Modified fuzzy prediction enabled automatic evolutionary clustering algorithm for dynamic image segmentation

Temporal images and video segmentation is a difficult task due to noise, occlusions, motion blur and abrupt scene changes. Deep neural networks have shown promise, but their success depends heavily on large amounts of labelled data or pre-training. In contrast, unsupervised clustering avoids this dependence but often suffers from limited robustness. The alternate approach of a meta-heuristic-based clustering algorithm in a dynamic environment is an under-explored area. To address these challenges, a combination of state-of-the-art clustering algorithm, Δ-MOCK (Multi-Objective Clustering with automatic K-determination) and dynamic evolutionary optimization for unsupervised segmentation of temporal images from a scene is proposed. The proposed algorithm Prediction-based Multi-objective Automatic Clustering (PMoAC) begins with superpixel generation and then applies multi-objective optimization to group superpixels based on variance and spatial connectedness. A central feature of PMoAC is a modified Fuzzy Prediction (m-FP) strategy, which extends conventional fuzzy inference by introducing weighted history and adaptive membership weighting to enhance the temporal adaptation. This allows the algorithm to anticipate evolving segmentation and adjust to environmental changes. Comparative experiments on 12 benchmark datasets with different linear and non-linear prediction schemes in the same approach, along with state-of-the-art algorithms: DNSGA-III, DMOEA/D and FDSPEA2-BR show that PMoAC with m-FP achieves consistently higher accuracy and efficiency as illustrated by 31.32% increase in mean Dice Similarity Coefficient (mDSC), 10.11% increase in mean hypervolume (mV) with almost 45% less overall runtime.