A database approach to hierarchical image storage and recognition

This paper presents the design of a model-based vision-recognition engine for implementing effective indexing schemes over a PC-based database system. The objective is to achieve robust recognition of planar contours or fragments of contours that are scale invariants of known models. Planar contours are partitioned into straight or slightly curved fragments (segments) and curved fragments (corners). The boundary problem is handled by using a constant-curvature criterion. To reduce the search space without pruning correct matches, a set of local and global descriptors is extracted from the model to enable indexed searching with different scaling, rotation, and translation from the model. Only a sub-set of matched classes with the highest frequencies are retained for finer matching which consists of distance matching, scale, and relative positioning. For partially occluded contours, fragments are individually matched by indexing over the local descriptors. Low-cost geometric inter-relationships among scene fragments are compared with those of the matched models, which allows consolidation of some matches, and invalidation of others. This approach enables a further reduction in the size of the matched classes. An evaluation of the method shows its ability to recognize entire contours, as well as partially occluded contours, from among a reasonably large dictionary. The approach proved to be tolerant with respect to partial occlusion, digitization noise, and slight variations in scale and spatial layout. The PC-based database approach to pattern recognition is shown to be efficient for real-time applications, and greatly simplifies the design.