GPU-Based Mura Defect Detection in TFT-LCD Panels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automatic optical detection methods for Mura defects in TFT-LCDs suffer from poor self-adaptation, high false detection rates, and inefficiency, especially when dealing with large uneven brightness areas or noise, and require significant calculation resources.
Innovation Solution
A GPU-based TFT-LCD Mura defect detection method that uses a studentized residual based double-second-order regression diagnosis model, excludes outliers and influential points, and employs double-N-order polynomial surface fitting, combined with Gabor filtering and morphological processing for accurate defect detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automatic optical detection method is used for Mura defect detection, then detection efficiency is improved, but detection precision deteriorates due to poor self-adaptation and high false detection rate
Solution Approach 1:
The patent transforms the detection approach by changing from traditional single-order polynomial fitting to double-N-order regression diagnosis model. This parameter change in the mathematical model enables better adaptation to different Mura defect types and sizes, improving both detection precision and efficiency simultaneously by accurately characterizing the complex brightness distribution patterns
Solution Approach 2:
The patent replaces the conventional CPU-based sequential processing with GPU parallel processing architecture. This substitution leverages the massive parallel computing power of GPU to perform simultaneous calculations for multiple image regions, resolving the contradiction between detection speed and precision by enabling complex algorithms to run efficiently at high speed
2Manufacturing precision
If B-spline surface fitting is used for Mura defect detection, then background fitting is performed, but calculation amount increases significantly and detection effect deteriorates for large uneven brightness areas
Solution Approach 1:
The patent segments the fitting process into two distinct stages: first performing second-order regression to remove large-scale brightness trends, then applying N-order polynomial fitting to capture local variations. This segmentation divides the complex calculation into manageable parts, reducing overall computational burden while maintaining high fitting accuracy for both large and small Mura defects
Solution Approach 2:
The patent implements a dynamic adaptive fitting strategy where the polynomial order N is adjusted based on the characteristics of the image region being analyzed. For regions with large uneven brightness areas, lower polynomial orders are used to avoid overfitting, while for regions with fine details, higher orders are applied. This dynamic approach optimizes calculation amount while preserving detection accuracy across different defect types
3Measurement precision
If CPU-based processing is used for defect detection, then detection accuracy is maintained, but detection time increases significantly for large-format and multiple images
Solution Approach 1:
The patent substitutes the CPU sequential processing system with a GPU parallel processing system. The GPU's architecture with thousands of cores enables simultaneous processing of multiple image regions and multiple images, maintaining the accuracy of complex regression calculations while reducing detection time by factors of 10-100x for large-format and multi-image scenarios
Data Source
AI summary
The present disclosure discloses a GPU-based TFT-LCD Mura defect detection method, comprising: (1) establishing a studentized residual based double-second-order regression diagnosis model based original image data to obtain double-second-order regression background data; (2) obtaining influence quantities of respective data points on fitted values according to the original image data and the double-second-order regression background image data; (3) excluding outliers and influential points in the original image data according to the influence quantities to obtain a new pixel point set; (4) establishing a double-N-order polynomial surface fitting model according to the new pixel point set to obtain double-N-order background image data; (5) obtaining a residual image R according to the double-N-order background image data and the original image data, and performing threshold segmentation on the residual image to obtain a threshold segmentation image; and (6) performing morphological processing on the threshold segmentation image to obtain an eroded and dilated image, thereby achieving effective segmentation of Mura defects with uneven brightness distribution.