Executive Summary
Identifying the authorship of fine art remains a significant challenge in computer vision due to the subjective nature of artistic style. This paper introduces Cluster Multiple Kernel Learning (CMKL), a novel algorithm designed to identify oil painters by efficiently fusing color, texture, and spatial features. By clustering candidate kernels and selecting only the most representative ones before training, the method achieves higher accuracy and significantly faster processing times than traditional Multiple Kernel Learning (MKL) approaches. On the Painting91 dataset, the method achieved a 54.5% accuracy rate, representing a substantial improvement in automated art attribution.
Introduction
While computer vision has mastered the classification of "objective" natural images—such as distinguishing between different species of animals—recognizing the "master’s touch" in fine art is far more complex. Artistic style is defined by subtle brushstrokes, specific color palettes, and unique spatial layouts that vary significantly even within a single artist's body of work.
To capture these nuances, researchers often turn to Multiple Kernel Learning (MKL). MKL allows a model to combine various data "views" (features) using different mathematical functions, known as kernels. However, traditional MKL faces two primary hurdles: redundancy and computational complexity. Many candidate kernels provide overlapping information, and processing a massive library of kernels is computationally expensive, hindering real-world applications. The CMKL algorithm aims to bridge this gap by creating a system that is both stylistically sensitive and computationally efficient.
Methodology
The proposed system utilizes a two-stage pipeline: Multi-Dimensional Feature Extraction followed by the Cluster Multiple Kernel Learning (CMKL) algorithm.
1. Multi-Dimensional Feature Extraction
To capture the comprehensive "fingerprint" of a painting, the system extracts 12 distinct types of features categorized into global and local views:
- Global Features: These describe the image as a whole. They include Local Binary Patterns (LBP) for texture analysis, GIST for spatial layout, Color Histograms in the CIE color space, and Canny Edge detection for structural lines.
- Local Features (Bag-of-Features): These focus on specific patches within the image. The system utilizes SIFT (Scale-Invariant Feature Transform) and SSIM (Structural Similarity) to capture fine-grained details like individual brushwork and local patterns.
2. The CMKL Algorithm
Rather than feeding an exhaustive set of kernels into a classifier, CMKL introduces a "pre-learning" filter to optimize the kernel set. The process involves:
- Kernel Generation: Creating a large pool of candidate kernels using various functions (e.g., Gaussian, Polynomial).
- Clustering: Using a K-medoids algorithm to group similar kernels based on their classification behavior. This behavior is measured using an Oracle matrix, which defines the ideal similarity between samples based on their labels.
- Selection: From each identified cluster, only the kernel with the highest individual classification performance is selected.
- Fusion: Traditional MKL (under $L_p$-norm constraints) is applied to this optimized, non-redundant subset to produce the final classifier.
Experimental Setup and Datasets
The researchers validated CMKL using two distinct testing environments:
- UCI Benchmarks: Four standard datasets (breast, bupa, pima, wdbc) were used to establish the algorithm's general classification robustness.
- Painting91: This primary benchmark consists of 4,266 works by 91 different artists. It is particularly challenging due to the high degree of stylistic variation across different periods of an artist's career.
Key Results and Insights
Efficiency and Accuracy
CMKL successfully broke the typical tradeoff between speed and precision. On the Painting91 dataset, CMKL-L10 achieved an accuracy of 54.5%, outperforming traditional MKL-L2 (52.9%). While 54.5% may appear modest, it is a significant achievement given the 91-class complexity and the subjective nature of the data.
Computational gains were even more pronounced. On the wdbc dataset, traditional MKL-L2 required 138.92 seconds for training, whereas CMKL-L2 completed the task in just 89.40 seconds—a 35% reduction in processing time.
The Importance of Color
One of the most striking findings was the role of color in artist recognition. Features like "Color LBP" and "Color GIST" significantly outperformed their grayscale versions. This confirms that an artist's specific use of color is one of their most distinctive and measurable stylistic markers.
Conclusion and Future Directions
The CMKL algorithm demonstrates that in kernel learning, diversity is more valuable than quantity. By removing redundant "weak learners" through clustering, the model becomes more robust and significantly faster. The authors even demonstrated the feasibility of this approach via a prototype mobile application for real-time art identification.
Future work will focus on several areas:
- Deep Learning Integration: Exploring whether Deep Neural Networks (e.g., ResNet) can automatically learn style features to replace hand-crafted ones.
- Dynamic Clustering: Developing methods to automatically determine the optimal number of clusters ($K$) based on data structure.
- Initialization Stability: Improving the stability of the clustering process to reduce sensitivity to initial starting points.
Summary updated