LOW RESOLUTION ATTRIBUTES CLASSIFICATION

• Main Author: Chuardi, Vincent
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/64114
• Institution: Institut Teknologi Bandung

Attributes Classification is a classification processes to classify attributes/things that is used by people, i.e. clothing (shirt, suit, jacket, jeans, short, dress, etc.) and accessories (hat, helm, etc.). In addition, attributes classification also describe attributes color. Important procedures on this paper is Super Resolution, Object Detection and Classification, and Color Classification. Chosen Super Resolution is FSRCNN type small. The scaling used is 4. Super Resolution model is pre-train model without changes. Execution time by FSRCNN-small is 5 milliseconds. Chosen Object Detection & Classification is YOLO version 4 standard. Model used by YOLOv4 is re-trained using custom dataset with appropriate clothing and accessory attribute classes. YOLOv4 with re-trained model has a 46% mAP. YOLOv4 has the highest mAP compare to other object detection methods with the same custom dataset. Color classification make use HSL color model to classify colors. In addition, this step also do segmentation to separate background with object attributes to get only the relevant attributes color. Canny and Hough Transform is used as the method to do segmentation. Execution time for the whole color classification is 15 milliseconds every 1000 pixel. System success with 83.8% precision score and 86.7% color description.