← Back

Weather Image Generation (Thesis)

用於生成不同天氣影像的端到端深度學習生成網路

Master's Thesis (Chinese) English Version
Master's Thesis Weakly Supervised Learning Triplet Content Loss
Weather Generation Visualization Network Architecture

Project Overview

This project, conducted as my Master's Thesis, introduces DGNDW, an end-to-end deep learning framework designed to synthesize diverse weather conditions. It addresses the scarcity of adverse weather datasets (fog, rain, snow) for training robust autonomous driving systems.

Core Techniques

The framework leverages Weakly Supervised Learning to translate clear images into weather-degraded counterparts. A key contribution is the novel Triplet Content Loss, which preserves the semantic structure of the scene while realistically rendering weather effects.

The generated data significantly improves the domain adaptation capabilities of downstream object detectors, bridging the gap between clear and adverse weather performance.

Related Publications

[1] Weather Image Generation (DGNDW)
Master's Thesis, CYCU, 2023
Authors: Shyang-En Weng, Shaou-Gang Miaou
Thesis Link