RESTORATION OF LOW-RESOLUTION IMAGES USING MACHINE LEARNING WITH CASCADED GENERATION AND DOMAIN-INTEGRATED TRAINING ON FACIAL IMAGES

• Main Author: Limanta, Felix
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39138
• Institution: Institut Teknologi Bandung

Images are becoming more prevalent in our daily lives—especially portraits and facial images. This leads to the increasing prevalence of image-or-video-based surveillance and security, yet in many instances the low capture resolution of the camera or the distance between the camera and the subject causes the facial images to become indecipherably tiny and/or blurry. Super-resolution refers to the process of upscaling an image from a low-resolution one to a high-resolution one while adding details, while face hallucination refers to the specific usage of super-resolution on the domain of human facial images. A common technique to narrow down the solution space for general image super-resolution is cascaded generation. No currently-published face hallucination method uses cascaded generation with a magnification scale larger than 8×. This undergraduate thesis discusses the application of cascaded generation on face hallucination. This undergraduate thesis compares three face hallucination architectures: direct generation, chained generation, and cascaded generation. This is carried out to determine whether cascaded generation on face hallucination is feasible. After the experiment, the model for cascaded-generation face hallucination is trained for two weeks. Due to the lack of training time, the performance of the trained model is inferior to current state-of-the-art methods but is still viable and shall perform better with more training. In addition, using the trained face hallucination model on non-facial images results in a distorted, but still recognizable image.