Facial Emotion Recognition (FER) plays a crucial role in intelligent human–computer interaction, healthcare monitoring, and behavioural analysis by enabling systems to interpret human emotional states from facial expressions. However, existing deep learning-based approaches, particularly those relying on ensemble convolutional neural networks and full-scale transformer architectures, often suffer from high computational complexity, redundancy in feature representation, and limited efficiency for real-time applications. To support these challenges, this paper suggests an efficient and innovative model that is commonly known as Adaptive Feature Selection with Lightweight Transformer (AFST) in the classification of facial emotions. The proposed solution initially uses a simple convolutional neural network that offers important spatial attributes of facial images, which makes for lower computing costs. Subsequently, an adaptive feature selection module is introduced to dynamically evaluate and retain only the most informative facial features, effectively eliminating redundant and irrelevant data. Such selective representation is then transformed using a lightweight transformer architecture, and the fine-tuned features are captured by global contextual relationships using an efficient attention mechanism. By limiting the attention computation to selected features, the model significantly reduces complexity while maintaining high discriminative capability. The proposed AFST model is evaluated on benchmark datasets including the CK+ and FER2013 benchmarks and AffectNet, and it has shown better performance in terms of accuracy, precision, recall and F1-score. The framework also produces significant decreases in computation costs and inference time and is thus applicable in real-time implementation. The results validate that the integration of adaptive feature selection with lightweight transformer learning provides a scalable, robust, and efficient solution for next-generation facial emotion recognition systems.