Segmentation of Digital Rock Images Guided by Edge Feature Using Deep Learning

Abstract

Summary Segmentation of digital rock images is a crucial and basic step in digital rock process, and equivalent elastic parameter and fluid properties calculated from the digital rock can be affected by the result of segmentation. Conventional segmentation algorithm based on thresholding algorithm cannot perform a satisfying result in small structure due to noise impact. To address issues, a modified guided by prior information, edge feature, is proposed to improve accuracy of small structure. Edge feature reflects information of the effect of transport, weathered, and eroded in the deposition process, but the shape of noise and artifacts can’t reflect these information, rather show regularity due to the influence of instruments, hence boundary feature can improve the discrimination of noise. Furthermore, conventional SegNet was used to compare with modified SegNet, the former obtains 90.21% accuracy using 38-layers network, proposed approach using prior information achieves 93.07% accuracy using 30-layers network, which demonstrates less computational time and better anti-noise property. In addition, connectivity was used to evaluate segmentation result, modified SegNet shows a better similarity with origin image.