Towards automation in the fish processing industry using machine learning

Abstract

This master project was inspired by challenges faced by commercial fisheries in the north of Norway of controlling food quality and food safety. In this thesis, four different ML models’ ability to do object and keypoint detection on specific anatomy parts of fish, has been studied. With the aim of recommending a suitable model to be part of a CV system for an industrial fish gutting machine that cuts open the fish belly between the pelvic fins and the anus. Requirement that the rotating knife shall not cut into the flesh behind the anus opening, and cut should end (or start) maximum 5 millimeters from the anus opening. Likewise, at the pelvic fins, the cut shall start (or end) 15 millimeters from target along the centerline of the fish, and a sideways offset of roughly ±5 millimeters can be acceptable, depending on the length of the fish.

The experiments were performed with two YOLOv7 and two Detectron2 models, YOLOv7 for object detection with bounding boxes, and Detectron2 for keypoint detections. The results showed that only one of the Detectron2 models was able to do keypoint detection repeatedly, but the achieved accuracy was not good enough. Both the YOLOv7 models were able to meet the cut length requirements and both got recommended for use in the suggested CV solution.

More work still remains before one of the YOLOv7 models can be taken in use, such as determining the object detection speed, finding a suitable embedded computer with GPU to run the CV system on, determining the best way of communication between the PLC in Folla and the CV system and finding a suitable location for a camera inside the Folla machine.