Face Identification System Using Convolutional Neural Network for Low Resolution Image

This research aims to determine the performance of face identification on closed circuit television (CCTV) cameras. There are two data classifications used, namely training data and testing data. The training data use the CASIA-Webface dataset. Meanwhile, the testing data consists of two data, namely the source data and the target data. The source data are form of photos taken using a Digital Single Lens Reflex (DSLR) camera, while the target data use video data taken with CCTV. The source data consists of 10 IDs, each of them has 1 image for each size, so the total images used in the source data are 50 IDs. While the target data are 20 IDs, each of them has 20 face images with low resolution characteristics, less light and face capture not parallel to CCTV. This research uses Convolutional Neural Network (CNN) method with ResNet50 architecture, ArcFace as a loss function in the training process and Cosine Similarity for the face identification process. ResNet50 and ArcFace use an embedding size of 512 and in the training process, ArcFace's scale and margin parameters are 64 and 0.5. The results indicate differences in accuracy, True Positive Rate (TPR) and False Positive Rate (FPR) from the face identification process between the image sizes used and the respective IDs in the source data. The method used had the highest performance for face image identification scenarios of 128 x 128 pixels with accuracy and FPR reached 99.30% and 0.02%. From the TPR, the method used had high performance at size of 512 x 512 pixels namely 91.50%.