Image Search Reducing Processing Load via Feature Segmentation

The patent segments the search target set into multiple groups based on object features (size, shape, color). The first reducer divides search targets into size-based groups, and the second reducer further segments them by shape characteristics. This segmentation allows the system to process only relevant groups for similarity calculation, reducing overall processing load while maintaining search accuracy through feature-based filtering.

The patent applies preliminary action by performing feature-based filtering before similarity level calculations. The index image selector and reducers pre-process the search targets by matching object features (size, shape, color) against the imaged object, narrowing down the candidate set beforehand. This preliminary filtering ensures that only highly relevant targets undergo computationally intensive similarity calculations.

The patent extracts only the essential depth information from the multiple imagers without processing all image data in full detail. The system uses the parallax information from the plurality of imagers to determine size and shape features, then extracts only the necessary depth characteristics for search target filtering, avoiding unnecessary processing of complete stereoscopic image pairs.

The system performs preliminary processing of stereoscopic images to extract size and shape features before the main search operation. By pre-processing the depth information from multiple imagers to establish object characteristics, the system reduces the computational burden during the actual search and similarity calculation phases.

The patent applies local quality by using different filtering strategies for different regions of the search space. The first reducer applies size-based filtering with broader criteria to maintain match completeness, while the second reducer applies more stringent shape-based filtering only to the already-narrowed candidate set. This graduated approach ensures that potential matches are not eliminated by overly aggressive early filtering.

The system performs partial filtering by applying feature-based reduction only to the extent necessary to achieve acceptable processing speeds. The reducers are configured to maintain a sufficient number of candidate targets, ensuring that while calculation speed improves, the candidate set remains large enough to include all potential matches for the final similarity evaluation.