End-to-end unsupervised learning of latent-space clustering for image segmentation via fully dense-UNet and fuzzy C-means loss

End-to-end unsupervised learning of latent-space clustering for image segmentation via fully dense-UNet and fuzzy C-means loss

연구 분야: Artificial Intelligence

학회: Machine Vision and Applications

Clustering is a fundamental unsupervised approach in machine learning for grouping tasks. Image segmentation is one of the main applications of clustering and a preliminary requirement for most high-level applications in computer vision and scene understanding. However, parameter tuning requirements of conventional unsupervised image segmentation approaches limit their application. Deep learning approaches are capable of diverse and discriminate feature learning, however supervised learning paradigm and computational complexity of deep neural networks (DNNs) induces bottlenecks for real-time applications. We present unsupervised learning paradigm for fully dense-UNet (FDU-Net) model training with loss constraints: Semantic loss, Fuzzy C-means Clustering (FCM) loss, and Total Variation (TV) loss. Semantic loss works by selecting maximum activation class for each pixel spatial location and Simple Linear Iterative Clustering (SLIC)-based spatial refinement provides a coherent feature representation for model optimisation. FCM loss is based on the objective function of the conventional unsupervised Fuzzy C-means algorithm loss function. TV loss computes and minimises the spatial discontinuities in the FDU-Net activation maps. Loss constraints operate in tandem to ensure the control of false positives and false negatives. We conduct extensive experiments to compare our proposed method with unsupervised conventional and contemporary deep learning-driven (DL) methods. We experimentally demonstrate that the proposed method yields competitive quantitative and, most importantly, qualitative segmentation results, on the unseen images from the BSDS500 benchmark dataset. During inference, the segmentation quality of the proposed approach results is more significant than the contemporary DL-based and conventional clustering methods while reducing the computation cost by several folds.