Jun Zhou

Date of Award

Summer 2022

Document Type

Open Access Dissertation


Computer Science and Engineering

First Advisor

Song Wang


The surface of many cultural heritage objects, such as pottery sherds found in the Southeastern Woodlands, were embellished with curve patterns. The original full designs of these patterns reflect rich historical and cultural information. However, in practice, most objects are fragmentary, making the complete underlying designs unknowable at the scale of the sherd fragment. The challenge to reconstruct and study complete designs is stymied because 1) most pottery sherds contain only a small portion of the underlying full design, 2) curve patterns detected on a sherd are usually incomplete and noisy, and 3) in the case of a stamping application, the same design may be applied multiple times with spatial overlap on pottery, resulting in a composite pattern. Our study aims to address these challenges and better identify and discover the full designs from fragmented pottery sherds. In this research, we study two important computer vision problems: design identification that identifies a sherd underlying design and sherd identification that clusters unidentified sherds to discover unknown designs. We focus both problems on curve patterns, where the underlying full design and the partial pattern on the sherd are curve structures, and develop new algorithms to address them.

For design identification, we formulate this problem as matching: a binary curve pattern image segmented from a sherd depth image is matched to each known full design, and the best matching proposes its underlying design. We develop two curvepattern matching algorithms for this purpose. First, we develop a new curve matching method by extending Chamfer matching, which decomposes a composite pattern into multiple candidate components as long as these components match a partial design. The optimal combination of these components defines a new matching cost. Second, we develop a new patch-based curve pattern matching method to locate the most similar regions between the sherd and the considered full design. Specifically, we apply uniform sampling for constructing patches and employ a learning-based curve feature descriptor to derive a heatmap for the local similarity between the sherd and the design. With this heatmap, we locate the best matching portions by region growing and define a new matching cost considering the overall similarity of these portions.

For sherd identification, we develop a new clustering algorithm to identify and group sherds with the same design. Given the segmented curve-pattern images of a collection of sherds, we first conduct patch-based pairwise matching between each pair of sherds to construct a similarity matrix. The pairwise matching is based on the best-matched patches between the two sherds to handle possible composite patterns. We build a fully connected graph based on this similarity matrix and partition the graph into subgraphs/clusters by adaptive thresholding. An iterative cluster refining strategy is developed, with curve-pattern stitching in the iteration, for identifying and refining the sherd clustering.

We collect a set of pottery sherds from the heartland of the paddle-stamping tradition with a subset of available paddle-stamped designs from southeastern North America to evaluate the developed algorithms. Moreover, we developed the Snowvision, a computer-aid system that includes sherd digitization, preservation, curve structure segmentation from a digitized sherd depth image, design identification, and sherd identification based on the developed algorithms.