Research On Surface Defect Detection Of Wood Based On Deep Learning

As society continues to develop and economic levels increase,the demand for wood as an important resource continues to rise.However,wood defects can negatively impact the quality of wood and reduce its value.Therefore,it is essential to detect and process defects during wood processing in order to increase wood utilization.Traditional visual inspection and sensor detection methods are no longer adequate for industrial defect detection,and China’s automated level in wood processing is relatively low.In recent years,deep learning algorithms and machine vision technology have made great progress in the field of defect detection,and compared to traditional object detection methods,these technologies have advantages such as strong feature learning ability,good adaptability,and the ability to process complex scenes,attracting the attention of researchers both domestically and internationally.YOLOv5(You Only Look Once version five)is recognized by scholars at home and abroad as the best model in the field of deep learning in recent years.To address the above issues,this article proposes a wood surface defect detection method based on YOLOv5 s,improving the algorithm’s adaptability and increasing detection accuracy and efficiency.Secondly,the wood surface defect detection system is transplanted to an embedded device in order to balance model lightness,detection speed,and recognition accuracy.Therefore,this paper employs deep learning algorithms for wood surface defect detection research,and the main research content and results are as follows:(1)A wood surface defect detection dataset that is matched with the experimental requirements was produced.The dataset contains 4359 images,and takes into consideration factors such as the small number of defect images for individual target classes to enhance detection robustness.At the same time,experiments were carried out on the self-made dataset,verifying that the proposed detection model can be used for wood surface defect detection.(2)A YOLOv5-based model improvement method is proposed to solve the problem of large deep learning model volume and poor real-time performance.Firstly,the Ghost Bottleneck module is used to replace the convolution module in YOLOv5 to reduce parameter volume.Secondly,the attention mechanism is added to the main network to improve detection accuracy.Finally,the CIOU loss function is introduced to improve the speed of model training convergence.The improved model reduces parameter volume and volume by 47.89% and 47.22% respectively,achieving an average precision of 93.54%.The mean average precision for the four classes of defect targets,dead knot,living knot,crack,and hole are 98.66%,99.06%,98.10% and 96.53%,respectively.While maintaining detection accuracy,the detection speed is increased by 5.9 FPS,and it has good generalization performance.This effectively solves the problems of low detection efficiency and poor generalization in existing methods.(3)The improved algorithm was experimentally ported from the PC to an embedded platform,achieving miniaturization and portability.The embedded platform reads the video stream through a camera and inputs it into the improved YOLOv5 model for inference after decoding and format conversion.NCS2 is used to optimize detection performance,drastically improving the model’s detection speed on the embedded platform.Secondly,third-party libraries such as Qt software,Open CV and Libtorch are used to develop a wood surface defect detection software with major functions such as defect detection and result display,which has good humancomputer interaction.After testing,the improvement model achieved a detection speed of 15.6FPS on the Raspberry Pi embedded platform,achieving accurate and real-time detection of wood surface defects.