Distortion-Invariant Pattern Recognition with Adaptive Correlation Filters

Pattern recognition based on correlation is one of the most useful techniques for many applications. Since the pioneer work of VanderLugt (1964), correlation filters have gained popularity thanks to their shift-invariance property, good mathematical basis, and easy implementation by means of digital, optical or hybrid optical/digital systems. However, conventional correlation filters are sensitive to intensity signal degradations (blurring and noise) as well as to geometrical distortions of an object of interest. Basically, blurring is owing to image formation process, and it can be produced by imperfection of capturing devices, relative motion between a camera and an input scene, propagation environment, etc. An observed input scene always contains noise produced by an imaging system (i.e. imperfection of imaging sensors) or by a recording medium (i.e. quantization errors) (Bertero & Boccacci, 1998; Perry et al., 2002). On the other hand, geometric distortions change the information content and, therefore, affect the accuracy of recognition techniques. Two types of geometric distortions are distinguished: internal and external distortions. The internal distortions are produced by the geometrics of a sensor; they are systematic and can be corrected by a calibration. External distortions affect the sensor position or the object shape; they are unpredictable (Starck et al., 1998). This chapter treats the problem of distortion-invariant pattern recognition based on adaptive composite correlation filters. The distinctive feature of the described methods is the use of an adaptive approach to the filters design (Diaz-Ramirez et al., 2006; González-Fraga et al., 2006). According to this concept, we are interested in a filter with good performance characteristics for a given observed scene, i.e., with a fixed set of patterns or a fixed background to be rejected, rather than in a filter with average performance parameters over an ensemble of images. Specifically, we treat two problems: reliable recognition of degraded objects embedded into a linearly degraded and noisy scene (Ramos-Michel & Kober, 2007) and adaptive recognition of geometrically distorted objects in blurred and noisy scenes (Ramos-Michel & Kober, 2008). The first problem concerns with the design of optimum generalized filters to improve the recognition of a distorted object embedded into a nonoverlapping background noise when the input scene is degraded with a linear system and noise. The obtained filters take into account explicitly information about an object to be recognized, background noise, linear system degradation, linear target distortion, and sensor noise. For the filter design, it is O pe n A cc es s D at ab as e w w w .ite ch on lin e. co m