Automated Image Screening Software for Parallel Hardware

This SBIR Phase I proposal focuses on the development of a prototype automatic image screening system that runs multiple image transform and feature extraction algorithms in parallel. Expected advances in high performance parallel desktop computer hardware provide the potential to automatically screen large numbers of high-resolution images at a reasonable cost. To harness this potential, we propose to design and prototype a software system that preprocesses images, then performs multiple different image transforms and feature extraction methods in parallel, and finally classifies image areas at multiple levels of discrimination based on the pool of extracted features. We will prototype parallel algorithms for a number of transforms (such as Gabor, wavelet, and Hough) and feature extraction methods (such as fractal dimension, entropy and other pixel gray level statistics). Wherever it is applicable, we will use features that are invariant to translation, rotation, and scale in the image plane. For classification, we will use a hierarchy of state-of-the-art neural network classifiers. We have already proven the feasibility and computational efficiency of this approach in automatic target recognition; we expect the main benefits to carry over to the domain of automatic image screening: extremely rapid training, self-organizing architecture, and ability to handle large numbers of classification labels.