ABSTRACT
In medical and material science, 3D reconstruction is of great importance for quantitative analysis of microstructures. After the image segmentation process of serial slices, in order to reconstruct each local structure in volume data, it needs to use precise object tracking algorithm to recognize the same object region in adjacent slice. Suffering from weak representative hand-crafted features, traditional object tracking methods always draw out under-segmentation results. In this work, we have proposed an adjacent similarity based deep learning tracking method (ASDLTrack) to reconstruct 3D microstructure. By transferring object tracking problem to classification problem, it can utilize powerful representative ability of convolutional neural network in pattern recognition. Experiments in three datasets with three metrics demonstrate that our algorithm achieves the promising performance compared to traditional methods.
