Raquel Dosil Lago

Research Interests


Past Research

3D Composite-Feature Detectors

Methodology for data-driven synthesis of composite feature-detectors using a decomposition-integration strategy. The image is decomposed into a set of elementary frequency features. To obtain composite features, simple features are integrated by means of unsupervised clustering. The criterion for binding is based on the concept of Phase Congruence and is reflected using image dissimilarity measures. It has been applied in different tasks of processing and analysis of 3D and 2D+t data:

  • Active model initialization in 3D medical images and feature enhancement in GPR and TMR images.
    Slides of my PhD thesis presentation (in Spanish, 32MB ppt).
  • Motion segmentation.
    Slides of ACIVS'06 presentation (20MB ppt).




Feature Preserving Diffusion Filtering

Anisotropic filter with a diffusivity tensor that depends on image features, specifically surfaces and corners. Applications

  • Enhancement of volumetric medical data, like MR, CT, etc. We achieve denoising without blurring of surfaces and rounding of corners.
    Slides of the VIIP'01 conference presentation (in English, 3,5MB ppt)
  • Regularization of distance maps to implicit surfaces. Natural neighbour interpolation produces many artifacts in the shape of the zero-level set when estimation of normals is poor. Tangential diffusion reduces artifacts but avoiding dilation or erosion of level sets due to the normal component of diffusion.
    Slides of the CEIG'04 conference presentation (in Spanish, 1MB ppt)

Superquadric Model Matching

Object identification and rough approximation to its surface by matching with an a priori model of a superquadric with global deformations. A priori models are generated by fitting to sample points of a training surface using a GA algorithm. Model matching is accomplished by optimization  of affine transformation parameters to fit image surface patches:

  • Initialization of 3D active models. Superquadrics with global deformations provide a rough approximation to the initial surface of an object, ensuring convergence of active models to the target surface en accelerating surface evolution.
    Slides of the CEIG'01 presentation (in Spanish, 1MB ppt)
  • Initialization with multipart models. CSG is used to construct multipart prior models, where each part is a superquadric with global deformations. An algorithm for the optimization of such model has been developed.
    Slides of the VIIP'02 presentation (in English, 11MB ppt)