1 January 2026 in Applied Computational Intelligence and Soft Computing
While significant progress has been made in face recognition under individual environmental constraints, comprehensive approaches that address multiple simultaneous challenges—such as occlusion and pose variation—remain relatively scarce in the literature. Most existing recognition models are tailored to mitigate specific conditions in isolation, thereby limiting their generalizability in real‐world, unconstrained environments. Concurrently, transfer learning has emerged as a dominant paradigm, with pretrained convolutional neural networks (CNNs) being fine‐tuned for downstream face recognition tasks due to their efficiency and superior performance. Despite the inherent robustness of deep learning models to certain distortions, their effectiveness can be substantially improved through appropriate preprocessing strategies, especially under compound constraints. In this study, we introduce a dual‐stage frequency‐domain preprocessing framework based on the discrete cosine transform (DCT) and fast Fourier transform (FFT), designed to enhance the performance of a hybrid SqueezeNet–PCA–SVM architecture for face recognition under dual constraints—synthetic occlusion and head tilt. To simulate real‐world variability, test images were artificially occluded at 20% and 30% levels and then reconstructed using a statistical imputation technique known as multiple imputation by chained equations (MICE). Experimental results show that integrating the DCT–FFT preprocessing pipeline with the SqueezeNet–PCA–SVM model significantly boosts recognition accuracy, from 70% to 76.25% at 20% occlusion recovery and from 32.5% to 63.75% at 30%. These findings affirm the efficacy of DCT–FFT as a noise‐reduction mechanism that enhances the discriminative quality of imputed images and underscore its value as a preprocessing enhancement for deep learning–based face recognition systems operating in multiconstrained environments.