Data Augmentation for Object Detection using Deep Reinforcement Learning

Sammanfattning: Data augmentation is a concept which is used to improve machine learning models for computer vision tasks. It is usually done by firstly, defining a set of functions which transforms images and secondly, applying a random selection of these functions on the images. Since the quality of training data is one of the, if not the most important factor to obtain a good model, this master thesis poses the question whether an intelligent deep reinforcement learning (DRL) agent can select augmentation functions in a better way. More specifically, can the agent select augmentations such that the performance of an object detection model increases? Besides improving the performance of an object detection model, the DRL agent provides insights in what constitutes good data augmentation. The project results in an agent which augments images such that mean average precision (mAP50) increases with 2.3% compared to a baseline detector, trained with random augmentations. This is a promising result that encourages further research on this area. To our knowledge, this is the first time a deep reinforcement learning agent has been used to improve an object detection model via better data augmentation.