Solder Joint Inspection Model Using Dual-Stage Image Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional neural model supervised learning for solder joint inspection on surface-mount devices is inefficient due to high resource consumption, difficulty in highlighting differences between normal and abnormal images, and insufficient accuracy, leading to high misjudgment rates and the need for additional manpower and time.
Innovation Solution
A method involving training two identification models to differentiate between images with and without solder joints, followed by masking and training a solder joint inspection model using labeled images to generate normal and abnormal solder joint images, enabling quick identification of abnormal solder joints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional neural model supervised learning is used to inspect solder joints, then the model can identify abnormal solder joints, but huge consumption of human and time resources is required to collect abnormal images
Solution Approach 1:
The patent applies preliminary action by pre-training two separate identification models (first identification model for solder joints present, second identification model for solder joints absent) before the actual inspection. This preliminary training phase enables the system to quickly generate training images without requiring manual collection during the inspection process, thus reducing time resource consumption while maintaining inspection accuracy.
2Measurement precision
If traditional neural model supervised learning is used, then solder joint abnormalities can be detected, but the difference between normal and abnormal pictures cannot be highlighted
Solution Approach 1:
The patent applies segmentation by dividing the training image generation process into two separate identification models: the first identification model segments images where solder joints are present, and the second identification model segments images where solder joints are absent. This segmentation allows the system to highlight differences between normal and abnormal pictures by comparing outputs from these two specialized models, thereby improving detection capability while preserving critical visual differences.
3Measurement precision
If traditional neural model training is used, then the model can be trained with available data, but the optimized loss function is not easily converged and misjudgment rate remains high
Solution Approach 1:
The patent applies preliminary action by pre-training two separate identification models with specialized loss functions before combining their outputs. The first identification model is trained with a loss function optimized for detecting solder joints present, while the second identification model uses a loss function optimized for detecting solder joints absent. This preliminary training approach enables better convergence and reduces misjudgment rates by addressing the training complexity issue through divided specialization rather than attempting to train a single model with all requirements.
Data Source
AI summary
A solder joint inspection model training method includes the steps of: training a first identification model according to first sample images to identify a surface-mount device with a solder joint in an image; training a second identification model according to second sample images to identify a surface-mount device without a solder joint in an image; inputting labeled original images to a trained first identification model to output first images; inputting the first images to a trained second identification model to output second images; masking the first images with the second images to generate images with normal solder joints and images with abnormal solder joints; and training a solder joint inspection model based on the images with normal solder joints and the images with abnormal solder joints.