Surface defect detection of particleboard is crucial for product quality control and the improvement of production efficiency. However, existing technologies suffer from issues including insufficient accuracy in multi-scale defect recognition, difficulty in matching model inference speed to the requirements of industrial online detection, and accuracy degradation caused by information loss during feature fusion, all of which have restricted the realization of high-quality production management and control. To address the above challenges, the efficient real-time detection (ERTD) architecture proposed in this study achieved rapid and accurate detection of multi-scale surface defects in particleboard through four technical breakthroughs. First, the ghost hierarchical generation lightweight multi-scale feature module was proposed to reduce computational costs while maintaining detection accuracy. Second, the arranged connections convolutional module was developed to deeply mine middle-layer features, and a deep network was constructed by leveraging the complementarity of shallow and deep features, as well as non-downsampling transmission to alleviate gradient vanishing. Meanwhile, the enhanced coordinate pyramid attention module was put forward to resolve the problem of disconnection between high-level and low-level features, thus improving the scale adaptability for small-target detection. Finally, the multi-direction lightweight multi-scale feature fusion architecture was designed to fuse features at all levels with equal priority, comprehensively enhancing the accuracy and efficiency of multi-scale detection. Results demonstrated that the ERTD performed excellently in the task of particleboard surface defect detection: compared with mainstream object detection models, its mean average precision was increased by 1.5%, and the number of parameters was reduced by 20%. In the deployment of practical production lines, the ERTD had completed pilot-scale validation in particleboard manufacturing enterprises in Jiangsu and Guangzhou (China) and was planned for mass application. It could accurately realize online classification and multi-scale recognition of defects such as big shavings, oil pollution, and sand leakage, with a detection accuracy of 97%. Through feature fusion and structural innovation, the ERTD broke through the limitations in multi-scale recognition and real-time performance, providing an efficient solution for surface defect detection.
Wang et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: