Integrating Low-Level Features Computation With Inductive Learning Techniques for Texture Recognition

Abstract

This paper presents a method for applying inductive learning techniques to texture description and recognition. Local features of texture are computed by two well-known methods, Laws’ masks and co-occurrence matrices. Then, a three-level generalization of local features is applied to create texture description rules. The first level generalization, the scaling interface, has been implemented to transform the numeric data of local texture features into their higher symbolic representation as numerical ranges. This scaling interface tests data consistency as well. The creation of description rules incorporating the inductive incremental learning algorithm is the second generalization step. The SG-TRUNC method of rule reduction is applied as the next hierarchical generalization level. This machine learning approach to texture description and recognition is compared with the classic pattern recognition methodology. The results from the recognition phase are presented from six classes of textures, characterized by smoothly changing illumination and/or texture resolution. The average recognition rate was 91% for the inductive learning approach, and all classes of textures were recognized. In comparison, the traditional k-NN pattern recognition method did not recognize one class of texture, and the average recognition rate was 83%. The proposed methodology smooths the recognition rates through the hierarchy of generalization levels, i.e. the next generalization step increases these rates for classes that were less easily recognized, and decreases these rates for classes that were more easily recognized.

Application of symbolic machine learning to the recognition of texture concepts

“…then the dynamic increase of the ttuncation degree (but performed in the lowest range of optimization values) should slightly increase the recognition effectiveness in this range. Sucb effect was observed in our fll'St experiments [10]. We concluded then that a recognition curve created for each class description and for different values of the truncation degree could be a useful classification feature for the ftnal decision making process.…”

“…Considering the application of leaming-based concept acquisition in texture domain, we have already shown the methodology and benefits of such an approach [10]. In one of our experiments, the system was able to improve the average recognition rate (for six classes of texture acquired from very poor image data) from 70% of correct recognitions obtained for the lc-NN pattern recognition method [5J, to 80% for the symbolic machine learning approach, and to 91 % for the symbolic machine learning approach incorporating optimization of texture class descriptions and a single matching.…”

“…This work follows our fonner observations and results [10]. Developed object recognition methodology is being applied to create an intelligent vision system for outdoor navigation of autonomous robots and for remote surveillance systems [1l].…”

On the need for integrated approaches to image understanding

“…A ML system tackling low-level imaging has been developed by Pachowitz [70] on texture recognition. In practice, the system applies the AQ15 inductive learning algorithm [60] to symbolic texture representations to derive discriminant descriptions for classification purposes.…”