<p>Deep learning-based emotion recognition technology has rapidly advanced, emerging as a key focus in image analysis. Dunhuang murals, invaluable global art treasures, depict historical figures’ emotions, whose accurate recognition is crucial for understanding their artistic and historical significance. This study proposes a PNASNet-CBAM-based method, incorporating the CBAM attention mechanism to classify mural figures’ emotional features, while analyzing their cultural evolution. The findings include: (1) PNASNet-CBAM achieves 76.16% emotion classification accuracy, achieving a macro-average F1 score of 0.812 and recall of 0.892 on the test set, surpassing mainstream models; (2) Expert-annotated emotional categories established authoritative standardized data through systematic preprocessing; (3) First application of deep learning revealed historical evolution of emotional expressions in Dunhuang murals, advancing cultural heritage research methodologies. These results demonstrate the model’s superior performance, validate the annotated dataset’s reliability, and pioneer technological applications for decoding artistic emotional transitions across dynasties.</p>