Image Fusion and Deep Learning Based Ear Recognition Using Thermal and Visible Images

Abstract

Advances in imaging and deep learning have fueled interest in ear biometrics, as the structure of the ear offers unique identification features. Thermal and visible ear images capture different aspects of these features. Thermal images are light-independent, and visible images excel at capturing texture details. Combining these images creates more feature-rich composite images. This study examines the fusion of thermal and visible ear images taken under varying lighting conditions to enhance automatic ear recognition. The image fusion process involved three distinct multiresolution analysis methods: discrete wavelet transform, ridgelet transform, and curvelet transform. Subsequently, a specially designed deep learning model was used for ear recognition. The results of this study reveal that employing the complex-valued curvelet transform and thermal images achieved an impressive recognition rate of 96.82%, surpassing all other methods. Conversely, visible images exhibited the lowest recognition rate of 75.00%, especially in low-light conditions. In conclusion, the fusion of multiple data sources significantly enhances ear recognition effectiveness, and the proposed model consistently achieves remarkable recognition rates even when working with a limited number of fused ear images.