劣化場景下的視覺研究

Abstract

在數位時代中，對於自動駕駛、醫療診斷等多種應用領域，高品質視覺數據的需求顯得尤為重要。然而，由於惡劣天氣、低光照、模糊及噪聲等因素，真實環境中捕獲的影像常常受損，這不僅妨礙了人類的判斷能力，也減弱了機器學習模型在進行物體檢測、語義分割等下游任務的效能。面對這些挑戰，從精準的影像質量評估開始，到運用先進的影像恢復技術變得至關重要。隨著深度學習技術的進步，我們現在有能力利用深度學習模型從數據中學習複雜的模式和特徵，自動檢測和修正影像質量問題，達到顯著的進步。

本研究通過深度學習技術，針對受損影像系統性地開發了從影像質量評估到二維影像及三維體積重建的相關應用。此過程涵蓋了從初步的影像質量評估，到二維影像的質量提升，進而到三維體積的重建技術，並最終探討了如何從機器的視角出發，對受損影像進行處理，以提升下游機器學習任務的性能。透過一系列廣泛的實驗，本研究不僅在多種受損情況下顯著提升了影像質量，也展示了我們的系統如何直接增強下游任務的性能，為複雜環境中高品質影像處理和分析提供了新的見解和工具。

In the digital era, the demand for high-quality visual data has become increasingly critical for a variety of applications, including autonomous driving and medical diagnostics. However, images captured in real-world conditions often suffer from degradation due to adverse weather, low lighting, blurriness, and noise. These factors not only hinder human judgment but also impair the performance of machine learning models in tasks such as object detection and semantic segmentation. Addressing these challenges, starting with precise image quality assessment to employing advanced image restoration techniques, has become essential. With advancements in deep learning technology, we now have the capability to use deep learning models to learn complex patterns and features from data, automatically detecting and correcting image quality issues for significant improvements.

This study systematically develops applications ranging from image quality assessment to the restoration of two-dimensional images and the reconstruction of three-dimensional volumes using deep learning technology for degraded images. The process encompasses initial image quality assessment, enhancement of two-dimensional image quality, and then reconstruction of three-dimensional volumes. It further explores how to process damaged images from a machine-centric perspective to enhance the performance of downstream machine learning tasks. Through a series of extensive experiments, this research not only significantly improves image quality under various degraded conditions but also demonstrates how our system directly enhances the performance of downstream tasks, providing new insights and tools for high-quality image processing and analysis in complex environments.