Vehicle Paint Defect Data Generation Using Style Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The vehicle painting process, particularly the top coating process, lacks sufficient data for automated defect inspection due to reliance on manual devices, while the middle coating process has abundant defect image data.
Innovation Solution
A method and device for generating vehicle paint surface data by performing style transfer on images from the data-abundant middle coating process to match the style of the data-scarce top coating process, using a style transfer network with an Adaptive Instance Normalization (AdaIN) layer to create additional defect images for top coating.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual devices for naked-eye visual inspection are used in the top coating process, then the inspection can be performed, but the data accumulation for training AI networks is insufficient
Solution Approach 1:
The patent creates synthetic defect images by copying and transforming defect patterns from the middle coating process. Style transfer networks generate realistic defect images for the top coating process by adapting visual characteristics from available middle coating data, enabling AI training without requiring actual top coating defect samples.
Solution Approach 2:
The patent performs preliminary data preparation by pre-processing and annotating middle coating defect images before they are needed for top coating inspection. Defect images are extracted, labeled, and stored in advance, then transformed when required for training inspection models in the top coating process.
2Quantity of substance
If style transfer is performed on all first process defect images, then more comprehensive second process defect images can be generated, but the processing time and computational resources increase
Solution Approach 1:
The patent selectively applies style transfer to representative defect images rather than processing every single defect image. By choosing a subset of diverse and representative defect cases from the middle coating process, the system generates sufficient training data for the top coating process while avoiding unnecessary computational overhead.
3Extent of automation
If vision inspection devices are deployed in the middle coating process, then automated inspection capability is improved, but the same technology cannot be directly applied to the top coating process due to data scarcity
Solution Approach 1:
The patent creates a universal defect database that can serve multiple coating processes. By generating synthetic defect images through style transfer, the same AI inspection model architecture and training framework can be applied across different coating processes (middle coating, top coating, and potentially others) without requiring process-specific defect data collection systems.
Data Source
AI summary
A method of generating vehicle paint surface data includes: obtaining first process paint surface images of vehicles in a first process among processes for producing the vehicles. The method also includes storing, as first process defect images, images that contain paint surface defects, from among the first process paint surface images. The method additionally includes obtaining second process paint surface images of the vehicles in a second process that is performed after the first process. The method also includes generating second process defect images by performing a style transfer on the first process defect images to match a paint surface style of the second process, by using some or all of the second process paint surface images.