Image clustering is an unsupervised learning task that primary function is to categorise and extract the image's valuable attributes. Many techniques like data augmentation and normalization or meta-heuristic algorithms like swarm intelligence can be used to improve the clustering performance.
This research proposes an hybrid Ant Colony Optimization (ACO) algorithm, named CLAntIMG, based on the KMeans method to improve image clustering performances grouping structured and unstructured data components. The algorithm takes a distance matrix and an initial pheromone matrix as input and returns the best route found by the KMeans algorithm, representing the centroids of the clusters. This method computes pairwise distance matrices using different metric distances (Euclidean, Manhattan, Chebyshev). In addition, a proposed matrix containing dimensional characteristics with extracted features based on Local Binary Pattern (LBP) is applied to extract features from the images, and the resulting dimensional characteristics. These LBP features provide texture information that can enhance the clustering process by considering both spatial and texture-based dissimilarities between data points.
The efficiency and robustness of the proposed CLAntIMG algorithm compared with other algorithms demonstrate its ability to find optimal cluster assignments even with different initial parameters. The performance of each cluster is evaluated using the Silhouette score, the Calinski-Harabasz index and the normalized mutual information (NMI) leading to promising results, notably with the Chebyshev distance, reaching an NMI of 0.94 on the Iris dataset.