Segmentation of butterflies from complex agro-ecological images using quantum mechanics and spatial refinement

Abstract

The presence of butterflies is an important indicator of environmental health and stability. Its pollination ability increases agricultural yields. However, some butterflies may also cause crop degradation. Butterfly segmentation in complex agro-ecological environments is therefore a preliminary step whose accuracy affects species recognition. We propose in this paper a segmentation method composed of a sequence of three steps. First, we combine background connectivity with local distinctiveness measures to locate roughly butterfly areas. Second, we deploy an adaptation of the Schrödinger equation on a graph propagation process that checks regional similarities to those areas. Finally, we perform a spatial refinement that optimizes the butterfly detection. Both objective and subjective comparative experiments suggest that our method performs well against image artifacts. Obtained results outperform those of some state-of-the-art methods that reach high performance.