Shedding New Light on Traditional Image Clustering: A Non-Deep Approach With Competitive Performance and Interpretability
Jie Chi Yang, Sheng-Ku Lin · 2024
Image clustering, a fundamental task in computer vision, entails grouping images into distinct categories based on their intrinsic properties and similarities. Traditional (non-deep) image clustering models often struggle to achieve high accuracy due to variations in pose, illumination, or occlusion within image datasets, which frequently lead to multi-modal clusters. In recent years, deep neural networks, with their robust representation learning capabilities, have demonstrated considerable accuracy in image clustering tasks. However, the high computational costs and lack of interpretability of deep models have limited their practical application. In this paper, we introduce the MaxFeature Torque Clustering (MFTC) model, a non-deep approach designed as a transitional solution that bridges the gap between traditional and deep image clustering models. MFTC stands out for its accuracy, outperforming conventional image clustering methods, and provides greater interpretability, an aspect often lacking in deep models. Across six publicly available image datasets, the non-deep MFTC model achieved accuracy comparable to or better than previous state-of-the-art (SOTA) deep image clustering models. The codes are available1.