An Object Detection Model based on Augmented Reality for Iraqi Archaeology

Abstract

The culture of Iraq, which boasts a rich history, serves as evidence of the magnificence of human civilization. Nonetheless, safeguarding and highlighting this valuable cultural legacy has become a significant worry in a time characterized by technological progress. Augmented Reality (AR) offers a powerful tool for preserving and presenting historical sites. The aim of this research is to leverage AR technology as a means to ensure the continued preservation and dynamic presentation of Iraq's cultural heritage. This study explores the capabilities of CNNs as the basis of AR's development. CNN is used as an essential initial step in constructing AR systems. The proposed model utilizes a pre-trained backbone network to extract complicated spatial features from input images; additional convolutional and fully connected layers are introduced to refine these features. A new custom class called “AnchorBoxes”, dynamically generates predefined anchor boxes for each feature map. Since there is not an appropriate Iraqi archeology dataset available for training deep learning models, a dataset of 2188 color images was collected. Spanning ancient Iraqi ruins, celebrated monuments, and real-time scenes combined with various objects. This dataset is subjected to manual annotation, wherein bounding boxes and labels are assigned to objects in each image. Results from the regression analysis emphasize the model's proficiency in estimating object bounding box coordinates with good precision and regression loss equal 0.008, facilitating locate-accurate object localization. The classification outcomes illuminate the model's ability to assign class labels to detected objects with high confidence. The mAP for the trained model was 0.84 and the classification loss was 0.02

Keywords

• Anchor boxes, Classification, Computer Vision, Localization, Object detection.