Semantic Segmentation in Low Light Disaster Scenes
Author
Cui, Bo
Term
4. semester
Education
Publication year
2022
Abstract
This thesis addresses how mobile robots can perform reliable semantic segmentation in low-light disaster scenes, where methods trained on bright images often degrade. To tackle both image quality and the lack of suitable data, the work proposes a two-step approach: (1) synthesizing a low-light disaster dataset (LLDSD) from PST900 using gamma darkening and noise modeling, and (2) a pipeline that first enhances low-light images with a DSLR image enhancement network and then applies a semantic segmentation model. The segmentation component is a modified PSPNet with a loss function tailored to the label characteristics in the used data, combined with pre-training, layer freezing, and data augmentation. Experiments show that the modified PSPNet outperforms the original PSPNet on the datasets used, and that the ‘enhancement followed by segmentation’ sequence increases accuracy on low-light images. Contributions include a practical dataset synthesis procedure, an integrated enhancement–segmentation pipeline, and empirical evidence of improved performance in dark disaster environments.
Dette speciale undersøger, hvordan mobilrobotter kan foretage pålidelig semantisk segmentering i katastrofescener med lav belysning, hvor standardmetoder trænet på lyse billeder ofte fejler. For at adressere både billedkvalitet og manglen på passende data foreslår arbejdet en to-trins tilgang: (1) syntese af et lav-lys katastrofedatasæt (LLDSD) ud fra PST900 ved hjælp af gamma-formørkning og støjmodellering, og (2) en pipeline, der først forbedrer lav-lys billeder med et DSLR-billedforbedringsnetværk og derefter anvender en semantisk segmenteringsmodel. Segmenteringsdelen bygger på en modificeret PSPNet, hvor tabsfunktionen tilpasses etikettegenskaberne i de anvendte data, kombineret med prætræning, frysning af lag og dataaugmentation. Eksperimenter viser, at den modificerede PSPNet overgår den oprindelige PSPNet på de anvendte datasæt, og at sekvensen ‘forbedring efterfulgt af segmentering’ øger nøjagtigheden på lav-lys billeder. Bidragene omfatter en praktisk datasætsynteseprocedure, en integreret forbedrings-segmenteringspipeline og empirisk evidens for forbedret ydeevne i mørke katastrofemiljøer.
[This apstract has been generated with the help of AI directly from the project full text]