Author: Richard Segovia
Supervisors: Ramon Baldrich
Presentation time: 11:15
Virtual Room: 1.4 | Live presentation URL
In this work, we present the development of a novel solution for hyperspectral waste segmentation based on a modified U-NET with a ResNet18 backbone. This project is part of a collaboration between PICVISA, a company that specializes in residue (waste, garbage) sorting systems, and the Computer Vision Center. PICVISA provides the image dataset and groundtruth acquired with different hyperspectral cameras. The main goal of the project is to extend existing technologies on RGB images to nonlimited band images. For this, we first explored various methods to correctly normalize the images of the dataset in order to feed them to a neural network. Next, we modified a network that was originally designed to work with RGB images (three channels) to process hyperspectral images (multiple channels) as well. These modifications include the usage of 1×1 convolutions (network-in-network), wider decoder layers, and wider encoder layers. Finally, we found it necessary to group some classes that are variants of the same material because the confusion matrix showed that the model struggled to differentiate these classes. We conclude that the additional information that hyperspectral images provide helps to improve the segmentation. Our results showed that models trained with hyperspectral images perform better than the ones trained with RGB images.