Robust and accurate automatic detection of anatomical features on organic shapes is a challenging task. Despite a rough similarity, each shape is unique. To cope with this variety, we propose a novel clustering-based feature detection scheme. The scheme can be used as a standalone feature detection scheme or it can provide meaningful starting points for surface analyzing-based detection algorithms. The scheme includes the identi cation of a representative set of shapes and the usage of a specialized iterative closest point algorithm for the registration of shapes, which is followed by the projection of the features using the transformation matrix of the registration. Evaluation is based on a large set of expert annotated shapes and showed superior performance compared to state-of-the-art surface analyzing methods. Accuracy increased of 32 % and detection of all features is ensured.
Shape analysis typically involves the abstraction of complex structures by
removing redundant details. It captures the essence of geometry via a representative
set of distinctive features, such as 3-D points, planes, curves or areas. These
features can be employed for classification, registration and for driving the
automation of shape transformations. The problem is very challenging due to the
variability of organic surfaces. In this work, we focus on the robust and
accurate detection of anatomical features found on ear impressions. An example ear
impression with annotated features is given in Fig. 1. Theses features provide
the basic structure for the automatic design of customized in-the-ear hearing
aids [
Previous work in this field was reported by Paulsen et al., they employed
anatomical features (landmarks) to build a statistical shape model of the human
ear canal for shape analysis and the automatic design of customized hearing
aids [
Our CBD scheme is based on three methods: (i) robust alignment of ear impressions, (ii) identification of a representative set of ear impressions, and (iii) feature projection from one impression to another. 2.1
Similar to Zouhar et al. [
The initial centerline is computed by equidistantly slicing the mesh parallel to a plane defined by the open contour at the bottom of the impression (Fig. 1). (a) Frontal view (b) Backward view Fig. 1. 3-D mesh representation of an ear impression (ear negative) consisting of the external ear and the ear canal. Some of the later considered features are annotated. For each slice the center of mass is computed resulting in an ordered set of points l1; : : : ; lN . Afterwards the centerline is refined and improved by applying internal and external forces as defined below
Eext;i =
In eqs. (1) to (3), xr;i denotes the intersection of a random ray r with the mesh
emitted from the centerline point li, and N is the number of rays. The final
update rule of a centerline point li is a weighted combination of the internal
and external force, where = 0:04 and = 1:0. The centerline points are
updated according to eq. (3) until convergence. An example for the initial and
final centerline is given in Fig. 2. A critical step of the ICP is the matching
of point pairs, which can be computational expensive. We can use the fact
that the centerlines are ordered from top to bottom to employ an efficient point
matching technique. We shift the centerlines along each other. In each shift
step the overlapping parts are extracted. The point matching is then reduced
to matching the index i. For every step, we compute the ICP (point-to-point
error metric) and store the transformation matrix along with the registration
error. To solve the ICP, we utilize the SVD-based strategy proposed by Arun
et al. [
To detect the features on a so far unknown ear impression Snew it is registered with each surface in SRep. In the following only surface Sref 2 SRep with the smallest registration error is considered. The features Fref of Sref are transformed using Tref resulting in a new feature set Fr′ef. Since Fr′ef typically will not end up directly on the surface Snew a final projection is necessary. Currently, we employ only simple projection algorithms. In case of feature points or point sets, the projection is achieved by projecting the given point on the closest point of Snew. In case of feature planes, no projection is necessary. 3
The clustering of a set of 473 ear impressions resulted in a representative set of 10 ear impressions. Major distinctive properties of the clusters are thickness of the ear canal, direction of the ear canal, thickness and length of the helix and the shape of the anti-tragus concavity (Fig. 1).
For the evaluation of the feature detection we restricted ourselves to the
feature points listed in Tab. 1. We compared our CBD scheme with a
surfaceanalysis-based (SBD) one developed by Baloch et al. [
A novel clustering-based feature detection scheme specifically adapted to ear impressions was presented. The scheme involves the identification of a representative set of ear impressions and the development of a robust alignment procedure for ear impressions. The feature detection is carried out by registering a denotes the mean distance in mm and new impression with the impressions of the representative set. Followed by a projection of the features taken from the impression with the best registration result. We could show that our scheme is superior compared to an approach that analyzes the surface for peaks, concavities, ridges and bumps. On average an improvement of 1mm could be achieved. Due to the template based approach, the detection rate is 100 %. The tolerance rate, defined as detecting a feature in a certain area around the labeled feature, could be improved about 9 %. Furthermore, the standard deviation of the detection error is greatly reduced, which corresponds to a more robust detection of the features.
So far, the presented results are preliminary and restricted to feature points. The next steps in our work are extending the implementation to feature planes and feature areas and better feature projection algorithms.