Stacked Denoising Autoencoder for Semiconductor Image Noise Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional noise filtering methods in semiconductor defect inspection struggle to selectively suppress nuisance signals without compromising defect of interest (DOI) signals, especially in Advanced Wafer Level Packaging, due to challenges in distinguishing between defects and noise induced by grain roughness.

Innovation Solution

A deep learning-based approach using a stacked denoising autoencoder is employed to separate defects of interest from nuisance signals by training a deep learning module with paired noisy and de-noised images, applying techniques like pixel blob selection, smoothing, dilation, erosion, blurring, template replication, flood fill, median filtering, and low-pass filtering to reduce characteristic noise.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If conventional noise filtering methods (median filtering, low-pass filtering) are used, then noise is reduced, but defect of interest signals are also compromised

Engineering Contradiction:
ImprovenoiseVSAvoiddefect detection accuracy
Core Design Contradiction:
Object-affected harmful factorsVSMeasurement precision

Solution Approach 1:

The system performs preliminary classification of image regions into DOI regions and nuisance regions before applying noise filtering. By pre-identifying which regions contain defects of interest using template matching and region classification algorithms, the system can then selectively apply filtering only to nuisance regions, preserving DOI signals while reducing noise in appropriate areas.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system applies different processing strategies to different regions of the image based on their classification. DOI regions receive minimal or no filtering to preserve signal integrity, while nuisance regions undergo aggressive noise filtering. This localized approach allows the system to optimize noise reduction in each region according to its specific characteristics and requirements.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If aggressive noise filtering is applied to suppress nuisance signals, then nuisance classification improves, but DOI signals are compromised

Engineering Contradiction:
Improvenuisance classification accuracyVSAvoiddefect detection reliability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The system performs preliminary classification of image regions into DOI regions and nuisance regions before applying noise filtering. By pre-identifying which regions contain defects of interest using template matching and region classification algorithms, the system can then selectively apply filtering only to nuisance regions, preserving DOI signals while reducing noise in appropriate areas.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system applies different processing strategies to different regions of the image based on their classification. DOI regions receive minimal or no filtering to preserve signal integrity, while nuisance regions undergo aggressive noise filtering. This localized approach allows the system to optimize noise reduction in each region according to its specific characteristics and requirements.

Inventive Principle:
Principle #3Local quality

3Object-affected harmful factors

If deep learning-based denoising is used, then selective noise suppression is achieved, but computational complexity increases

Engineering Contradiction:
ImprovenoiseVSAvoidcomputational complexity
Core Design Contradiction:
Object-affected harmful factorsVSDevice complexity

Solution Approach 1:

The system segments the image processing task into distinct stages: template matching, region classification, and selective denoising. By dividing the workflow into modular components, the system can apply computationally intensive deep learning denoising only to identified nuisance regions rather than processing the entire image, significantly reducing overall computational complexity while maintaining effective noise suppression where needed.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3891491B1Image noise reduction using stacked denoising auto-encoder
Publication Date: 2025.08.20 KLA CORP
  • EP3891491B1 patent drawingFigure 1A~1B
  • EP3891491B1 patent drawingFigure 2A~2B
  • EP3891491B1 patent drawingFigure 3

AI summary

Embodiments disclosed herein may comprise receiving a run-time image of a run-time die and, with a deep learning module, identifying a characteristic noise in the run-time image, and modifying the run-time image to reduce the characteristic noise, thereby generating a de-noised run-time image. Such embodiments may be performed as methods, by systems, or from non-transitory computer-readable storage media on one or more computing devices. An image sensor of a metrology tool may capture the run-time image of the run-time die. The metrology tool may include a run-time die disposed on a specimen, a run-time image sensor, and a processor in electronic communication with the run-time image sensor. Embodiments may further comprise receiving a training image of a training die, modifying the training image, and training the deep learning module to identify the characteristic noise in the run-time image and modify the run-time image.