Deep learning in Dynamic Imager - A convolutional neural network module

Abstract

This thesis investigates the extent of which deep learning methods can be used for automatic detection of salmon in images. It also investigates the extent of which a module with deep learning functionality can be integrated into an already existing program.

To solve these tasks, a comparison of state-of-the-art object detectors using convolutional neural networks is conducted. A dataset consisting of underwater images of salmon is annotated, then used to train a model based on the YOLOv2 (You Only Look Once version 2) object detection method. A module using the Darknet software framework is created for use with the image processing program Dynamic Imager. Finally, the trained models are evaluated based on their accuracy and speed.

The best performing model achieve an average precision score of 80.2\%. Models tested using a graphical processing unit achieves a prediction time of 0.8 frames per second. Models tested in Dynamic Imager without access to a graphical processing unit achieve a prediction time of 0.1 frames per second.