Images are complex phenomena possessing greater information than definable by words alone. Thus, libraries retain original images in addition to verbal files which index and describe them. A large collection of images in electronic form challenges traditional storage since the image collection is """"""""virtual"""""""" and only accessible through its text index, no matter how inadequate that may be. As each new image is added to the collection, it is clearly impossible to textually define all possible features that might be relevant for later searches. Moreover, a diverse image collection assembled and assigned verbal indices by different individuals will incorporate unwanted variations. Thus, an image index based solely on semantics will necessarily be poor, incomplete and non-robust. New methods must be explored to index these collections on a non-textual pictorial content basis. This project proposes an approach to image indexing which combines semantic assignment to create coarse groupings with mathematical image processing operations to create ranking by pictorial similarity based on numerical descriptors of image features. Reassembly of pictures thus grouped could offer a powerful browsing tool. Ranking may be interactively extended to a selected subgroup of the image collection, or automatically on the basis of a statistical measure. A combination of these search methods permits iterative convergence to smaller, manageable subsets of images. Early efforts will be performed on a large electronic image collection (the Video-disc Echocardiography Encyclopedia) which resulted from preliminary work with the National Library of Medicine. That collection is composed of 54,000 images already highly indexed by a text database of 33 fields. Rapid operator-directed digital image processing of these digitized images will reduce structurally relevant features to sketches or cartoons. This secondary geometrical file will act as an indexing pointer to the original images. A browser would find within the collection a sample image (accessed by the text index) which embodies a feature he is interested in retrieving. The system would then display other images from the collection ranked in order of mathematical similarity to the prepositus image. Iterative selection and automatic reordering of these image groups would allow convergence on progressively smaller subsets of images possessing the desired feature.
| Tagare, H D (1997) Deformable 2-D template matching using orthogonal curves. IEEE Trans Med Imaging 16:108-17 |
| Robinson, G P; Tagare, H D; Duncan, J S et al. (1996) Medical image collection indexing: shape-based retrieval using KD-trees. Comput Med Imaging Graph 20:209-17 |
