A comprehensive benchmark for single image deraining networks
Computer vision systems can be greatly affected by adverse weather conditions, such as rain and haze. The success achieved by popular models in common high-level vision tasks typically relies on clean weather images. However, in real world, such clean condition is not always available. In this conte...
| Autor: | |
|---|---|
| Formato: | tesis de maestría |
| Estado: | Versión publicada |
| Fecha de publicación: | 2019 |
| País: | Brasil |
| Recursos: | Universidade de São Paulo (USP) |
| Repositorio: | Biblioteca Digital de Teses e Dissertações da USP |
| Idioma: | inglés |
| OAI Identifier: | oai:teses.usp.br:tde-20082025-192226 |
| Acesso em linha: | https://www.teses.usp.br/teses/disponiveis/45/45134/tde-20082025-192226/ |
| Access Level: | acceso abierto |
| Palavra-chave: | Convolutional neural networks Deep learning Deraining Machine learning Redes neurais convolucionais |
| Resumo: | Computer vision systems can be greatly affected by adverse weather conditions, such as rain and haze. The success achieved by popular models in common high-level vision tasks typically relies on clean weather images. However, in real world, such clean condition is not always available. In this context, many single image deraining algorithms have been proposed in order to remove image degradation caused by the presence of rain in the scene. This work presents a comprehensive study and evaluation of recent single-image deraining algorithms and their current limitations as well as conclusions drawn from a thorough investigation. We provide a robust and comprehensive analysis to guide a model proposal capable of overcoming the limitations of current state-of-the-art deraining algorithms. We collected a large-scale dataset including synthetic rainy images and real world rainy images separated by the rain type formation. Besides, we annotated real world rainy images to evaluate the raining and deraining impact on the detection task. This task-driven approach is a novelty on this work and it provides future research directions. |
|---|