An Active and Transfer Learning Method for Instance Segmentation using Mask-RCNN

Abstract

State-of-the-art instance segmentation is one of the hottest topics in image recognition. Image recognition makes use of convolutional neural networks. Training convolutional neural networks require a vast amount of data and computational excelling machines. This is often not feasible for simple tasks. This master thesis investigates multiple machine learning methods to assist the user to label data and to train convolutional neural networks, without a need for large datasets and the newest computer. This master thesis proposes a tool that can train neural networks within two hours and gives promising results for small datasets.