Segmentation plays an important role in imagebased plant phenotyping applications. Deep learning has led to a dramatic improvement in segmentation performance. Most deep learning-based methods are supervised and require abundant application-specific training data. Considering the wide range of plant phenotyping applications, such data may not be always available. To mitigate this problem, we introduce a segmentation method that exploits the power of deep learning without using any prior training. In this paper, we specifically focus on flower segmentation. Recurrence of information inside a flower image is used to train an image-specific deep network that is subsequently used for segmentation. The proposed method is self-supervised as it exploits the internal statistics of input image without using any prior labeled data. To the best of our knowledge, this is the first unsupervised deep learning-based method proposed for single-image flower segmentation.