Combined Defect and Location Neural Networks for 3D Object Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to accurately identify defects in three-dimensional objects using two-dimensional images, as they fail to effectively capture points, lines, and curves inherent in these objects.
Innovation Solution
Combining a defect neural network with a location neural network, where the defect neural network identifies defect types and the location neural network determines precise areas of interest using geometric shapes labeled with distance-based value distributions, enabling accurate defect detection in three-dimensional objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single neural network is used to identify defects in three-dimensional objects from two-dimensional images, then the device complexity is reduced, but the measurement precision of defect location and type identification deteriorates
Solution Approach 1:
The patent divides the defect detection task into two separate neural networks: a defect neural network that identifies defect types, and a location neural network that precisely locates defects using geometric shape labels (points, lines, curves). This segmentation allows each network to specialize in one aspect, improving overall measurement precision while maintaining manageable device complexity through modular architecture
Solution Approach 2:
The patent transitions from traditional two-dimensional image analysis to incorporating three-dimensional geometric information by labeling defects with geometric shapes (points, lines, curves) that represent the spatial structure of defects on three-dimensional objects. This dimensional enhancement enables the location neural network to accurately capture the spatial characteristics of defects that cannot be represented by simple bounding boxes
2Ease of operation
If traditional two-dimensional image analysis is used, then the ease of operation is maintained, but the ability to detect points, lines, and curves on three-dimensional objects deteriorates
Solution Approach 1:
The patent changes the labeling parameters from traditional bounding box coordinates to geometric shape parameters (points, lines, curves) that better represent three-dimensional defect structures. This parameter transformation allows the model to capture essential geometric features of defects while maintaining the simplicity of two-dimensional image processing workflows
Data Source
AI summary
In an example embodiment, two different neural networks are used in conjunction with each other to identify defects in items appearing in images. More particularly, a defect neural network is trained on images labeled with an indication of a defect type (or no defect) appears in an item in the image (with optionally an identification of the area in which the defect appears). Separately, a location neural network is trained on images that have a specific geometric shape (e.g., point, line, circle) labeled in the image. The labels identify not just the pixels that are part of the shape, but also provide a value that indicates a degree of the identification of the location. Surrounding pixels to the identified shapes are then given values according to a distribution that is based on distance from the shape, with the peak values being those pixels that are exactly on the identified shape.