Globoki modeli za segmentacijo beločnice

HAFNER, ANDREJ (author), Peer, Peter (mentor), Vitek, Matej (comentor)

Semantična segmentacija je proces označevanja delov slike na nivoju slikovnih elementov. Na podlagi rezultatov lahko razberemo, kaj je pomensko vsebovano na posameznih predelih slike. To delo obravnava segmentacijo beločnice očesa s pomočjo trenutno najbolj uspešnih arhitektur nevronskih mrež. Javno dostopne implementacije arhitektur SegNet, DeepLabv3+, HRNetV2 in UPerNet predelamo in naučimo za segmentacijo beločnice na podatkovnih množicah SBVPI in MASD. Nato ocenimo njihovo uspešnost pri binarni klasifikaciji posameznih slikovnih elementov kot beločnica ali ozadje. Na podlagi metrik mIoU, natančnosti, priklica in f1-ocene se za najbolj uspešnega izkaže model UPerNet, kar v nadaljevanju pokažemo v kvalitativni analizi. Modele testiramo tudi na podmnožici testne množice iz tekmovanja v segmentaciji beločnice SSBC 2019. Rezultate primerjamo z zmagovalnim modelom U-Net, tukaj pa zmaga model SegNet.

• Language: Slovenian
• Keywords: biometrija, segmentacija beločnice, globoke nevronske mreže, konvolucijske nevronske mreže
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2020
• PID: 20.500.12556/RUL-114893
• COBISS.SI-ID: 1538558915
• Publication date in RUL: 25.03.2020

Secondary language

• Language: English
• Title: Sclera segmentation using deep neural networks

Abstract

Semantic segmentation is the process of labeling the images pixel-wise. The result is a mask, from which we can depict what is located in certain parts of the image. In this work we dive into semantic segmentation of sclera, using state-of-the-art neural network arhitectures. Publicly available implementations of SegNet, DeepLabv3+, HRNetV2 and UPerNet are adapted and trained on the SBVPI and MASD datasets. We measure their success at the binary classification of the pixels as sclera or background. For this we use performance metrics mIoU, precision, recall and f1-score. We find the model UPerNet most succesful at this task, which is also show in the qualitative analysis. Models are also tested on a subset of the test set of the sclera benchmarking competition SSBC 2019. The results are compared to the winning model, where Segnet takes the lead.

• Keywords: biometry, sclera segmentation, deep neural networks, convolutional neural networks