University of Oulu

Using encoder-decoder architecture for material segmentation based on beam profile analysis

Saved in:
Author: Bennadji, Bilel1
Organizations: 1University of Oulu, Faculty of Information Technology and Electrical Engineering, Computer Science
Format: ebook
Version: published version
Access: open
Online Access: PDF Full Text (PDF, 12.1 MB)
Pages: 63
Persistent link: http://urn.fi/URN:NBN:fi:oulu-202008212844
Language: English
Published: Oulu : B. Bennadji, 2020
Publish Date: 2020-08-21
Thesis type: Master's thesis
Tutor: Heikkilä, Janne
Reviewer: Heikkilä, Janne
Pedone, Matteo
Description:

Abstract

Recognition and segmentation of materials has proven to be a challenging problem because of the wide divergence in appearance within and between categories. Many recent material segmentation approaches treat materials as yet another set of labels like objects. However, materials are basically different from objects as they have no basic shape or defined spatial extent. Our approach roughly ignores this and can primarily take advantage of limited implicit context (local appearance) as it seems during training, because our training images that almost do not have a global image context; such as (I) where the used materials have no inherent shape or defined spatial extent like apple, orange and potato approximately have the same spherical shape; (II) besides, images where taken under a black background, which roughly removes the spatial features of the materials.

We introduce a new materials segmentation dataset, which was taken with a Beam Profile Analysis sensing device. The dataset contains 10 material categories, and it has image pair samples consisting of grayscale images with and without the laser spots (grayscale and laser images) in addition to annotated segmented images.

To the best of our knowledge, this is the first material segmentation dataset for Beam Profile Analysis images. As a second step, we proposed a deep learning approach to perform material segmentation on our dataset; our proposed CNNs is an encoder-decoder model, which is based on the DeeplabV3+ model. Our main goal is to obtain segmented material maps and discover how the laser spots contribute to the segmentation results; therefore, we perform a comparative analysis across different types of architectures to observe how the laser spots contribute to the whole segmentation. We built our experiments on three main types of models that use a different type of input; for each model, we implemented various types of backbone architectures. Our experiments results show that the laser spots have an efficient contribution on the segmentation results. GrayLaser model achieves a significant accuracy improvement compared to other models, where the fine-tuned architecture of this model has reached an accuracy of 94% over MIoU metric, and one trained from the scratch has reached an accuracy of 62% over MIoU.

see all

Subjects:
Copyright information: © Bilel Bennadji, 2020. This publication is copyrighted. You may download, display and print it for your own personal use. Commercial use is prohibited.