Machine Learning for a Toolkit for Image Mining
Abstract
Given the exponential growth in the number of images stored in databases around the world, there is a clear need for computer-assisted means of searching images. Keywords and computed indices, referred to as metadata, can be used to tag images for subsequent retrieval. Most potential applications, however, will require searching for images on the basis of image content, but existing techniques for content-based search are either too slow, customized at great expense for a single application, or simply do not work well. An environment has been created enabling a user with limited computer skills to interactively train a computer algorithm to recognize patterns of spectral or textural features in imagery. The trained algorithm can then be exported as an independent agent to search large databases for matching image patterns. Following their retrieval, a user can further refine agent performance by indicating mistakes. The final product is a search toil capable of prescreening and highlighting images, significantly reducing the workload of analysis attempting to detect regions or objects in imagery. An approach to interactive learning has been developed as part of this environment. Based on techniques of knowledge-based image processing, this approach uses interest images to provide a means of continuous feedback of algorithm performance to the user, who in turn responds by indicating where the algorithm is making mistakes. The set of user-indicated examples and counterexamples form the inputs to a new machine learning algorithm called functional template learning. This algorithm is competitive with other machine learning techniques in terms of classification accuracy and exhibits several advantages in terms of speed and understandability that make it particularly well suited us interactive, supervised machine learning and autonomous searching of databases.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 06, 1995
- Accession Number
- ADA293137
Entities
People
- Richard . J. Sasiela
- Richard L. Delaney
Organizations
- Massachusetts Institute of Technology