Communications and Signal Processing Seminar

New Texture Similarity Metrics

David NeuhoffProfessorUniversity of Michigan, Department of Electrical Engineering & Computer Science

This talk introduces the problem of developing objective assessments of image similarity that reflect human judgments, and describes two recently developed approaches to assessing similarity of textured images. The first, called STSIM (Structured Texture Similarity Metric), lies in the family of Structured Similarity (SSIM) metrics. It is based on a subband decomposition and sliding-window local statistics within the subbands. The second, called LRI (Local Radius Index), is a local pattern based metric operating in the spatial domain. It is complementary to metrics based on the commonly used Local Binary Patterns (LBP) feature. The accuracy of these metrics is assessed based on how well they enable content-based retrieval of identical textures, texture classification, and coding/compression of textured images. The new metrics attain state-of-the-art performance, with the LRI-based metrics being computationally much simpler.

Sponsored by

Unversity of Michigan, Department of EECS