Performance

In a normal software, a calculation is performed on the basis of data that is entered into the software. In machine learning, software and algorithms evolve as they are “trained” with new datasets. This training becomes a form of pattern recognition, where the algorithms learn what to look for based on prior examples. As new training data and examples are provided to the machine learning software, the algorithms generate more and more precise results.

Let’s say that the algorithms are trained to recognize salmon lice. The algorithms are given training data, consisting of many images of salmon with annotations of lice in different stages. Through feedback in the form of these examples, the algorithms become increasingly accurate in recognizing salmon lice. It is important that the photos are taken under real conditions in the cage: changing lighting conditions, different angles of the fish, for example.

The quality of the algorithms is important. Algorithms may need to be adjusted along the way to emphasize biological factors in the data that make the results more accurate. For example, consider an algorithm that recognizes individual fish based on unique spot patterns. If algorithms are to recognize the same fish over time, these algorithms must also have the capacity to understand how these spot patterns change with the growth of the fish.

Images are taken with an underwater camera in the cage. Per Erik Hansen, Product Manager for Lice and Welfare at Aquabyte, mentions that the company uses a standard two-lens camera with particularly good optics to be able to measure 3D distance to the fish. The camera continuously takes images, and is placed in an optimal location such that as many fish as possible can swim past. Software running on the camera filters out unused images and analyzes the best ones.

“Customers who have used the system for a long time report that the results turn out to be in line with reality,” Per Erik Hansen continues. Farmers get daily lice numbers for each cage, with lice counts on far more fish than possible with manual counting. “The software distinguishes individual fish from each other, such that the same fish is only counted once. This contributes to Aquabyte delivering lice numbers with great accuracy,” notes Hansen.

Precise counting of salmon lice and accurate estimation of weight data and distribution are two use cases that have received fully developed solutions. Fish welfare is another important application area, including detection of winter sores, deformations, and other external changes on the fish. By giving the farmer much better insight and basis for decision making, it will be possible to improve daily operations, and in general achieve more efficient and sustainable fish farming.