Improving Biometric Accuracy Using Ear Canal Images: Removing Overexposed and Underexposed Images Through a Lightweight Algorithm

Extensive research has been conducted on high-security personal authentication using DNNs, such as facial, fingerprint, and voice recognition. However, these biometric methods have accuracy issues under certain conditions, such as in dark environments like night-time, when gloves or masks are worn or in the presence of background noise. To address these challenges, our study focuses on the use of images of the ear canal which is a highly individual characteristic that can be easily captured and is unaffected by the environment. Previous research involved manually removing noisy images, such as those overexposed or underexposed, before conducting accuracy evaluations, achieving an F1-Score of 1.0 with clean data. Additionally, systems that utilize DNNs suffer from high power consumption due to GPU usage and expensive computational resources. In our research, we aim to enhance accuracy through an automated detection of overexposed and underexposed images using a simple algorithm. We also opted for the EfficientNet model which is high computational efficiency. Our results showed an improvement in the F1-Score to 0.99968, an increase of 0.0019 compared to before noise removal. Furthermore, by conducting a five-fold cross-validation not previously performed, we demonstrated no decrease in accuracy, thus indicating the versatility of using ear canal images for personal authentication. The misclassified images we analyzed included significant blurring near walls, faintly shadowed images, and blurred images. Future work will explore methods to resolve these issues and examine more lightweight models suitable for operation on edge devices.