With the availability of more powerful computers it is nowadays possible to perform pixel based operations on real camera images even in the full color space. New adaptive classification tools like neural networks make it possible to develop special-purpose object detectors that can segment arbitrary objects in real images with a complex distribution in the feature space after training with one or several previously labeled image(s). The paper focuses on a detailed comparison of a neural approach based on local linear maps (LLMs) to a classifier based on normal distributions. The proposed adaptive segmentation method uses local color information to estimate the membership probability in the object, respectively, background class. The method is applied to the recognition and localization of human hands in color camera images of complex laboratory scenes.

Visual stimuli lead to neural activity in the retina that is propagated in separate ON and OFF pathways to the cortex. Most models of biological early vision recombine these activity streams by a linear integration at the simple cell level. Based on empirical as well as theoretical investigations we propose a nonlinear recombination circuit that is selectively responsive to contrast magnitude as well as to the sharpness of luminance transition. Simulations with artificial and camera images show a higher positional selectivity for local contrasts than an equivalent linear device. In a multiscale hierarchy the nonlinear circuit produces a unique maximum response in scale-space where scale directly relates to the width of the luminance transition. In order to investigate the biological relevance of the proposed neural circuit, we measured the model sensitivity to luminance gradient reversal in bar stimuli. Our simulations show strong similarity to simple cell recordings in the feline striate cortex. This result further supports the evidence for nonlinear interaction at the simple cell level.

Incrementally constructed cascade architectures are a promising alternative to networks of predefined size. This paper compares the direct cascade architecture (DCA) proposed in Littmann and Ritter (1992) to the cascade-correlation approach of Fahlman and Lebiere (1990) and to related approaches and discusses the properties on the basis of various benchmark results. One important virtue of DCA is that it allows the cascading of entire subnetworks, even if these admit no error-backpropagation. Exploiting this flexibility and using LLM networks as cascaded elements, we show that the performance of the resulting network cascades can be greatly enhanced compared to the performance of a single network. Our results for the Mackey-Glass time series prediction task indicate that such deeply cascaded network architectures achieve good generalization even on small data sets, when shallow, broad architectures of comparable size suffer from overfitting. We conclude that the DCA approach offers a powerful and flexible alternative to existing schemes such as, e.g., the mixtures of experts approach, for the construction of modular systems from a wide range of subnetwork types.