Welding Keypoint Detection Using Transformed Training Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training keypoint detection models in welding systems fail to improve accuracy in irregular situations, and using images from normal operation as supervised data increases processing load.
Innovation Solution
Generate supervised data by transforming images from normal operation to simulate irregular situations, using techniques like image transformation and generative AI models to create a detection model that can accurately detect keypoints in various conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If images from normal operation are used as supervised data, then processing load is reduced, but detection accuracy in irregular situations deteriorates
Solution Approach 1:
The system performs preliminary action by generating transformed images that simulate irregular situations before actual welding operations. The image generation unit creates synthetic supervised data including transformed welding images with various irregularities (position shifts, rotations, scaling) to prepare the detection model for handling diverse scenarios. This preliminary preparation allows the model to be trained on a broader range of conditions without requiring actual irregular operation data, thus reducing processing load while maintaining detection accuracy.
Solution Approach 2:
The system applies parameter changes by transforming the normal welding images through various geometric transformations (translation, rotation, scaling) to generate synthetic irregular situations. The image generation unit modifies parameters such as position coordinates, orientation angles, and size ratios of keypoints in the training images to create diverse training data. This enables the detection model to learn robust feature recognition across different conditions without requiring additional physical data collection, thereby maintaining accuracy while reducing processing burden.
2Ease of manufacture
If detection model is trained only on normal operation images, then training data acquisition is simple, but adaptability to irregular situations deteriorates
Solution Approach 1:
The system uses copying by creating synthetic copies of normal welding images with applied transformations to represent irregular situations. The image generation unit generates multiple transformed versions of each training image, creating a expanded training dataset that includes position-shifted copies, rotated copies, and scaled copies. This copying approach allows the model to learn from diverse scenarios without requiring physical collection of irregular operation data, maintaining ease of data acquisition while improving adaptability.
Solution Approach 2:
The system performs preliminary action by pre-generating transformed training images that simulate various irregular situations before actual deployment. The image generation unit proactively creates synthetic training data with different geometric transformations to prepare the detection model for handling diverse welding conditions. This preliminary preparation ensures the model is trained on comprehensive data without requiring complex data collection processes during irregular operations, thus maintaining simplicity while enhancing adaptability.
Data Source
AI summary
An information processing device includes one or more hardware processors. The hardware processors generate at least one piece of second supervised data by using at least one piece of first supervised data. The first supervised data includes at least one welding image capturing a welding target, and true values of positions of a plurality of keypoints within the welding image. The second supervised data includes a transformed image obtained by transforming the welding image so as to change relative positions of the keypoints, and true values of changed positions. The hardware processors learn a detection model by using the first supervised data and the second supervised data. The detection model receives input of the welding image and outputs the positions of the keypoints.