Tecniche di Deep Learning applicate all'allevamento

Abstract: Although Deep Learning (DL) is increasingly being adopted in many sectors, farming is still an almost unscathed niche. This is mainly because of the humongous distance in knowledge between the experts of DL and those of farming itself. It's first of all a communication issue, and only in a second place a matter of reluctance to changes. Tackle those issues and you will find that also this sector can greatly benefit from the application of these new technologies. This thesis if therefore a collection of applications of DL to different topics in farming. This has been made possible by the key role of figures who are placed right in the middle and act as intermediaries between the experts to identify targets and measures of success. In our case, this role is covered by the Farm4Trade startup, which is also the main funder of this PhD. The first covered topic is the automatic cattle re-identification from images and videos. We show how DL methods designed for humans can be adapted to work in a completely different setting. As a feedback loop, methods developed for cattle have been reapplied on people and vehicles with successful results. The second topic is the automatic detection and tracking of pigs in the farm. The target here is to detect and classify individual behaviours and how they change through time. A collection of state-of-the-art DL techniques has been chain together while each individual piece has been analysed on its own to ensure good final performance. Finally, we jump at the end of the production chain to study how to apply DL to slaughtered pigs' carcasses image to detect and segment lungs lesions. These are reliable indicators of a bacterial pathology affecting the animal prior to its death. Results achieved during this PhD show how the whole sector of farming can benefit from the application of artificial intelligence algorithms.

Citation:

Bergamini, Luca "Tecniche di Deep Learning applicate all'allevamento" 2021

 not available

Paper download:

• Author version: