Investigating the Influence of Colour Spaces on Convolutional Neural Networks in Open-ended 3D Object Recognition

Mein, Anne-Jan (2021) Investigating the Influence of Colour Spaces on Convolutional Neural Networks in Open-ended 3D Object Recognition. Bachelor's Thesis, Artificial Intelligence.

Preview	Text AI_BA_2021_ANNE-JANMEIN.pdf Download (1MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (126kB)

Abstract

Due to the rising use of service robots there has been an increased interest in high performing 3D object recognition architectures for open-ended environments. These architectures have successfully been created with the use of Convolutional Neural Networks utilising an open-ended learning approach. Adding colour-information to the object representation has demonstrated to increase performance. This study aims to examine the influence of colour-spaces on neural networks with regards to open-ended object recognition. This has been done by converting the colour-information to the following colour-spaces: RGB, LAB, HSV, XYZ and YUV, as well as grayscaled images. These representations were then given as input for networks: MobileNetV2, vgg16 and ResNet50. Three rounds of experiments were performed with the first two rounds utilising an offline- and last round an online evaluation. The first round to determine the best hyperparameter configuration for each network The second round to compare the colour-spaces, resulting in each network using a different colour-space to reach their highest performance. The online evaluation showed that vgg16 combined with the YUV colour-space achieved the highest object recognition performance in an open-ended setup. These results indicate that choosing the choosing the correct colour-space for an object recognition architecture utilising a Convolutional NeuralNetwork can lead to a perrmance increase with regards to open-ended object recognition.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Mohades Kasaei, S.H.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	07 Jul 2021 10:34
Last Modified:	07 Jul 2021 10:34
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/25023

Actions (login required)

View Item