A Deep Learning Algorithm for Particle Segmentation of Aerosol Images
Journal of Flow Visualization and Image Processing
First page number:
Last page number:
This paper addresses the task of particle segmentation in high-resolution aerosol images. A particle-boundary model is introduced to predict particles and their boundaries, simultaneously using a fully convolutional neural network. Given a RGB color image, the model directly outputs a predicted particle map and a boundary map. A simple and parameter-free watershed post-processing procedure is performed on the predicted particle and boundary map to produce the final segmented particles. An overlapping 256 × 256 patch extraction method is also designed for seamless prediction of the particle in large images. During the training phase, a combined loss function is used to alleviate the problem of unbalanced classes. We compare two weight initialization schemes: He's normal initializer and one pretrained on the 2018 Data Science Bowl. Experimental results show that the pretrained network is able to converge much faster to its steady value in comparison with the nonpretrained network and achieve higher intersection over union scores. Our method also outperforms the traditional Otsu and triangle methods and gets better segmentation results on raw and contrast adjusted test images.
Computer vision; Image segmentation; Deep learning; Fully convolutional neural network; Aerosol image processing
Computer Sciences | Environmental Health
Chen, L. A.,
A Deep Learning Algorithm for Particle Segmentation of Aerosol Images.
Journal of Flow Visualization and Image Processing, 26(3),