Combining Feature- and Correspondence-Based Methods for Visual Object Recognition

Abstract

We present an object recognition system built on a combination of feature- and correspondence-based pattern recognizers. The feature-based part, called preselection network, is a single-layer feedforward network weighted with the amount of information contributed by each feature to the decision at hand. For processing arbitrary objects, we employ small, regular graphs whose nodes are attributed with Gabor amplitudes, termed parquet graphs. The preselection network can quickly rule out most irrelevant matches and leaves only the ambiguous cases, so-called model candidates, to be verified by a rudimentary version of elastic graph matching, a standard correspondence-based technique for face and object recognition. According to the model, graphs are constructed that describe the object in the input image well. We report the results of experiments on standard databases for object recognition. The method achieved high recognition rates on identity and pose. Unlike many other models, it can also cope with varying background, multiple objects, and partial occlusion.

Using Growing Neural Gas Networks to Represent Visual Object Knowledge

“…ically assembled object models [14]. Its properties include the unsupervised structural organization of object components according to their visual resemblance, and the use of this structure for matching novel components.…”

“…This measure allows for smooth similarity potentials with fairly wide maxima [14]. In comparison with previous and similar approaches (e.g., Kohonen Feature Map, and Growing Cell Structure), GNG is more flexible since no dimensionality assumptions need to be made, and it allows continuous learning by adding neurons and synapses until a performance criterion is met.…”

“…Finding the best matching model features for a set of novel features F R extracted from test object views is at the core of the object recognition system [14] and the categorization system [13]. In our work, they are matched through a recall procedure, where the previously learned information inferred from the self-organized Neural Map favors the selection of some model features over others until the best matching one is found.…”

“…These systems use either a featurebased [7,2] or a correspondence-based approach [14,5,15]. In both, the processing of an object view relies on the extraction of image features together with the use of stored object models derived from training object views.…”

“…The resulting artificial systems are either limited to represent a narrow range of object types, or they are in conflict with neuropsychological and neuroanatomical observations [4], as well as with the results of psychophysical experiments [9] and computational studies [11]. A recently proposed artificial system dynamically generates its object models using a more balanced approach [14]. This method yields high recognition rates, still it performs rather poorly on categorization and further research on its visual object knowledge representation is needed to improve its performance and robustness.…”