The world's first sheep facial recognition software, developed in Dunedin, is set to be prototyped this year.
Sheep NN, a project created by artificial intelligence and machine learning company Iris Data Science, has received a $40,000 grant from Callaghan Innovation towards the $100,000 project that will take the model toprototype by the end of the year.
Iris Data Science was co-founded by Greg Peyroux and Benoit Auvray, who have been working on the project to cheaply re-identify sheep, potentially removing the need for ear-tags while also solving other farm management and broader issues.
The team recently began collecting data and developing similar technology for other livestock, including cattle and goats, and expected that would be available shortly after the sheep prototype.
Once prototyped, then it could be put in front of industry and potential investors and it was hoped to be able to do that early next year, Peyroux said yesterday.
The project created interest when it was displayed at MobileTECH 2019 in Rotorua in April - an annual event showing digital technologies for the agricultural, horticultural and forestry sectors - and at TEXpo during Dunedin Techweek.
Sheep face images were collected from around Southland, Canterbury and Otago, and fed into a machine-learning model. It slowly learned by itself to identify sheep by finding recognisable features.
Since taking the first pictures more than a year ago, the company had collected "thousands'' of images and hours of high-resolution video footage from farms to create a deep learning identification pipeline that would be further developed in the coming months.
Benoit Auvray and Greg Peyroux, of Iris Data Science, who have been working on sheep facial recognition. Photo / Stephen Jaquiery
As more farmers moved towards management technologies such as digital scales and automatic drafting gates, a reliable low-cost method of identification was essential, Peyroux said.
Sheep were originally chosen as the company wanted to be first in the world to develop the technology for sheep recognition; now people had told them that there were other gaps, such as goats.
Interest had been show by the likes of software and hardware manufacturers, mostly overseas, particularly Australia.
The challenge had been to get the technology working on farm as soon as possible, so people could see it running, and the funding would help them do that, he said.
Auvray said the funding would help with overcoming the challenges of applying a deep neural network to detect sheep faces in an image complicated by changing head pose, background and lighting conditions.
Future applications for the technology were broad and included tracking animal locations to prevent stock rustling, monitoring animal behaviour, estimating weight, diseases, welfare, or other characteristics, or estimating parentage without the need to observe lambing or do DNA parentage testing.
Other projects included a pasture quality system and optimising fertiliser application.
Environmental impact conversations were becoming "louder and louder" all the time, Peyroux said.
