Evaluating freshness of produce using transfer learning

Automated quality control of produce such as fruits and vegetables is of great importance to industry. In particular, the ability to evaluate the state of decay for various produce items would allow for efficient sorting of produce such that the freshest items could be more quickly shipped to consumers. Unfortunately, training an accurate classifier for determining how decayed produce is can require a large amount of data. This problem is further exacerbated by the large variety of produce available as different items would exhibit decay in different ways. In this paper, we propose an algorithm that can learn an accurate ranking classifier for sorting produce using only a small amount of data. We achieve this through our proposed transfer learning algorithm that is able to automatically select good preexisting source task training data to supplement insufficient training data in the given target task. We show how much our algorithm improves over standard training on real images of produce items captured at various stages of decay.