Deep Learning based Computer Vision

Published on
June 14, 2017

In the past few years we have been devoted to create new methods for computer vision in horti- and agriculture. By using convolutional neural networks (Deep Learning), we can analyse the class and quality of plants, fruit and vegetables in the greenhouse or on the field on a per-pixel level.

Specifically, in the past few months we accomplished the feature of recognising individual plant parts. Using this method, we can localise with high certainty where the stem, fruit and leafs are positioned. This enables us to create a local geometry of the plant. Until now this level of detail could not be realised robustly, but Deep Learning now enables us to perform this analysis on a large scale and speed.

For applications one primarily must think in the direction of harvest robotics. This next generation mechanisation requires exact localisation of plant parts in order to position the end-effector (the grabbing and cutting tool) successfully at the fruit.

But also for phenotyping, extensively used in plant breeding to obtain plant features, this new technique will become a valuable tool. Currently, properties of the plants are measured by hand. However, with Deep Learning these features will be able to obtained automatically in the near future.

A third application is the detection of diseases. Currently already measuring systems navigate the greenhouse and the field. With advanced computer vision on hyperspctral images we can now determine on a per-pixel level if diseases or plagues are present in the crop.

The research of this new technique does require some computing power. Therefore we acquired GPU accelerated computers. Also a lot of data is needed to train the neural networks. Because this is manually annotated, we furthermore are researched methods to partially automate this part.

This project received funding in the Horizon 2020 program for research and innovation from the European Commission under contract number 644313.