparison for Fine-Granular Fruit Shelf-Life Prediction

Jayita Dutta, Manasi Patwardhan, Parijat Deshpande, Beena Rai

Keywords: Fine granular classification, Shelf-Life Prediction, pairwise comparison, Fruits 360, Vgg16, reduction of food waste

TL;DR: Fine-Granular Classification fruit Shelf-Life Prediction

Abstract: Globally, 30% of the produced fruits gets wasted annually across the supply chain. To control this huge wastage and corresponding economic losses, real time prediction of shelf life of fruits is essential. This process of reporting shelf life of fruits is often carried out manually using invasive techniques by domain experts, which becomes infeasible when the fruits are getting transported over long distances across the supply chain. To automate this process, we use non-invasive vision-based technique to predict the current age of the fruit from which the shelf-life can be computed. To achieve this we train a model to capture the visual degradation features of fruits. However, such models require large amounts of annotated images for training with fine-granular `days-old' labels. Curating such dataset either by expert annotations or laboratory experiments is expensive. Also, the annotated datasets available for this task are scarce. To address this challenge, in this paper, we avail the accessibility of online time-lapse videos of fruits to auto-synthesise a dataset for the task of pairwise comparison of video frames, whose labels are generated based on their sequence in the video. We transfer-learn the knowledge gained by a model, trained with this data on the comparison task, to improve the performance of the fine-granular classification task where labels are depicting the age of a fruit image. We empirically showcase that with this approach, the performance for the classification task improves by a margin of 16% - 23% percentage in terms of fine-granular classification accuracy for distinct fruits.