University of Twente Student Theses


Self-supervised semantic segmentation based on self-attention

Pandey, Vaidehi (2021) Self-supervised semantic segmentation based on self-attention.

[img] PDF
Abstract:The unavailability of a significant number of annotated images is affecting the medical imaging area of computer vision. This unavailability is because of the fact that labeling a medical image is a time consuming task as it requires careful analysis of the whole image and can only be done by an expert. Supervised learning method will not give good results due to lack of input-output pairs. Self-supervised learning gives us a way out by transforming the images themselves to labels thus not requiring human-annotated labels in larger amounts. This paper proposes a self-supervised strategy for representation learning which can be further used for other downstream tasks. In our method we integrate colorization task into BYOL which is a contrastive learning method. The resulting self-supervised method is trained on cem500k dataset with two different encoders namely resnet50 and stand alone self-attention. The encoders trained through our self-supervised training method achieved comparable results to the encoders trained with the original BYOL. Further, the self-attention model pre-trained using our method performed better than the rest of the encoders on the semantic segmentation task. We analyzed the Class Activation Map(CAM) and found that the self-attention encoder(pre-trained using our method) activates visually important regions on the image.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page