Transfer Learning for Few-Shot Plants Recognition: Antarctic Station Greenhouse Use-Case

Liliya Lemikhova, Sergey Nesteruk, Andrey Somov

Abstract: In this paper, we apply computer vision for plant recognition at the Antarctic station greenhouse, a training facility for future space colonization missions. Our experiments rely on transfer learning and explore the importance of the pre-training data domain. We show that a common approach of using models pre-trained on the Imagenet dataset can be further improved using publicly available domain-specific datasets. The classification results of 17 plant varieties with the ResNet50 model increase the F-score from 75% to 82 % using only 3 training images. We also achieve 78% top-3 accuracy without any training data.