Feature proposal model on multidimensional data clustering and its application

Xi Li, Huimin Ma, Xiang Wang

Abstract: Highlights • We propose a feature proposal model for clustering based on multidimensional data. • We propose a feature non-maximum suppression method to achieve dimension reduction. • We design a grouped weighted feature clustering model for effective classifying. • Our method achieves comparable performance on several classification datasets. • Our method can be used in the application of salient object segmentation. Abstract Unsupervised learning based on clustering is an important field of artificial intelligence. In recent years, most studies focused on improving the performance of clustering algorithms by balancing the number of samples in each class or removing the noise features by iterative optimization. However, these methods are not suitable for some practical classification tasks, such as image segmentation and salient region proposals. Due to pixels or superpixels belonging to different objects in different images are not balanced, and it is difficult to use a limited number of features to describe classifying different objects, these algorithms cannot achieve high-performance classification in these tasks. In this paper, we put forward a feature proposal (FeatPro) model that can dynamically select useful features for different tasks. Furthermore, we design a feature non-maximum suppression (FNMS) algorithm, in order to achieve high performance classification by using as few features as possible. Experiments show that our method obtains competitive results on standard classification datasets and in the application of salient object segmentation.