Analytical Image Models and Their Applications

Anuj Srivastava, Xiuwen Liu, Ulf GrenanderECCV 2002 · DOI

TL;DR

This paper introduces Bessel K forms, a novel analytical probability model for image spectral components, offering a more interpretable and efficient alternative to purely numerical dimension reduction techniques. By modeling image features with these parametric forms, the authors demonstrate improved performance in tasks like texture synthesis, clutter classification, and object recognition.

What Problem Does This Paper Address?

Computer vision and image understanding heavily rely on statistical models to interpret images. A major challenge is the high dimensionality of image data, necessitating dimension reduction. Traditional methods often fall into two categories: purely numerical, bottom-up approaches (like PCA) that are computationally efficient but lack physical interpretability, and top-down, object-representation-based methods (like deformable templates) that are more interpretable but computationally expensive. This paper aims to bridge this gap by developing analytical probability models for image spectral components that are both computationally efficient and offer a degree of physical interpretability, enabling better image understanding.

Methods

The core contribution is the introduction and application of Bessel K forms. The process begins by decomposing an image into its spectral components using a bank of filters (e.g., Gabor, Laplacian of Gaussian). For each filtered image, the marginal probability distribution of the filter outputs is modeled using a two-parameter analytical form, the Bessel K form. This form is derived from a generator model and is characterized by parameters p and c.

Key methodological aspects include:

  • Analytical Modeling: Instead of relying solely on numerical histograms or statistical moments, Bessel K forms provide a closed-form analytical expression for the probability density.
  • Parameter Estimation: The Bessel parameters (p, c) are estimated efficiently from the filtered images using moment estimators (kurtosis and variance).
  • Pseudo-Metric for Image Comparison: A novel L²-metric is derived for comparing Bessel K forms. This metric is then extended to a pseudo-metric for comparing entire images by considering the distances between their spectral representations across multiple filters. This allows for quantifying image similarity in a theoretically grounded way.
  • Applications: The framework is applied to several computer vision tasks, including texture modeling and synthesis, clutter classification, hypothesis pruning for object recognition, and object classification.

Datasets

The paper utilizes several publicly available datasets for evaluation:

  • Groningen image database: Used for natural image texture modeling and synthesis.
  • FSU IR face database: Used for evaluating IR face image modeling.
  • Brown range database: Used for range image analysis.
  • Columbia Object Image Library (COIL-100): A standard dataset for 3D object recognition, used for object classification experiments.
  • Equinox company's infrared face dataset: A large dataset of long-wave infrared face images used for object classification.

No new benchmarks were introduced; rather, existing datasets were used to demonstrate the effectiveness of the proposed Bessel K representation.

Key Insights and Ideas

  • Bridging Numerical and Analytical Approaches: Bessel K forms offer a way to combine the efficiency of spectral analysis with the interpretability of analytical models, moving beyond purely numerical representations.
  • Physical Interpretability of Parameters: The Bessel parameter p is shown to be related to the "distinctness" and "frequency of occurrence" of image features (like edges), providing a link between model parameters and object characteristics.
  • Efficient Image Comparison: The derived L²-metric on Bessel K forms provides a principled way to quantify image similarity, which is crucial for tasks like classification and recognition.
  • Broad Applicability: The framework demonstrates effectiveness across diverse image types (video, IR, range) and tasks, suggesting its general utility in image understanding.

Key Results

  • Goodness of Fit: Estimated Bessel K forms closely match observed histograms for various image types, as evidenced by visual inspection and quantitative measures like KL divergence. For natural images, the average KL divergence was limited to 0.0657, and for IR face images, it was 0.0478.
  • Texture Synthesis: The model successfully captures perceptual characteristics of textures, enabling systematic synthesis of images with matched histograms.
  • Object Classification Performance: Bessel K representation, when used with a nearest neighbor classifier, achieved state-of-the-art or comparable results on the COIL-100 dataset, often outperforming PCA, ICA, and SVMs, especially with fewer training views. For instance, with 36/36 training/test images, Bessel Forms achieved 99.89% accuracy, surpassing other methods.
  • Infrared Face Recognition: Near-perfect recognition rates were achieved on a large infrared face dataset using the Bessel K representation.

Limitations and Future Work

  • L²-Metric Constraint: The L²-metric on Bessel K forms is only applicable for p > 0.25. The paper suggests workarounds like approximating p or using numerical integration for cases where p < 0.25.
  • Filter Dependency: The image pseudo-metric d₁ is dependent on the choice of filters used. The authors acknowledge that this Euclidean form might not always be appropriate and suggest exploring alternatives like the maximum of components.
  • Nearest Neighbor Classifier: While effective, the nearest neighbor classifier used in some applications might not generalize optimally, especially with limited training data. The authors suggest that more sophisticated classifiers could yield significant improvements.
  • Challenging Recognition Scenarios: The paper notes that the performance of Bessel representations in "challenging object recognition situations" remains to be fully tested, implying this as a future research direction.
  • Further Applications: The authors suggest that the ability of Bessel K forms to prune hypotheses is promising for more detailed models in object recognition.

Summary updated

BibTeX
@inproceedings{srivastava2002analytical,
  author = {Anuj Srivastava and Xiuwen Liu and Ulf Grenander},
  title = {{Analytical Image Models and Their Applications}},
  booktitle = {European Conference on Computer Vision},
  year = {2002},
  doi = {10.1007/3-540-47969-4\_3},
}
Figure