Automated emotion detection through facial analysis represents a crucial technological advancement for interpreting human expressions from digital images. Current classification approaches face challenges in achieving precise accuracy because they extract insufficient distinguishing characteristics needed to differentiate various emotional states. To address these issues, we introduce a Parallel Deep Convolutional Neural Network (PDCNN) incorporating Gaussian Error Linear Unit (GELU) activation for enhanced emotion classification performance. Our methodology utilizes two established benchmark collections: FER-2013 and Japanese Female Facial Expression (JAFFE) databases, which undergo preprocessing through Contrast Limited Adaptive Histogram Equalization (CLAHE) and dataset balancing via Generative Adversarial Network (GAN) based augmentation techniques. Subsequently, the enhanced dataset undergoes feature extraction using DenseNet-121 architecture, which captures representations through densely connected layers before processing through our dual-stage classification framework. This system leverages both localized and comprehensive facial characteristics, integrating them via a fusion mechanism for precise emotional categorization. Our experimental validation demonstrates that the PDCNN with GELU approach achieves superior performance with 94.67% accuracy on FER-2013 and 99.10% on JAFFE datasets, outperforming traditional CNN, Convolutional Relational Network (CRN), ResNet-50, Wavelet DCNN, and Three Channel-CNN methodologies.