Performance Improvement of MobileNetV2 Through Edge-Preserving Bilateral Filter Approach in Face Recognition

Abstract

The feature extraction process. This study proposes an image preprocessing approach using a Bilateral Filter to enhance the performance of MobileNetV2. The dataset consists of facial images from 82 respondents (comprising a total of approximately 3,444 images, with approximately 42 images captured per respondent across various poses and expressions) collected via Teachable Machine, with dimensions standardized to 224x224 pixels. A quantitative experimental method was conducted, beginning with preprocessing steps such as cropping and resizing. The dataset was partitioned into a 70:20:10 ratio for training, validation, and testing. Furthermore, data augmentation was applied, including position shifts, rotation, zoom, and shear. The model was trained for 50 epochs using the Adam optimizer. The testing results indicate that the integration of the Bilateral Filter significantly improves classification accuracy, increasing it from 69.92% without a filter to 94.31% with the filter. The model demonstrated high precision for most subjects, although variations in accuracy occurred in certain classes due to data diversity. An accuracy improvement of 24.39% confirms that the Bilateral Filter is effective in reducing noise while maintaining the facial edges crucial for Convolutional Neural Network (CNN) feature extraction. This integration provides an optimal solution for face recognition that is both accurate and efficient for implementation on resource-constrained devices.