Comparison of one-stage and two-stage strategies of machine learning model for rock art object detection.

Abstract

Tracing rock art documentation has traditionally relied on manual identification by field specialists, which can be laborious and expensive. Integrating Artificial Intelligence (AI) and Machine Learning (ML) in archaeology has introduced new possibilities for enhancing rock art research. ML, a sub-field of AI, focuses on training models using data. At the same time, Deep Learning (DL) advances this approach with Artificial Neural Networks (ANN). Although ML methods have proven helpful for structured datasets, their application to rock art research remains largely unexplored. Object detection, a critical task in computer vision, involves predicting object locations and classes in images. State-of-the-art object detectors include two-stage detectors like Faster R-CNN and Mask R-CNN, which generate region proposals for classification and bounding-box regression, achieving high accuracy at the expense of speed. Conversely, one-stage detectors like YOLO treat object detection as a regression problem, sacrificing some accuracy for improved speed. Striking a balance between accuracy and speed poses a challenge in rock art research. In this context, a promising approach involves training a deep neural network to regress the difficulty scores produced by human annotators. An experiment was conducted using both one-stage and two-stage detectors on large-scale rock art images. The study presents the results and reflects their implications for rock art research. The findings emphasize the potential of AI, ML, and advanced computer vision techniques to revolutionize the field of rock art research, offering enhanced capabilities for analyzing, preserving, and interpreting these invaluable archaeological artifacts. Through automation of the detecting process and a decrease in the need for manual labor, these technologies can increase efficiency and accuracy in rock art studies.