Enhancing Image Aesthetic Quality using Deep Learning Based on Image Color Aesthetics Assessment

The proliferation of digital photography has driven the demand for automated tools to enhance the aesthetic quality of images. While deep learning frameworks, such as the Neural Image Correction & Enhancement Routine (NICER), offer promising and interpretable solutions by optimizing explicit editing parameters, they can produce unrealistic results with extreme color transformations when guided solely by general-purpose aesthetic models. This paper identifies the root cause of this failure mode and proposes a targeted solution: regularizing the enhancement process by integrating a specialized model for color aesthetic evaluation. This model, trained on human perceptions of color harmony, acts as a perceptual regularizer to guide the optimization process. We propose and evaluate two integration strategies: a full replacement of the general aesthetic model and a hybrid approach that leverages both models in parallel. A subjective user study (N=23) reveals that our color-guided approach, particularly the full replacement strategy, yields enhancements rated as significantly more realistic and aesthetically pleasing than the baseline. Statistical analysis confirms the significance of these findings. This work demonstrates that incorporating domain-specific perceptual models for aspects like color aesthetics is a crucial step toward developing automated image enhancement systems that better align with human visual preferences.