A deep learning-based algorithm for online detection of small target defects in large-size sawn timber

As a core material in wood structure buildings, the surface quality and grade of timber are crucial to the safety of these structures. The objectives of this study are a) to develop an online detection system for small target defects in large-size Pinus densiflora sawn timber using machine vision/deep learning technology; b) to enhance timber productivity and quality by efficiently and accurately detecting defects using predictive models generated by a deep learning-based algorithm. The predictive models generated by deep learning with a YOLO-integrated network structure are utilized in this study. The proposed methods, including image stitching, segmentation, and fusion techniques based on SIFT features, enable the input of large-size sawn timber oversize images while preserving the integrity of the information. The efficient layer aggregation network enhances machine vision defect detection on timber sawing lines, adapting to variable environments with a focus on Pinus densiflora timber. The results indicated that the machine vision defect detection device is able to predict candidate bounding boxes and class probabilities for multiple types of knots in complex, naturally characterized materials, even under conditions of background noise and interfering factors. Comparing the detection results of the proposed system with the statistical outcomes of manual visual inspections under production conditions involving long hours and large quantities of sawn timber yielded an identification and detection accuracy of 90.37 %. The system's speed for detecting knot defects on the surface of sawn timber can reach 40 m/min, making it suitable for practical application of wood product processing lines.