Ear Recognition Based On Watershed Algorithm And Neural Networks

Ear recognition is a new technology of biologic recognition. Compared with other objects of biologic recognition, human ear has unique physiological and structural advantage. It can be a beneficial supplement for other biometric features, or be solely used at some situations as well.Currently, most researches are focused on characteristics of human ear. However, interferential factors induced by non-ear area result in invalidation of characteristics information. Therefore, a marked watershed algorithm was employed in this paper to select human ear image from the valid area. In this way, extraction of characteristics just deals with ear area. Consequently, validity of characteristic information will be improved.Before using marked watershed algorithm, the original ear image is preprocessed with magnitude of gravitational force field transformation. The magnitude of gravitational force field transformation is originally used in extraction of characteristics for ear image, but experiment shows that it can strengthen edges of image and restrain noises to some extent.However, from a practical effect, it is not satisfactory for the traditional watershed algorithm directly to be used in the preprocessed images. There are many excess segmented areas. Therefore, through a horizontal comparison of a variety of edge detection algorithms, marked watershed algorithm was adopted finally. Because statistics of images were used to obtain initial numerical of marks, the marked watershed algorithm led to good results, successfully separating an effective ear area from image.On the basis of effective ear area, texture and invariant moments, two kinds of algorithm of characteristics extraction are used for representation of valid ear area. Because the two algorithms are based on statistics of area, they are not sensitive to rotation. This advantage could mitigate impact of recognition results due to translation, zoom and rotation of image.In order to verify effectiveness of texture and invariant moments of ear area, BP feed forward neural networks is adopted as classifier in this paper. These two characteristics vectors were tested respectively in two sections on BP feed forward neural networks. And through analyzing testing data, recognition performances of the two characteristics were compared.From experimental results, under conditions of 200 training samples, these two characteristics maintained correct identification of a high rate at 99 and 100 percent. For noised images, the correct identification rate fell to a certain extent of 96.5 and 99 percent. This is still a high level. Especially, for rotated images, the correct identification rates of the two algorithms have no or little decline. However, for non-registered images, refusal rate failed to achieve a high correct level as the registered images 96 and 92 percent. Overall, ear recognition based on watershed algorithm and neural networks is satisfactory. And it is expected to provide valuable ideas for the development and application of human ear recognition technology.