A multiresolution neural network approach to invariant image recognition
(EN)
Kollias, SD
(EN)
Triple-correlation-based representations of images have recently been combined with neural network architectures to derive invariant, with respect to translation, rotation and dilation, robust classification of images. Multiresolution image analysis is used in this paper to reduce the size of these representations in an optimal way, based on autoassociative linear networks. Hierarchical neural networks are then proposed as an efficient architecture for classification or retrieval of multiresolution invariant image representations. An effective procedure for designing and training such networks is also described and simulation results are presented which illustrate the capabilities of the proposed approach.Triple-correlation-based representations of images have recently been combined with neural network architectures to derive invariant, with respect to translation, rotation and dilation, robust classification of images. Multiresolution image analysis is used in this paper to reduce the size of these representations in an optimal way, based on autoassociative linear networks. Hierarchical neural networks are then proposed as an efficient architecture for classification or retrieval of multiresolution invariant image representations. An effective procedure for designing and training such networks is also described and simulation results are presented which illustrate the capabilities of the proposed approach.
(EN)
