In recent years, the use of deformable models (also called deformable templates) for recognizing non-rigid patterns and objects has aroused a lot of interests. Existing deformable models are powerful in that a template can deform itself to match with different variants of an object, if the template is initialized properly. Since most existing deformable models have relatively simple structure, initialization can usually be done using simple rigid template matching methods. However, most real-world objects have much more complex structure and deformable shape, and they usually appear with other (possibly noisy) background objects. These make the problem very challenging in practice....[ Read more ]

In recent years, the use of deformable models (also called deformable templates) for recognizing non-rigid patterns and objects has aroused a lot of interests. Existing deformable models are powerful in that a template can deform itself to match with different variants of an object, if the template is initialized properly. Since most existing deformable models have relatively simple structure, initialization can usually be done using simple rigid template matching methods. However, most real-world objects have much more complex structure and deformable shape, and they usually appear with other (possibly noisy) background objects. These make the problem very challenging in practice.

In this work, we propose a featzlre-based elastic model for line-drawing matching and classification, which can be seen as a major extension of the elastic net proposed by Durbin and Willshaw in 19S7. Model matching is divided into three phases, namely, initialization, deformable template matching, and refinment. We introduce a feature distance measure, which is combined with the original spatial distance measure to achieve more reliable point correspondence between the template and the input image. The feature distance measure is of pivotal importance in template initialization and object detection. By modifying the matching strategy of the original elastic net, our elastic model is less sensitive to backround distractions and hence is more desirable for extracting an object from multiple objects. In addition, the structure of an object is explicitly encoded in our elastic model formulation. Hence we can get richer information about the matched objects. When tested on strokebased characters, a refinement phase is introduced to give better correspondence between the matched strokes. This refinement phase is an iterative process of stroke matching evaluation, splitting, and deformable template matching.

We have performed a feasibility study on our approach by studying a difficult off-line character recognition problem: recognizing (i.e. detecting, extracting, and then classifying) character components, called radicals, in handwritten Chinese characters. Our approach shows encouraging results when tested on artificially generated character images with distortions, broken strokes, and other noises. When tested on real handwriting data, the recognition result is not as good due to feature extraction errors and inadequate modelling (since no model training is performed). Nevertheless, the elastic model has demonstrated its potential for complex deformable object recognition. Our experience with the real handwriting data has also led us to identify several important research issues for future study along the direction of deformable pattern recognition.