An accurate object detection of wood defects using an improved Faster R-CNN model
Xianghe Zou, Chongyang Wu, Hongen Liu, Zhangwei Yu, Xianyan Kuang
Abstract
Wood defect classification and location are crucial for automatic repair and reuse in wood product manufacture, which requires high accuracy and efficiency in the engineering practice. In this paper, a novel modeling method is proposed for faster R-CNN to improve the object detection of wood defect. An improved ResNet-50 network, a focal loss function, and a soft non-maximum suppression (soft NMS) were adopted to replace the counterparts of the initial Faster R-CNN. Experiments were conducted on a wood defect dataset with seven types of defects. The experimental results indicate that the using of the improved ResNet-50 network brings a significant increase in mean average precision (mAP) by 4.4% and reduces the detection time by 3.6% than that of the traditional Faster R-CNN on wood defect detection. Besides, mAPs are 3.4% and 2.8% higher than that of the traditional Faster R-CNN as the employment of the focal loss function and soft NMS, respectively. Finally, the proposed model obtains better mAP than the classic target detection networks. Results demonstrate the effectiveness and feasibility of the proposed modeling method in engineering applications.