Research on Rebar Counting Based on YOLO-Rebar Model

Abstract

Smart site is a general trend in the development of the construction industry, aiming to design and manage the site in a controllable, visualized and data-oriented way, which is of great significance to strengthen the safety and civilized construction management of the site. An improved YOLO-Rebar model is proposed by this paper for counting rebar based on the fact that there are not ground true boxes located at YOLOv3 13*13-sized channel for detecting large objects in images. The 26*26-sized and 52*52-sized channels for predicting medium and small objects of images are reserved in YOLO-Rebar, so that convolution layers are reduced by 17. Trainable parameters is reduced by 41,657,874. Moreover, the average training time of different epochs decrease by 30%, GPU memory consumption decreases by 43% and the size of model weight files cut down by 68%. The experiments show that, when epoch > 30, YOLO-Rebar achieved the same Rebar counting performance of YOLOv3 with less training time, lower GPU memory occupation and smaller model weight file size.