Deep machine learning has revolutionised computer vision in recent years, achieving performance levels across a variety of tasks that surpass previous approaches and often exceed human performance. Convolutional Neural Nets (CNNs) of various forms are of particular interest, and many applications of existing architectures in the life and biological sciences have been reported.
Images of plants, however, raise particular challenges that differ from those faced in the wider computer vision community. Objects of interest are often very small but embedded in very large images, with the contextual information needed to correctly interpret those images distributed across them. Deployment of CNNs can also be problematic for a community focussed on analysing plants in their natural habitat – fields.
This talk will describe novel convolutional neural net architectures developed for shoot and root image analysis at the Computer Vision Laboratory, University of Nottingham, and the infrastructures and methods used to support their development and deployment.
Authors: Tony Pridmore, Andrew French, Michael Pound