Small datasets for fruit detection with transfer learning.

Abstract

A common approach to the problem of fruit detection in images is to design a deep learning network and train a model to locate objects, using bounding boxes to identify regions containing fruit. However, this requires sufficient data and presents challenges for small datasets. Transfer learning, which acquires knowledge from a source domain and brings that to a new target domain, can produce improved performance in the target domain. The work discussed in this paper shows the application of transfer learning for fruit detection with small datasets and presents analysis between the number of training images in source and target domains. This investigation is based on three datasets: two containing tomatoes and one containing strawberries. Experimental results indicate that transfer learning can enhance prediction with limited data.