PROJECT TITLE :

Fundamental Visual Concept Learning From Correlated Images and Textí_

ABSTRACT:

The visual notions in heterogeneous web media, such as objects, situations, and activities, cannot be dissected semantically. Learning fundamental visual concepts (FVCs) is a key part of understanding any visual material, as well as applications such as retrieval, annotation, and so on. FVC learning is formulated in this work, and a method to this problem called nearby concept distributing is proposed in this paper (NCD). The visual patches in images are nodes in a concept graph that creates inter-image and intra-image edges between visual patches in different images and between visual patches in the same image. Semantic information from images is distributed to visual patches based on the idea graph's fitness, uniqueness, smoothness, and sparseness, without the use of pre-trained concept detectors or classifiers. It is possible to learn all ideas with an arbitrarily high probability as the size of the data increases, according to our analysis of the proposed approach. Using three publicly available datasets, we demonstrate the effectiveness of the NCD strategy. In the experiments, we found that our method outperforms current methods for learning visual concepts from linked media.