A block structured neural feature extraction system is proposed whose distortion tolerance is built up gradually by successive blocks in a pipeline architecture. The system consists of only feedforward neural networks, allowing efficient parallel implementation. The feature extraction is based on distortion-tolerant Gabor transformation and minimum distortion clustering by hierarchical self-organizing feature maps (SOFM). Due to unsupervised learning strategy, there is no need for preclassified training samples or other explicit selection for training patterns during the training. A subspace classifier implementation on top of the feature extractor is demonstrated. The current experiments indicate that the feature space has sufficient resolution power for a small number of classes with rather strong distortions. The amount of supervised training required is very small, due to many unsupervised stages refining the data to be suitable for classification.