Lightweight face recognition based on PCA coupled with PSO-based feature optimization in resource-constrained environments

Abstract

We present a streamlined PCA-PSO structure of lightweight face recognition in this paper, which combines principal component analysis (PCA) to reduce dimensions with particle swarm optimization (PSO) to adaptively select features. In contrast to traditional PCA-based models, our model dynamically chooses the most informative subset of the 44 features and results in a small but informative representation. The approach has a high recognition accuracy of 99.99 on ORL and 98.45 on LFW, with low latency (approximately 3 seconds per epoch) and low memory footprint, and without the need to use a graphic card to run it. Extensive computational studies have shown that the proposed pipeline is much faster and consumes less memory than PCA-ACO and lightweight convolutional neural network (CNN) models like MobileNetV2 and Squeeze Net and is best suited to run in real-time in CPU-limited environments. The statistical tests prove the betterment of our method compared to the past methods, where the changes are statistically significant. This research offers an efficient, simple, yet scalable alternative to deep learning-based recognition systems, especially when embedded integration is a key requirement, by offering rapid convergence, adaptive feature selection, and compact representation.