基於消息理論之無參考視覺品質評估於相機雜訊影像

Abstract

由於所有相機拍攝的影像不可避免地會受到各種雜訊的影響，無參考雜訊影像品質評估在多媒體處理領域具有重要的實際價值。我們提出的方法基於以下兩個核心理念： 1. 影像的客觀品質主要取決於人類視覺系統 (HVS) 感知的有效資訊量與雜訊。2. 影像雜訊水平估計的精確性對於雜訊失真影像的品質預測性能至關重要。我們設計了一個基於人類視覺系統的影像資訊容量，利用影像功率譜、雜訊估計和對比敏感度函數 (CSF)，以確定雜訊失真影像中的最大感知資訊量。此外，本文還提出了一種基於離散小波變換 (DWT) 的精確影像雜訊估計算法。影像在不同雜訊類型下的變化，則可透過局部均值減法和對比度正規化 (MSCN) 係數的分布來捕捉影像自然特徵。我們採用了柯爾莫哥洛夫-阿諾德網絡 (KANs) 進行訓練，將特徵向量映射到對應的主觀評分。實驗結果顯示，在 TID2008、 TID2013 和 KADID-10k 資料庫上，我們的方法 ICNEN-IQA 在雜訊影像品質預測方面優於其他最先進的影像品質評估方法。

Since all camera-captured images are inevitably subjected to various types of noises, no-reference noisy image quality assessment holds significant practical value in the field of multimedia processing. Our proposed method is based on the following two ideas: 1. The objective image quality is largely determined by the amount of useful information perceived by the Human Visual System (HVS) and noises. 2. Accurate image noise level estimation is the key to the perceptual quality of a noisy image. We design a HVS-based Shannon Information Capacity which leverages image power spectrum, noise estimation and a Contrast Sensitivity Function (CSF) to determine the maximum perception information in noise-distorted images. An accurate DWT-based image noise estimation algorithm is also proposed in this paper. Additionally, the variations in images under different types of noises can be captured via naturalness measurements, as characterized by the distributions of locally mean subtracted and contrast normalized (MSCN) coefficients. Kolmogorov-Arnold Networks (KANs) are used for training to map our feature vectors into corresponding subjective scores. Experimental results on TID2008, TID2013, and KADID-10k databases demonstrate that our method using image information capacity, noise estimation, and naturalness measurements, ICNENM-IQA, outperforms other state-of-the-art image quality assessment methods for noisy image quality prediction.