基於自適應視窗大小與霧霾合成技術的監督式與半監督式影像除霧演算法

Abstract

霧霾影像常伴隨嚴重的色彩偏移與細節喪失，進而顯著影響如物件偵測、追蹤與場景理解等下游視覺任務的表現。因此，作為影像復原的經典任務之一，影像除霧仍是電腦視覺領域中的重要研究議題。 本論文提出兩種影像除霧方法。第一個方法為自適應視窗大小暗通道先驗網路（Adaptive Patch Size Dark Channel Prior Network, APSDCP-Net）。該方法為一結合傳統暗通道先驗與深度學習模型架構的混合式網路設計，可針對每個像素自適應輸出平滑權重，加權整合多個視窗大小下的暗通道結果，進而提升透射率圖（transmission map）的估測準確性與整體除霧效果。 第二個方法為針對真實世界應用所設計之半監督式除霧框架，包含三個核心模組。首先，我們提出一套進階的霧霾資料合成流程。基於對現有真實世界資料集的分析，針對大氣光顏色、不均勻光照及非均勻霧分布，提出更擬真的模擬策略。其次，我們設計一個雙階段除霧網路，將除霧任務拆解為霧霾去除與雜訊抑制兩個子任務。最後，我們引入半監督式學習策略，利用未標註的真實霧霾影像所產生的偽標籤對模型進行微調。實驗結果顯示，本研究提出的資料合成方法顯著提升模型對真實霧霾影像的泛化能力，且整體架構在多項量化指標與視覺品質上皆優於現有先進方法。

Hazy images often suffer from severe color distortion and detail loss, which significantly degrades the performance of downstream vision tasks such as object detection, tracking, and scene understanding. Therefore, image dehazing, as a classical task of image restoration, remains a crucial topic in computer vision. In this thesis, we propose two methods to address the image dehazing problem. The first method, Adaptive Patch Size Dark Channel Prior Network (APSDCP-Net), is a hybrid dehazing network that integrates the traditional dark channel prior with a deep-learning architecture. The model computes dark channels using multiple patch-sizes and adaptively combines them with learned soft weights, resulting in more accurate transmission map estimation and enhanced dehazing performance. The second method is a semi-supervised dehazing framework designed for real-world applications. It incorporates three core components. First, an advanced haze synthesis pipeline is proposed based on empirical analysis of existing real-world datasets. Our pipeline improves data realism by modeling key characteristics such as varicolored haze, non-uniform illumination, and non-homogeneous haze distributions. Second, a two-stage dehazing network is designed to decouple the problem into two subtasks: haze removal and artifact suppression. Finally, a semi-supervised learning strategy is introduced to fine-tune the model using pseudo-labels generated from unlabeled real hazy images. Experimental results show that our synthesis pipeline significantly improves the model's generalizability to real-world hazy images, and the overall framework outperforms existing state-of-the-art methods in both quantitative metrics and visual quality.