Computational Modeling of Topographic Arrangements in Human Visual Cortex

摘要

The majority of the current systems for image recognition and classification put images into discrete semantic categories pre-defined by researchers. On the one hand, the classification accuracy is strongly dependent on how researchers define ground truth. It's unavoidable that performances of such systems are compromised by the semantic gaps between actual image content and textual description by human interpreters. On the other hand, such algorithms ignore the relationships among the images within the same category, which are of great importance in analyzing the training database as well as understanding query visual instances. This paper proposes an image analyzing and understanding model that describes both the inter-class relationship and intra-class variation of the database. Inspired from topographic arrangements found in biological cerebral cortices, we develop an algorithm that organizes visually similar images into the same local neighboring regions, without human labeling or annotation. Furthermore, images of greater resemblance will be close to each other while dislike images are located far away on the map. The Self-Organizing Map is a variation of neural network that uses competitive learning scheme to learn internal relationship of input data without supervision. The framework is implemented and evaluated using a standard, publicly available image database. Experimental results demonstrate meaningful and effective mapping of input images and the topological arrangement of SOM.