Skip to Main content Skip to Navigation

Machine learning for image segmentation

Abstract : In this PhD thesis, our aim is to establish a general methodology for performing the segmentation of a dataset constituted of similar images with only a few annotated images as training examples. This methodology is directly intended to be applied to images gathered in Earth observation or materials science applications, for which there is not enough annotated examples to train state-of-the-art deep learning based segmentation algorithms. The proposed methodology starts from a superpixel partition of the image and gradually merges the initial regions until anactual segmentation is obtained. The two main contributions described in this PhD thesis are the development of a new superpixel algorithm which makes use of the Eikonal equation, and the development of a superpixel merging algorithm steaming from the adaption of the Eikonal equation to the setting of graphs. The superpixels merging approach makes use of a region adjacency graph computed from the superpixel partition. The edges are weighted by a dissimilarity measure learned by a machine learning algorithm from low-level cues computed on the superpixels. In terms of application, our approach to image segmentation is finally evaluated on the SWIMSEG dataset, a dataset which contains sky cloud images. On this dataset, using only a limited amount of images for training our algorithm, we were able to obtain segmentation results similar to the ones obtained with state-of-the-art algorithms.
Document type :
Complete list of metadatas

Cited literature [129 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, March 18, 2020 - 8:46:08 AM
Last modification on : Wednesday, October 14, 2020 - 3:52:39 AM
Long-term archiving on: : Friday, June 19, 2020 - 12:59:59 PM


Version validated by the jury (STAR)


  • HAL Id : tel-02510662, version 1


Kaiwen Chang. Machine learning for image segmentation. Machine Learning [stat.ML]. Université Paris sciences et lettres, 2019. English. ⟨NNT : 2019PSLEM058⟩. ⟨tel-02510662⟩



Record views


Files downloads