The paper proposes a probabilistic formulation for semantic image annotation and retrieval, addressing the limitations of both supervised and unsupervised learning approaches. The proposed method, called Supervised Multiclass Labeling (SML), treats each semantic concept as a class in a multiclass classification problem. This approach retains the optimality of supervised learning while avoiding the computational burden of estimating multiple non-class distributions. SML represents images as bags of localized feature vectors and uses Gaussian mixture models (GMMs) to estimate class densities. The method is efficient and accurate, achieving higher accuracy than existing methods at a fraction of their computational cost. It is also robust to parameter tuning and performs well on large-scale databases. The paper includes a detailed description of the algorithms, experimental protocols, and results comparing SML with other methods.The paper proposes a probabilistic formulation for semantic image annotation and retrieval, addressing the limitations of both supervised and unsupervised learning approaches. The proposed method, called Supervised Multiclass Labeling (SML), treats each semantic concept as a class in a multiclass classification problem. This approach retains the optimality of supervised learning while avoiding the computational burden of estimating multiple non-class distributions. SML represents images as bags of localized feature vectors and uses Gaussian mixture models (GMMs) to estimate class densities. The method is efficient and accurate, achieving higher accuracy than existing methods at a fraction of their computational cost. It is also robust to parameter tuning and performs well on large-scale databases. The paper includes a detailed description of the algorithms, experimental protocols, and results comparing SML with other methods.