Despite rapid progresses in image retrieval techniques, visual object search, whose goal is to accurately locate a target object within a collection of images, remains a challenging problem.
This is due to the fact that target objects such as logos usually occupy only a small portion of an image with background clutter, and can differ significantly from the query in scale, rotation, viewpoint and colour. These variations lead to difficulties in object matching, thereby raising the need for highly discriminative visual features for accurate matching. Instead of matching features individually, using the spatial context (e.g., a group of local features) has shown to enable more discriminative matching.
This invention proposes a randomized approach to deriving spatial context, achieved by spatial random partition. The similarity between spatial contexts is measured by the average matching scores over multiple random patches.
Technology Features & Specifications
- Each image is represented as a collection of local interest points. Each local descriptor is quantized to a visual word using a vocabulary of V words. Using a stop list analogy, the most frequency visual words that occur in almost all images are discarded. All feature points are indexed by an inverted file so that only words that appear in the queries will be checked
- Each image is randomly partitioned into M*N non-overlapping rectangular patch. Such partitioning is performed K rounds independently. The sized and aspect rations of the patches in the subset are random due to the independent random partitions, making the spatial context of each pixel at different scale being taken into consideration for query
- Due to the independence of each round of partition, the patches from different partition rounds can be processed in parallel. It increases the speed and efficiency of image search and image matching
- Visual object search on social and e-commerce sites
- Context aware advertising
- Search engine for image matching
Market Trends and Opportunities
- Ability to rapidly search images or videos on the web or video feeds through object recognition system
- Removes the need to manually input text metatags for pictures and videos
- Improves the ability of advertisers to target contextually relevant ads
- Anomaly detection capability for security type applications
Make an Enquiry