Cluster CV2: a Computer Vision Approach to Spatial Identification of Data Clusters

This work shows a novel application based on techniques of Computer Vision and Machine Learning to identify k clusters into a data set with overlapping issue. Used in area of unsupervised data clustering, where separation between groups is tricky. Through pair-to-pair distance calculations upon original data, is gotten a Distances Matrix as representative information of data. This matrix contains visual information, then using morphological operators extract relevant features for individual identification of groups in data set. Next, matrix decomposition performed to covariance matrix, being calculation of data elements for each cluster in order to project data into a new linear space. So, overlapping and separation distances among clusters are corrected without loss information. Results present correct identification of k clusters, without loss information, and eliminating data overlap. Clustering validation metrics such as Silhouette and Precision was used to test the methodology.