Contact Lens Package Defect Detection Using Synthetic Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current quality control methods for contact lens packages are inadequate in detecting physically and digitally implanted foreign matter and defects, such as holes and edge defects, with sufficient accuracy and efficiency.
Innovation Solution
The development of systems and methods that utilize machine learning algorithms, specifically convolutional neural networks (CNNs), trained on datasets containing augmented images of contact lens packages with implanted foreign matter and defects. These models are configured to detect the existence of foreign matter and defects, and once validated, they analyze captured images to output a quality control metric indicating an accept or reject condition of the package.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional quality control inspection methods are used for contact lens packages, then the process is simple and fast, but the detection accuracy for foreign matter and defects is insufficient
Solution Approach 1:
The patent creates synthetic copies of defective contact lens packages by digitally implanting foreign matter and defects into images of normal packages. This synthetic data copying approach enables the training of detection models without requiring physical production of large numbers of defective packages, thereby improving detection accuracy while avoiding the complexity of physical defect generation systems
Solution Approach 2:
The patent performs preliminary training of machine learning models using synthetic defect data before actual quality control inspection. By pre-training models with artificially generated defective images, the system achieves high detection accuracy without requiring complex real-time analysis of rare physical defects during production
2Measurement precision
If large-scale physical production of defective packages is conducted to train detection models, then the model accuracy improves, but the time and resources required increase significantly
Solution Approach 1:
Instead of producing large numbers of physical defective packages for model training, the patent digitally copies and manipulates images of normal packages to create synthetic defective images. This copying approach generates unlimited training data instantaneously, achieving high model accuracy without the time-consuming process of physical defect production
Solution Approach 2:
The patent replaces the mechanical process of physically creating defective packages with a digital image processing system. By substituting physical defect generation with computational algorithms that digitally implant foreign matter and defects into images, the system eliminates the time and resource requirements of large-scale physical production while maintaining model training effectiveness
3Reliability
If traditional inspection methods are used, then the processing speed is fast, but the ability to detect rare defects is insufficient
Solution Approach 1:
The patent performs preliminary training of detection models using extensive synthetic defect data before deployment. This pre-training ensures the model has already learned to recognize various defect patterns, enabling it to reliably detect rare defects during actual inspection without sacrificing processing speed, as no additional analysis time is required during production
Solution Approach 2:
The patent transforms the detection problem by changing the parameters of training data rather than altering the inspection process itself. By modifying image parameters through digital implantation of foreign matter and defects at various positions, sizes, and types, the system creates diverse training scenarios that enhance the model's ability to detect rare defects while maintaining fast processing speeds during actual inspection
Data Source
AI summary
A method for quality control of contact lens packages comprises receiving a first data set comprising a plurality of images of a contact lens package having physically implanted defects; receiving a second data set comprising a plurality of images of a contact lens package having digitally implanted defects; testing and training, on the first and second data set, a model to determine a validated one or more quality control models; capturing image data of a package of a contact lens; analyzing, based on the validated one or more quality control models, the image data; and causing, based on the analyzing, output of a quality control metric indicative of at least an accept or reject condition of the package.