Hybrid Pixel-Superpixel Structures for Enhanced Image Segmentation: Integrating Boundary Information in Deep Learning Models
This project explores novel approaches to image segmentation using U-Net, leveraging superpixels to enhance accuracy. The first part investigates augmenting standard image inputs by encoding
and integrating superpixel information, including an extension that reintroduces this information
throughout the encoder. While results show that these methods can offer consistent improvements over the baseline, the gains are modest and suggest room for further optimization. The
second part introduces a hybrid data structure, the Superpixel-Integrated Grid (SIGrid), which
embeds superpixel boundary, shape, and color descriptors into a regular n × n grid. SIGrid enables more efficient training on smaller architectures while achieving noticeably higher segmentation accuracy, highlighting its potential as a lightweight and effective input representation. The
code developed for this project can be found at: https://github.com/JackRobs25/Honors
and integrating superpixel information, including an extension that reintroduces this information
throughout the encoder. While results show that these methods can offer consistent improvements over the baseline, the gains are modest and suggest room for further optimization. The
second part introduces a hybrid data structure, the Superpixel-Integrated Grid (SIGrid), which
embeds superpixel boundary, shape, and color descriptors into a regular n × n grid. SIGrid enables more efficient training on smaller architectures while achieving noticeably higher segmentation accuracy, highlighting its potential as a lightweight and effective input representation. The
code developed for this project can be found at: https://github.com/JackRobs25/Honors
