O-Net: annual rings CT image segmentation method based on attention and residual path

Abstract

Annual rings can record natural information such as regional temperature, humidity, and rainfall. However, they are often damaged by cracks, knots, and wormholes, making it difficult to obtain accurate data from them. In this study, O-Net is developed for the segmentation of annual rings in wood transverse sections containing cracks, wormholes, and knots. The O-Net framework is constructed by integrating parallel U-shaped encoder-decoder paths, utilizing convolutional kernels of two different sizes to capture multi-scale features of annual rings. Subsequently, the Efficient Channel and Spatial Attention (ECSA) mechanism is incorporated into the encoder to enhance the model’s ability to extract annual rings features near defects from both channel and spatial dimensions. Finally, a residual path is introduced through ablation experiments to refine the model’s architecture. The results demonstrate that O-Net effectively distinguishes between crack and knot boundaries and accurately segments annual rings, even in the presence of cracks and minor wormholes. O-Net is capable of extracting transverse sectional annual rings from wood containing cracks, wormholes, and knots, addressing the limitations of traditional methods in accurately extracting damaged annual rings with defects.