Hideo Umeki

Toward Video Bookmarking Search: Search-Target Inference from TVWatching

Our goal is to provide a one-button interface called Video Bookmarking Search, which efficiently retrieves information desired by a user with a simple user interface and an intelligent background process without distracting the users attention from watching TV. We propose a method that infers keywords a user wants to search when watching TV by the Named Entity Recognition and the Semantic Role Analysis techniques for closed captions of Japanese TV shows on city life, travel, or cuisine to perform keyword selection for appropriate keyword-based Web search. According to experiments using actual TV shows with over 200 bookmarks, our method achieved 74.6% recall in the 10 retrieved results of searchable keywords that users want to search.The extended fuzzy description logics extend the classical description logics by introducing cut concepts and cut roles to support expressive representation and reasoning for fuzzy knowledge. But the current extended fuzzy description logics lack reasoning algorithms with TBoxes. The paper proposes the sat-domain problem of the extended fuzzy description logic EFALCR+ with restricted TBoxes, and gives an optimized reasoning algorithm, which is proved to be sound, complete and with a worst complexity of EXPTime; then proves the complexity of the problem is EXPTime- complete.

Dynamic link matching for multiple object recognition

We present an extended version of dynamic link matching for the recognition of all objects similar to a stored model. With a given complex image containing multiple objects, our system can detect all of those matched with a model regardless of their type of geometric transformation and distortion. This is achieved by allowing simultaneous strengthening of multiple bundles of interlayer links (dynamic links) when local features in the model have more than one counterpart in the image layer. To extract the matched regions, in which dynamic links are established as geometrically invariant mappings, we propose a neural system based on edge dynamics in the edge-mapping parameter space.

Object recognition by dynamic link matching with multiple blob formation

We present a one-to-many object matching system based on neural dynamic link architecture. When an input image containing multiple objects is given, if some of them are similar to a stored model, the system can establish geometric transformation-invariant mappings between the model and the corresponding object regions in the input image. This can be achieved by extending the fast dynamic link matching (FDLM) algorithm to allow multiple blob formation. Numerical simulations of neural layer dynamics indicate that multiple blobs can be developed where the layer input is sufficiently strong against the background level. To extract matched regions from the input layer, we consider each neural layer as a graph and introduce another neural system based on local edge mappings. This system can roughly detect the matched regions with neighborhood-preserving mappings without global cost functions.