像素
人工智能
计算机视觉
计算机科学
脉冲噪声
非本地手段
滤波器(信号处理)
降噪
双边滤波器
中值滤波器
椒盐噪音
模式识别(心理学)
图像处理
图像(数学)
图像去噪
作者
Srinivasa Rao Gantenapalli,Praveen B. Choppala,James Stephen Meka,Paul D. Teal
标识
DOI:10.1109/icort56052.2023.10249109
摘要
This paper deals with impulse noise reduction in digital color images, with an application to radar imagery. Impulse noise is a random perturbation of the pixel intensity in either one or all of the color channels of the image. It is caused due to sensor sensitivities or atmospheric effects and impedes image perception and quality. The vector median filters are by far the most popular of the filters for denoising (or reducing) impulse noise in images. These filters work on the principle of similarity detection within a window, or block of pixels to determine a suitable pixel to replace the test pixel, which is usually the centre pixel. These filters, however, are time consuming as the window is made to sequentially slide across all the pixels in the image. A recent variant of this class of filters are the peer group filters, in particular the fast peer group filter. This filter applies the similarity detection procedure over only a set of those pixels that are peers (or members) to the test pixel. This membership criterion facilitates in accelerating the filtering process through minimising the computation to within the member pixels in the window. This approach, albeit being the fastest to date in the class of vector median filters, is still computationally expensive as it involves sequential windowing across the entire window. This paper proposes a faster variant to this technique. The key idea here is to apply a heuristic anomaly detection technique over the rows of the image to determine the anomalous (noisy) pixels and then apply the peer group filter over only those anomalies. This approach reduces the order of computation and ensures accelerated filtering by virtue of processing only those pixels that are determined to be noisy. The efficacy of the proposal is measured using a combined measure of time and accuracy.
科研通智能强力驱动
Strongly Powered by AbleSci AI