摘要：The speckle noise is considered one of the main causes of degradation in ultrasound image quality. Many despeckling filters have been proposed, which are always making a trade-off between noise suppression and loss of information. A class of despeckling methods based Non-Local Means (NLM) algorithm is known to efficiently preserve the edges and all fine details of an image while reducing the noise. The core idea of NLM filter is to estimate the denoised pixel by performing a weighted average of similar patches in the neighborhood around the noisy pixel. However, the presence of noise degrades the similarity measurement process of the NLM and thereby decreases its efficiency. In this work, a novel despeckling scheme for ultrasound images is proposed, by introducing the kernel principal component analysis (PCA) to the NLM and computing the similarity in a high dimension kernel PCA subspace. The kernel representation is robust to the presence of noise and it can give better performance even under high noisy conditions. And it takes into account higher-order statistics of the pixels which can lead to accurate edge preservation. In this work, a novel despeckling scheme for ultrasound images is proposed using the kernel PCA-NLM extended to speckle noise model. The visual inspection and image metrics will show that the proposed filter is very competitive with respect to one of state-of-the-art methods, the Optimized Bayesian Non Local Means filter (OBNLM), in terms of low contrast object detectability, speckle noise suppression, edge’s preservation.
