Loading...
Publication
Rule-Based Generative Modeling and Synthetic Image-Driven Deep Learning Classification in Ceramic Industry Applications
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 7.06 MB | Adobe PDF |
Authors
Abstract(s)
The ceramic industry, with its deep historical roots and complex production methodologies, has undergone significant transformation through the adoption of modern technologies. Moving away from labor-intensive techniques, the industry has embraced advancements such as computer-assisted design and computer-assisted manufacturing, which have markedly improved production efficiency, quality control, and customization capabilities. Nevertheless, the growing demand for personalized and bespoke ceramic products introduces new challenges, particularly in automating design generation and production line workflows, which are increasingly reliant on deep learning classification techniques.
This dissertation introduces an approach for generating new three-dimensional ceramic models through a rule-based system by integrating procedural generation methods. This system ensures structural integrity and manufacturability in the creation of diverse ceramic pieces and supports the generation of unique ceramic collections and the replication of ceramic pieces. Given the handiness of the method a three-dimensional ceramic tableware dataset was created and made available to the use in several fields of research. The system was made available through a web-based application which enhances user interaction and integrates augmented reality offering a more immersive experience allowing users to visualize the generated models overlaid in real environments.
With the advancements in the design phase of ceramics and with the potential faster introduction of new models into production, such as those generated by the rule-based system, acquiring real-world images for production line classification systems has become increasingly challenging due to the inherently labor-intensive nature and time-consuming of the task, hindering efficient dataset creation and deep learning model training, leading to more inefficient production lines. To address this issue, this dissertation presents CeramicFlow - an innovative computer graphics rendering pipeline designed to create synthetic images by employing three-dimensional design models of ceramic pieces. By leveraging domain randomization techniques, CeramicFlow generates synthetic image datasets that are validated for real-world ceramic classification tasks. The findings demonstrate that synthetic images can effectively support dataset development and reduce dependency on real-world data for deep learning applications in ceramic classification. The resulting Synthetic CeramicNet dataset provides a valuable resource for future research. The proposed methods show significant promise to be adapted to other industrial fields, potentially transforming how these industries approach automated design and production.
Description
Keywords
Indústria de cerâmica Deep learning Modelação generativa Sistema baseado em regras Imagens sintéticas