Enhancing Generalized Autoencoder for Improved Supervised Dimensionality Reduction in Image Recognition

Objectives: The article proposes enhancements to the Generalized Autoencoder (GAE) framework for Dimensionality Reduction (DR).The first objective is to improve the reconstruction process so that recognition accuracy increases. The second objective is to embed class label information into the framework. This turns GAE into a supervised method with better interclass discrimination. Method: Two improvements to the GAE are proposed: (i) an adjustment to reconstruction weights to capture neighbourhood relations in high-dimensional spaces effectively, and (ii) a strategy that incorporates class label information directly into the reconstruction set. The approach is evaluated using standard benchmark datasets for face recognition and handwritten digit classification. The approach is evaluated on eight benchmark datasets: four standard face databases (ORL, AR, UMIST, YaleB) and four handwritten digit databases (MNIST, Gujarati, Devnagari, Bangla). For a fair comparison, recognition performance is studied under varying dimensionality of the reduced feature space, different neighborhood sizes. Findings: Experimental results show that the modified GAE significantly improves recognition performance compared to the original GAE by 1-5%. The supervised version demonstrates stronger discriminative power across all datasets with better recognition rates by 3-7%. Novelty: The proposed approach adds class label information into the reconstruction set, turning GAE into a supervised framework. At the same time, it refines the weight adjustment process. Together, these two changes improve both accuracy and discrimination in dimensionality reduction tasks. Keywords: Dimensionality Reduction, Auto Encoder, Laplacian EigenMaps, Locally Linear Embeddings, Locality Preserving Projection, Neighbourhood Preserving Projection