PGNet v2: Enhancing Aesthetic Image Critique Generation with Self-Resurrecting Activation and Gaussian Gated Units

With the widespread use of smartphones and digital cameras, automated aesthetic critique generation has become essential for helping users capture visually appealing photos. Traditional methods rely on rule-based techniques or handcrafted aesthetic features, which lack flexibility and struggle to provide context-aware critiques. In contrast, deep learning enables data-driven analysis, allowing models to learn complex aesthetic patterns and generate insightful feedback. However, existing deep learning approaches suffer from limited critique diversity, poor multimodal fusion, and difficulty capturing fine-grained aesthetic attributes, reducing their effectiveness. To address these challenges, we propose PGNet v2, a transformer-based aesthetic critique generation system. PGNet v2 employs SwinV2 as a visual encoder and GPT-2 as a decoder, integrating visual and textual features through a cross-attention mechanism. By the use of FFN_GEGLU, PGNet v2 enhances text modeling and expressiveness, improving critique diversity, while Self-Resurrecting Activation Unit (SRAU) strengthens multimodal fusion, ensuring more coherent and contextually relevant critiques. Additionally, SwinV2's hierarchical feature extraction enables the model to capture fine-grained aesthetic attributes, improving the depth and accuracy of generated feedback. Experimental results demonstrate that PGNet v2 outperforms PGNet v0, AMAN, and CNN-LSTM across 35 evaluation metrics, achieving a 94% improvement rate. The model excels in distinguishing high- and low-quality images, adapting to diverse lighting conditions, and generating more insightful and contextually relevant critiques. These findings confirm PGNet v2's superiority over existing methods, making it a valuable tool for real-world photography enhancement.