Transfer learning from indoor to outdoor in cultivar testing: application to russeting detection

El Abidine M. Z., DUTAĞACI H., Rousseau D.

Acta Horticulturae, vol.1360, pp.245-251, 2023 (Scopus) identifier

  • Publication Type: Article / Abstract
  • Volume: 1360
  • Publication Date: 2023
  • Doi Number: 10.17660/actahortic.2023.1360.31
  • Journal Name: Acta Horticulturae
  • Journal Indexes: Scopus, CAB Abstracts, Food Science & Technology Abstracts, Veterinary Science Database
  • Page Numbers: pp.245-251
  • Keywords: horticulture, russeting, supervised classification, transfer learning
  • Eskisehir Osmangazi University Affiliated: Yes


Postharvest measurements in digital horticulture are nowadays well-calibrated. They are done indoors, i.e., in a controlled environment, with chosen lighting systems and optimized positioning of the sensors to probe the harvested items. In outdoor conditions, the deployment of digital horticulture with sensors and cameras remains challenging due to uneven illumination of the scenes or the spatial 3D complexity of the plants. In this contribution, we consider the important problem of russeting detection on apples. This is an important trait that impacts the quality of the fruit for the consumer and is commonly measured in cultivar testing trials. We investigate the possibility of using indoor postharvest russeting data (RGB images of individual apples) to help the detection of russeting on apple images in orchards. This is obtained with a shallow learning approach. First, a supervised model quantifies the amount of russeting in indoor data based on the low-cost apple sorting machine described in Couasnet et al. (2021). The indoor data set is then merged with the outdoor data set in the training phase to enhance the presence of russeting texture information through deep features. The outdoor data set includes images with several apples encompassed in a single image. The apples are first detected with a standard object detection algorithm (YOLOv4 Tiny). Inference on these detected objects is then boosted with the use of the indoor data set.