An Algorithm Based on DAF-Net++ Model for Wood Annual Rings Segmentation

• DOI: 10.3390/electronics12143009
• Publication Date: 2023-07-10T04:47:35Z
• Authors: Zhedong Ge; Ziheng Zhang; Liming Shi; Shuai Liu; Yisheng Gao; Yucheng Zhou; Qiang Sun
• Keywords: Net (polyhedron)

The semantic segmentation of annual rings is a research topic interest in wood chronology. To solve the problem being difficult to segment dense areas and greatly affected by defects such as cracks wormholes, this paper builds DAF-Net++ model which based on U-Net whose backbone network VGG16 filled with jump links, CBAM DCAM. In model, used enhance extraction ability image features, links are fuse information different levels, DCAM provides weighting guidance for shallow solves loss down-sampling information. Taking Chinese fir experimental object, 1700 CT images transverse section were obtained medical equipment 120 them randomly selected dataset, was expanded cropping rotation, among others. training rings, finally performance evaluated. method freeze followed thaw training, takes Focal Loss function, ReLU activation Adam optimizer. results show that, MIoU 93.67%, MPA 96.76%, PA 96.63%, Recall 96.76%. Compared other models U-Net, U-Net++, DeepLabv3+, etc., has better performance.