YOLO-ResNet: A New Model for Rebar Detection
Yaoshun Li, Lizhi Liu · 2021
Based on the YOLOv3 model, this paper proposes a YOLO-ResNet model for rebar detection. YOLO-ResNet model is based on ResNet-18 (Residual Network 18), retains the first four downsamplings, and adds CBAM (Convolutional Block Attention Module) attention model to the input and output parts to obtain a $52\times 52$-sized feature map for rebar detection by feature fusion process. Firstly, this paper clusters the width and height of the rebar in the training set to obtain new anchor prior boxes, then replaces the Darknet-53 network of YOLOv3 with an improved backbone network, and finally introduces the FocalLoss to penalize the easily divided negative samples so as to make the model converge faster. Compared with YOLOv3, the number of convolutional layers of YOLO-ResNet is reduced by 45 layers, the total training parameters are reduced by 94.35%, the training time of the model is reduced by an average of 40% under different training times, the GPU memory usage during training is reduced by 56%, and the model weight size has been reduced by 94%. We use 4 evaluation indicators of TrainTime, BestLoss, mAP and Accuarcy for the model. The experimental results show that the YOLO-ResNet model has better performance, less training time, lower GPU memory usage, and smaller model weight files than the YOLOv3 model without reducing the detection accuracy in the rebar detection.