Segmentation of wood CT images for internal defects detection based on CNN: A comparative study

•CT scanning and CNN models were used for internal wood defects segmentation.•Five CNN models were evaluated for segmenting three common internal wood defects.•The U-net model achieved the highest segmentation accuracy for wood defects.•YOLOv8-seg balanced detection and segmentation accuracy and efficiency. Computed Tomography (CT) scanning enables the timely detection of internal defects in wood, effectively increasing the utilization rate of wood and reducing processing costs. Efficient and accurate detection for internal defects in wood CT images remains to be a significant challenge. The purpose of this study is to find an efficient way for detecting and segmenting internal wood defects to meet diverse performance requirements. Three common internal wood defects were studied: knots, decay, and hollows. CT scanning was used to scan two logs, and convolutional neural network (CNN) algorithms were applied to detect and segment defects based on CT images. Five CNN models (U-net, DeeplabV3+, PSPnet, HRnet, and YOLOv8-seg) were evaluated for their accuracy and efficiency in detecting and segmenting different internal wood defects. Results shown that U-net achieved the highest segmentation accuracy but with the lowest efficiency. YOLOv8-seg model exhibited the best detection and segmentation efficiency, with segmentation accuracy second only to U-net. When considering both accuracy and efficiency, the models were ranked as follows in descending order: YOLOv8-seg, HRnet, DeeplabV3+, U-net, and PSPnet. Additionally, performance of all models varied across the three defects studied. All models detected knots more rapidly and accurately compared to decay and hollows. YOLOv8-seg model demonstrated superior detection and segmentation performance overall and is recommended for internal wood defect detection and segmentation tasks. U-net excelled in segmentation accuracy, suitable for scenarios where precision is a priority. This study provided a better and more flexible way for detecting and segmenting internal wood defects to help solve the practical problems in wood utilization.