Erkennung von Defekten mittels eines Text-Zu-Bild-Diffusionsmodells

Details

Veröffentlichungs-Nr.

EP4756738

Anmeldetag

20. November 2025

Veröffentlichung

10. Juni 2026

Rechtsraum

EP

IPC

G06V10/82, G06V20/60

Abstract

Methods for fine-tuning a convolutional neural network of a Text-To-Image Diffusion Model within a context of recognizing defects of manufactured products within images of those products are disclosed. Images of manufactured images that have various scratches, dents, or other defects are provided to the model along with a word or phrase indicating that there is a defect. The model then learns to identify the portion of the overall image that includes the defect. The learning of this type of task is based on the use of segmentation masks that correspond to the images, which are then used along with cross-attention maps of the model in order to calculate an average defect mask loss parameter of the model. By computing this parameter and applying it when updating weights of the model, the model can be fine-tuned to detect defects of manufactured products.

Anmelder

Firma

Robert Bosch GmbH

Land

🇩🇪 Deutschland