Improved Hieroglyph Representation for Image Retrieval

Abstract

In recent years, an interdisciplinary effort between archaeologists and computer vision experts has emerged to provide image retrieval tools that facilitate and support cultural heritage preservation. The performance of these tools largely depends on the hieroglyph representation quality. In the literature, the most successful hieroglyph representation for retrieval following the BoVW model includes a thinning hieroglyph process and selects interest points through uniform random sampling. However, thinned hieroglyphs could have noise or redundant information, and a random set of interest points could include non-useful interest points that are different in each iteration. In this article, we propose improving this hieroglyph representation by pruning thinned hieroglyphs and introducing an improved interest-point selection. Our experiments show that our proposal significantly improves the hieroglyph retrieval results of state-of-the-art methods.