Neural Networks for Cross-Section Segmentation in Raw Images of Log Ends

In this paper, wood cross-section (CS) segmentation of RGB images is treated. CS segmentation has already been studied for computed tomography images, but few study focuses on RGB images. CS segmentation in rough log ends is an important feature for the both assessment of wood quality and wood traceability. Indeed, it allows to extract other features like pith, eccentricity (distance between the pith and the geometric centre) or annual tree rings which are related to mechanical strength. In image processing, neural networks have been widely used to solve the problem of objects segmentation. In this paper, we propose to compare different state-of-the-art neural networks for CS segmentation task. In particular, we consider U-Net, Mask R-CNN, RefineNet and SegNet. We create an imageset which has been split into 6 subsets. Considered neural networks have been trained on each subset in order to compare their performance on different type of images. Results show different behaviors between neural networks. On the one hand, overall U-Net learns better on small dataset than the others. On the other hand, RefineNet learns well on huge dataset. While SegNet is less efficient and Mask R-CNN does not provide a detailed segmentation. This offers a preliminary result on neural network performances for CS segmentation.