MSU Technologies

Compression of Image Ensembles Using Tensor Decomposition

Introduction

Large still-image databases are often used for e-commerce, medical imaging, technical drawing, biometrics, and security-related applications. Efficient compression and retrieval methods are needed to store these images over the long-term. A key requirement for still-image compression is the ability to access and decode individual images within a collection, without having to access/decode other images in the database. The individual access requirement eliminates popular video coding techniques as potential candidates for still-image database compression. Often there is a significant amount of repeated and redundant data within the collection.

Description of Technology

This technology is a method of matrix decomposition that can be used to compress a collection of images of the same type using a tensor to exploit redundancy between images in a collection. An optimization method for rank allocation and a method to individually estimate rank-one tensors progressively are used to enhance the Canonical-decomposition Parallel-factor (CP) method. These improvements to the standard CP method allow the inventors to apply the new compression scheme to databases of similar still-images with redundant data and significantly reduce the storage required to save the database, while also maintaining access to individual images within the databases without having to decode all of the images in the database. A variable rank parameter is used to take advantage of redundant data and ensure optimal results.

Key Benefits

Random image access – this method allows random access to any image within the collection without the need to reconstruct other images in the same collection
More efficient compression – this method exploits the strong correlation that might exist among images within an ensemble

Applications