OPTIMIZATION OF FACIAL IMAGE RESOLUTION ON CONVOLUTIONAL NEURAL NETWORK FOR PRESENCE BIOMETRIC SYSTEMS

Facial recognition-based biometric systems are widely used in attendance and security applications because they do not require physical contact and are easy to implement on various devices. However, image resolution can affect the accuracy and computation time of facial recognition systems. This study developed a system with varying architectures and input resolutions, namely 512×512, 256×256, 128×128, 64×64, and 32×32 pixels. The dataset consisted of 480 images, including 240 face images and 240 non-face images. Evaluation was conducted using accuracy and training and testing computation times. The results showed that the 512×512 pixel resolution yielded the highest training accuracy of 55.00%, and the 512×512-pixel resolution required the longest training time of 504.23 seconds. In testing using new data, the 256×256 pixel resolution demonstrated optimal performance with an accuracy of 72.50% and a computation time of approximately 0.1602 seconds. Based on these results, the 256×256 pixel resolution can be recommended as the preferred choice for implementing a CNN based facial recognition system with limited computational resources.