Toward accurate detection of small objects in rail construction: A deep learning perspective

Object detection is a fundamental task in computer vision that involves locating and recognizing all instances of objects within images or videos. Recent advances in deep learning have significantly enhanced object detection accuracy, particularly in complex environments such as rail construction. One of the pressing challenges in this domain is the accurate detection of small objects (e.g., clips, bolts, and nuts), which is critical for ensuring asset integrity, safety and the effective development of asset registers for maintenance. This paper addresses the following research question through a systematic review and meta-analysis of existing literature: Which deep learning algorithms are most effective in accurately detecting small objects (assets) during rail construction to support maintenance and management throughout a line's operational life? We examine the deep learning algorithms applied to object detection in rail infrastructure to answer this question and the key parameters and metrics used to benchmark their performance. We analyze the unique challenges of applying these architectures to rail construction environments. The conventional performance assessment of object detection models often provides an incomplete picture while relying on established metrics like accuracy, recall, precision, mean average precision, and intersection over union. Such evaluations tend to focus exclusively on final performance scores, overlooking the model's intrinsic complexity and the characteristics of the dataset it was trained on. Thus, a comprehensive analysis must extend beyond surface-level metrics to examine the model's architectural complexity and the dataset characteristics (i.e., quality, scale, and complexity). The contributions of this study are twofold as it: (1) provides new insights into the nature of deep learning architectures for small object detection in rail construction; and (2) identifies the limitations associated with applying small object detection frameworks to three-dimensional point cloud data are identified paving the way to make advances to accurately to detect small objects. • Reviews the object detection and deep learning (DL) literature about the accuracy of detecting small objects in rail environments. • Deep learning algorithms used for object detection of small objects performance is benchmarked and evaluated. • Challenges of deploying varying DL architectures to rail construction and the implications for future research are identified and discussed.