Recognizing and Incrementally Evolving Texture Concepts in Dynamic Environments: A Model Generalization Approach
Abstract
The paper presents a novel approach to the invariant recognition of objects (through texture) in dynamic environments. The proposed approach assumes that (1) the system has to recognize objects on each image of a sequence of images, (2) the images demonstrate the variability of conditions under which objects are perceived (resolution, lighting, positioning), (3) both an observer and the objects can move, (4) the extraction of texture attributes and training examples can be imperfect. and (5) the system has to work autonomously (i.e., without the aid of a teacher). We propose to utilize images of a sequence to adapt system models to perceived variabilities of texture characteristics. Such an adaptation integrates recognition and segmentation processes of computer vision with incremental knowledge acquisition processes of machine learning. While the initial acquisition of texture models is driven by a teacher, the evolution of these models is performed over a sequence of images without the help of a teacher. Initially acquired texture descriptions are applied to recognize and extract objects on the next images. The effectiveness of such recognition and object extraction is monitored. When this effectiveness decreases, the system selects new training data and activates learning processes to improve its models. This paper presents both an outline of the iterative evolution methodology and the investigation of an incremental model generalization approach as a part of the evolution of texture models. Experiments were run in a partially-supervised mode rather than a fully autonomous model evolution. The experiments are compared based on the following three system configurations: (i) a one-level control structure, (ii) a two-level control structure, and (iii) a two-level control structure with data filtering. Obtained results are evaluated using the criteria of system recognition effectiveness, recognition, stability and predictability of evolved models.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 1991
- Accession Number
- ADA529590
Entities
People
- Peter W. Pachowicz
Organizations
- George Mason University