K-최근방 이웃 방법을 사용한 장면 분류 시스템의 문턱값 접근을 통한 실제 환경에서의 성능 향상 방법

Abstract

An Input image is blocked into several blocks and features are extracted from these blocks. Blocks are classified by K-NN classifier using training data with predefined labels, and the most frequently selected block label becomes the label of the image. K-NN based scene classification system is not perfect in a practical situation because there are lots of ambiguous images which even a man cannot tell (indoor from outdoor), (city from landscape), (sunset from mountain&forest), (forest from mountain). Thresholding approach is added to explicitly say that ambiguity exists, and this image has ambiguous label. This increases performance and completeness of previous K-NN based scene classification system.