Development and Parallelization of an Improved 2D Moving Window Standard Deviation Python Routine for Image Segmentation Purposes

作     者：Marcos R. de A. Conceição Luis F. F. de Mendonça Carlos A. D. Lentini Marcos R. de A. Conceição;Luis F. F. de Mendonça;Carlos A. D. Lentini

作者机构：Oceanography Department Geoscience Institute of the Federal University of Bahia (UFBA) Salvador Brazil Earth and Environmental Physics Department Physics Institute of the Federal University of Bahia (UFBA) Salvador Brazil Tropical Oceanography Group (GOAT) Salvador Brazil 

出 版 物：《Computational Water, Energy, and Environmental Engineering》 (水能与环境工程（英文）)

年 卷 期：2020年第9卷第3期

页      面：75-85页

学科分类：08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

主　　题：Digital Image Processing Image Segmentation Standard Deviation Python Machine Learning 

摘      要：Two additional features are particularly useful in pixelwise satellite data segmentation using neural networks: one results from local window averaging around each pixel (MWA) and another uses a standard deviation estimator (MWSD) instead of the average. While the former’s complexity has already been solved to a satisfying minimum, the latter did not. This article proposes a new algorithm that can substitute a naive MWSD, by making the complexity of the computational process fall from O(N2n2) to O(N2n), where N is a square input array side, and n is the moving window’s side length. The Numba python compiler was used to make python a competitive high-performance computing language in our optimizations. Our results show efficiency benchmars