UNSUPERVISED SEGMENTATION OF CORAL REEF IMAGES BY USING COLOR AND TEXTURE FEATURES

Abstract

Segmentation of natural scenes is an essential task in image processing. It finds a place in many image applications such as retrieval, indexing, classification, surveillance and content-based image retrieval. However, there is clear lack of image segmentation techniques in the literature studies related to underwater coral reef images segmentation. This thesis presents new methods to automate the segmentation of underwater coral reef images based on image processing techniques with the combination of color-texture features. The unsupervised segmentation of color-texture regions using J-value segmentation (JSEG) algorithm is one of the most popular and robust unsupervised segmentation algorithms. The JSEG algorithm consists of two stages; color quantization and spatial segmentation. However, the major problem of JSEG algorithm is over-segmentation. The unsupervised image segmentation method groups local pixels that are homogeneous in low-level features into non-overlapped larger regions that may potentially correspond to objects or their parts without any training examples. The over-segmentation occurs when many segments map to a single object. This thesis proposed a modified JSEG algorithm to solve the problem of over segmentation when applying it to underwater coral reef images.