Loading...
Publication
OPTIMIZING IMAGE-BASED TASKS IN MANUFACTURING WITH RGB-D FUSION
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 40.57 MB | Adobe PDF |
Authors
Abstract(s)
The manufacturing industry is undergoing a significant transformation with the
onset of the Fourth Industrial Revolution. A key aspect of this shift is the integration
of advanced technologies, such as smart sensors and automation, into production
processes. Within this context, 3D cameras have become invaluable, enabling
manufacturers to capture precise surface measurements of the products. In response,
the computer vision community has begun exploring new methods to combine
depth information with color data, enhancing existing solutions for classification and
design generation. By harnessing these advancements, manufacturers can streamline
production lines, reduce waste, and elevate product quality.
This dissertation presents two key innovations for classification and generation tasks:
(1) a novel branched Convolutional Neural Network (CNN), which achieves stateof-
the-art performance in RGB-Depth (RGB-D) image classification, and (2) a novel
branched Generative Adversarial Network (GAN), inspired by the same branched
architecture, that delivers state-of-the-art results on the Stanford Cars dataset. The
core idea of this branched approach is to specialize each branch to handle a specific
modality.
In the experiments, the classification performance improved by approximately 1%,
while achieving nearly three times the speed of the next best method. For image
generation, results varied depending on the dataset. On the Stanford Cars benchmark,
the model showed slight improvements in image quality and better diversity.
Description
Keywords
Branched CNN Branched GAN Image Classification Image Generation Image Manipulation RGB-D Fusion