Content-based binary indexing of coral reef stitched images

Abstract

The categorization and searching of images based on content is a fundamental part of image data analysis, but may prove tedious if done manually. Automating a task to separate images with relevant features of study may prove a useful tool in the research on large image datasets. This paper proposes a method for content-based indexing of underwater coral reef stitched images acquired by the Teardrop system into feature and non-feature classes by means of grey-level co-occurrence matrix (GLCM) features and a k-nearest neighbor classifier with a k-value of 5. Results of classification on three sites showed accuracy (ACC) values ranging from 80 to 90% when compared with the ground truth, while the average of both negative predictive value (NPV) and true positive rate (TPR) values were seen at 72.01 %. A total processing time of 1.48 minutes/100 images on average was achieved. Further research may improve the classifier through different classification methods or extend classification into higher-level semantic categories beyond the original binary classification