Abstract
Traditional "one-touch" image enhancement tools often rely on computationally expensive search-based methods or static filters that fail to generalize across diverse content. This paper introduces a novel framework that treats image enhancement as a three-way interaction between low-quality images, their enhanced versions, and the corresponding enhancement parameters. By integrating image features into a Matrix Factorization (MF) model with $L_{2,1}$-norm regularization, the proposed approach enables direct, high-speed parameter prediction. Evaluations on the MIT-Adobe FiveK dataset show that the model achieves a Root Mean Square Error (RMSE) of 0.0820, significantly outperforming standard MF and $k$-Nearest Neighbors ($k$NN) baselines while maintaining the efficiency required for mobile deployment.
Introduction
Automated image enhancement is a ubiquitous feature in modern mobile photography, yet many current solutions are either too simplistic or too resource-intensive. "Lazy" learners, such as $k$-Nearest Neighbors ($k$NN), require maintaining large on-device databases or constant high-speed connectivity to find similar images and replicate their settings. Alternatively, brute-force methods that score hundreds of variations are often too slow for real-time mobile use. This research addresses these limitations by proposing a structured model that predicts optimal enhancement parameters—brightness, saturation, and contrast—directly from an image's latent features without the need for external database queries at test time.
Methodology
The proposed system leverages Matrix Factorization (MF) to model the complex relationships inherent in image editing.
Three-Way Latent Interaction
The model defines interactions between three primary factors:
- Low-quality images ($U$)
- Enhanced versions ($V$)
- Enhancement parameters ($T$)
Feature Integration and Sparsity
Unlike standard MF, this model expresses the latent image factor ($U$) as a linear function of extracted image features ($F$). To ensure only the most relevant features are used, the authors employ a convex optimization scheme with an $L_{2,1}$-norm, which promotes feature sparsity and improves the model's robustness against noise.
Generalization and Inference
A constant offset ($Q$) is introduced to improve generalization to unseen images, preventing the model from simply memorizing training data. The model is trained using Gibbs Sampling to capture non-linear dependencies. At test time, predicting parameters for a new photo is reduced to a simple, fast matrix multiplication, making it highly efficient for mobile hardware.
Experimental Results
The framework was tested on synthetic data and the MIT-Adobe FiveK dataset, which contains 5,000 expert-enhanced photos. The authors extracted 1709-D feature vectors, including color histograms and localized contrast data, to represent each image.
Quantitative Performance
The proposed model achieved an RMSE of 0.0820 on the MIT-Adobe FiveK dataset, a marked improvement over standard Matrix Factorization (0.1251) and Multivariate Linear Regression (1.2420).
User Preference and Efficiency
In subjective tests with 11 participants comparing 210 pairs of images, the proposed approach's results were consistently preferred over $k$NN-based results. Crucially, the model's complexity at test time does not scale with the size of the training set, unlike $k$NN, making it a prime candidate for on-device processing.
Conclusion
By structuring image enhancement as a latent factor interaction problem, this research provides a path toward efficient, high-quality "one-touch" editing on mobile devices. While challenges remain in handling extremely complex scenes where expert opinions may vary, the current model offers a significant leap in both accuracy and computational efficiency over existing search-based paradigms. Future work may explore non-linear mappings for feature integration or applying this structured approach to localized edits.
Summary updated