The Tech TL;DR:

• Algorithmic Provenance: A deep-learning model trained on authenticated Raphael works identified the face of St. Joseph as a stylistic outlier, suggesting the involvement of pupils such as Giulio Romano.
• Architectural Shift: The research team pivoted from traditional full-frame image analysis to localized feature extraction, leveraging a Support Vector Machine (SVM) to classify individual figures within the canvas.

Architectural Breakdown of the Attribution Algorithm

The research, published in Heritage Science (2023), bypassed the limitations of standard neural networks by focusing on “deep feature analysis.” While traditional machine learning models often struggle with the small datasets inherent in historical art—where the “ground truth” is limited to a finite number of verified works—the team utilized a pre-trained ResNet50 architecture. This model, originally designed for large-scale image recognition, was fine-tuned to detect the specific, high-frequency patterns characteristic of Raphael’s brushwork.

According to mathematician and computer scientist Hassan Ugail, the computer identifies nuances invisible to the human eye. By isolating specific faces within the Madonna della Rosa, the algorithm achieved a 98 percent accuracy rate in identifying Raphael paintings. While the Virgin, the Child, and St. John were consistent with Raphael’s documented style, St. Joseph’s face failed to match the statistical profile of the master’s known output.

Implementation: Localized Feature Extraction

# Conceptual implementation for localized feature extraction
import torch
from torchvision import models

# Load pre-trained ResNet50 architecture
model = models.resnet50(pretrained=True)
model.eval()

def classify_roi(image_segment):
    # Normalize and transform segment for input
    input_tensor = transform(image_segment)
    with torch.no_grad():
        output = model(input_tensor)
    return output

# Apply to specific faces within the composition
faces = ["madonna", "child", "st_john", "st_joseph"]
for face in faces:
    result = classify_roi(get_roi(face))
    print(f"Attribution analysis for {face}: {result}")

Historical Context and Technical Implications

The Madonna della Rosa has been a subject of scholarly debate since the 17th century. While it was long considered a Raphael original, the lack of a clear provenance prior to 1667 in Spain left the door open for workshop attribution. The AI findings align with long-held suspicions from art historians who noted that the palette and execution of St. Joseph lacked the fluidity seen in the other figures.

The Trajectory of Computational Art History