Improved ResNet-50 model for identifying defects on wood surfaces.

Wood defect identification plays an essential role in guaranteeing the quality and efficiency of furniture manufacturing, which is a challenging task due to the data complexity and high-efficiency requirements. This paper proposes an improved ResNet-50-based modeling method to enhance the accuracy and efficiency of defect identification on wood surfaces. Firstly, a new optimization scheme integrating a convolutional block attention module and the cross-stage partial network (CSPNet) was developed for ResNet-50. Besides, the impacts of cross-stage parameter in CSPNet's classification performance are experimentally investigated, and it is demonstrated that the default cross-stage parameter in CSPNet is not always optimal. Furthermore, a ranger optimizer is proposed to obtain proper network parameters, which outperforms the traditional Adam optimizer with respect to training efficiency and prediction accuracy. Experiments are carried out on the image dataset collected from real-world wood surface defects. The experimental results illustrated that the proposed model achieves better defects on wood surfaces than state-of-the-art methods, with an overall detection accuracy of 86.25% in detecting knots, cracks, and color-related defects in images. These results demonstrate the effectiveness and feasibility of the proposed modeling and control method. [ABSTRACT FROM AUTHOR]