PENGEMBANGAN MODEL DETEKSI CACAT PERMUKAAN MOBIL MENGGUNAKAN JARINGAN SARAF TIRUAN

• Main Author: Made Atmavidya Virananda, I
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/68438
• Institution: Institut Teknologi Bandung

PT X is an automotive manufacturing company with production plants located in Indonesia, whose vehicles are sold domestically and exported internationally. Within their production line, there exists a final inspection station where each vehicle unit are examined for surface defects. Currently, this process is done visually by human operators, introducing several risks in terms of detection accuracy and speed: (1) accuracy is subject to human interpretation and fatigue and (2) the time needed for inspection per unit vehicle will decrease as production scales, compelling human operators to keep up with the rapid production line. These factors motivate the development of an automated visual inspection method, specifically with Artificial Neural Network models. This study utilizes 2 object detection models based on Convolutional Neural Network (CNN), which are You-Only-Look-Once (YOLOv5) and Single-Shot- Multibox-Detector (SSD), to detect surface defects through video and static images. The models are trained to detect 3 types of defects – scratch, ding, and dent – using artificial defect samples as training data and real defects as testing data. Based on our evaluation, YOLO is better at predicting surface defects on vehicles with a Mean Average Precision score of 32.1%. This score is obtained by limiting surfaceto- camera distance at 20cm and maintaining an aslant camera position. This study finds that model performance is optimal when it is tested against a sample defect from a distance of 20cm with a sloped angle.