The shape of the human ear is one personal characteristic that can be used in a biometric authentication system, using two-dimensional (2D) images obtained from the ear. Although biometric authentication based on physical properties, like fingerprints, eyes, and faces, or even biometric methods based on behavioral characteristics (for example online signature recognition, typewriting style), is more common, more studied, and also (in several cases) more effective than using the ears’ shape as a biometric characteristic, this paper presents an interesting approach based on local image feature extraction, transformation, and selection. The proposed method was applied to classify ear images, obtaining very good results, both in performance and robustness, when compared to other similar approaches (80 percent correct classification). It seems that this approach can also be useful when applied to other similar problems of pattern recognition using 2D images.

Shape analysis and authentication of the ear based on images is a hard task; several problems can significantly affect the classification/authentication task, for example: pose variations, changes in illumination and lighting conditions, occurrence of shadows and occlusions, and ear registration that is not very precise (problems in the selection of the region of interest). Unfortunately, as the authors indicate, 2D image-based ear biometrics usually obtain results that are dramatically poorer than those obtained by matching the three-dimensional (3D) data. Three-dimensional ear biometrics can be more robust than methods that use only the 2D image (projection) obtained from the original 3D structure.

To overcome these problems, the authors propose a new method based on local information. This avoids some problems related to ear registration, pose, and occlusions. They argue that their proposed method can significantly improve the performance of an image-based ear biometric system, and their experimental results, compared to other state-of-the-art methods, prove the performance, accuracy, and robustness of their new approach, based on local information.

The proposed method is a six-step process. First, image preprocessing is used to segment the images (based on ear landmarks), resize the segmented image to 150-by-100 pixels, execute a contrast-limited adaptive histogram equalization, and normalize the image.

Second, sub-window (SW) selection is used to divide the original image (150-by-100) using a sliding window, thus obtaining fifty 50-by-50 pixel SWs.

Third is feature extraction, a convolution of the SW (rescaled to 12-by-12) with a bank of Gabor filters (16 filters), capturing some spatial properties of the images (resulting in a vector with dimension 2304=12-by-12-by-16).

Fourth is feature transformation, the application of Laplacian eigenmaps (LEM) to project the input feature vector onto a 100-dimensional space. This subspace preserves local structure, and seems to have more discriminating power than the principal component analysis (PCA) approach.

Fifth is feature selection: the most discriminative SWs are selected by running sequential forward floating selection (SFFS). This method adds the SW classifiers one by one, selecting the best individual SW that contributes the most to outperform the previous set of classifiers. This algorithm composes a set with the best k classifiers (the minimum number k of classifiers can be defined by the user).

Lastly, step six is classification and fusion: the features extracted and transformed from each selected SW are used to train a nearest neighbor (NN) classifier. The implementation was done using MATLAB, and was based on well-known functions and some additional available code (as referenced in the text).

The experimental results demonstrate that the features extracted, transformed, and selected are robust, and support good performance when addressing these kinds of problem. The experiments demonstrate the validity and superiority of the local approach compared to global methods, as well as the robustness of the selected features when applied to different problems related to ear shape analysis and authentication (for example, pose, illumination, and occlusion). Finally, the proposed method seems to outperform several other ear authentication methods based on 2D images (with 80 percent correct classification), but the state of the art of 3D ear authentication methods can achieve a significantly better performance--roughly 97 percent correct classification.