OBJECT DETECTION AND CLASSIFICATION IN A VAGUE CONDITION USING DEEP LEARNING

• 主要作者: Wahyu Kartiko, Restu
• 格式: Final Project
• 語言: Indonesia
• 在線閱讀: https://digilib.itb.ac.id/gdl/view/75333
• 機構: Institut Teknologi Bandung

Object detection and classification is one of the sub-fields in deep learning-based image processing. Many deep learning models have proven accuracy and performance, such as faster R- CNN, YOLO, SSD. Many things affect image quality such as lighting or noise. Therefore, the image processing model must be robust so that it can detect objects under any circumstances. The presence of water spot noise and the blur caused by rainwater drop in the image makes the image unclear. Lack of light also causes the lines between the object and the background become blurry. Chia-Chi in his research stated that his model based on Faster R-CNN has good accuracies accros different objects, but the accuracy decreases when the picture has noise and blur caused by rain drop. In this study, an adaptation of YOLOv4 aims to create a model that is more accurate in detecting objects in a vague condition (rain and lack of light). The YOLOv4 pertrained model undergo fine- tuning process from a specialized dataset which contains images in rainy and low light conditions. The detected objects are limited to only 4 objects which are cars, buses, trucks, and humans. The experimental results show that the resulting model has a mAP value of 55.49%. In addition, when compared to the original YOLOv4 model, the resulting model has better accuracy in detecting objects in rainy and low light conditions.