Improvement of YOLO Detection Strategy for Detailed Defects in Bamboo Strips

It is difficult to detect several detailed defects when detecting surface defects in bamboo strips. The morphology of these defect characteristics exhibits relatively simple patterns but closely resembles the underlying fiber texture or coloration, as exemplified by cracks, mildew, wormholes, and burr formation. In this regard, this study proposes an improved model based on the YOLOv8 deep learning network. The improved model uses dynamic convolution and a Ghost module to improve the C3k2 modules in YOLOv8 to reconstruct its backbone and neck parts, where the research introduces the DySample module to replace the original upsample module to avoid the loss of feature information of targets after the network is used multiple times, further ensuring the detection effect of detailed features, as well as the EMA mechanism in the neck part. Experimental validation of the developed model demonstrated robust detection performance, achieving mAP values of 93.1%, 92.9%, 92.2%, and 92.2% for burr, mildew, cracking, and wormhole detection, with a total mAP of 92.6% and a precision of 81.5%; at the same time, the weight was decreased by 14%. The experimental results show that the improved model in this study has a certain detection effect on difficult-to-identify features on the surface of bamboo strips. This research demonstrates that employing YOLOv8 helps in detecting several challenging minor defects in bamboo strips.