The Booming Semiconductor Industry: Challenges and Innovations in Defect Detection
The semiconductor industry is experiencing unprecedented growth, driven by an increasing demand for electronics and technological advancements. As noted by Huang & Pan (2015) and ByungHun (2022), manufacturers are under immense pressure to enhance yield, reliability, and process efficiency. This need is further complicated by the introduction of various process-induced variations that manifest as surface anomalies during wafer fabrication. These anomalies, including particle contamination, pattern defects, and microscratches, can significantly impact the performance and reliability of semiconductor products.
The Need for Advanced Inspection Systems
To address these challenges, companies are moving away from manual inspections, which are time-consuming and error-prone. Research has increasingly focused on automated optical inspection (AOI) systems equipped with sub-micron resolution capabilities (Ma et al., 2023). However, certain types of defects—those with blurred outlines or low contrast against complex backgrounds—remain elusive for detection systems. This underlines the importance of improving defect detection capabilities in industrial settings, and current research is collaborating closely with major semiconductor manufacturers to devise innovative solutions.
Navigating the World of Image Preprocessing
An essential aspect of improving defect detection lies in image preprocessing, particularly in the realm of denoising. Traditional methods, such as wavelet-threshold-based denoising (Donoho & Johnstone, 1995) and Fourier transformation (Buades, Coll, & Morel, 2005), have generally succeeded in reducing additive white Gaussian noise. However, they often fall short in preserving critical image details and structural integrity, especially when it comes to intricate edges and textures (Zhang et al., 2017).
To overcome these limitations, advanced deep learning techniques, particularly generative adversarial networks (GANs), have emerged as promising alternatives. Models like the DN-GAN (Chen et al., 2020) and ESRGAN (Wang et al., 2018) have demonstrated superior performance in enhancing image quality, providing the high fidelity necessary for defect detection tasks.
Moreover, data augmentation strategies play a crucial role in identifying smaller defects that present limited feature information. Addressing issues of class imbalance, augmentation enhances model robustness, helping to improve the overall detection rates for rare defects (Kisantal et al., 2019).
Innovations in Object Detection Methods
With significant strides in object detection technologies, methods can now be categorized into two-stage and one-stage variants. The former, like the Faster Region-Based Convolutional Neural Network (Ren et al., 2015), offers high precision but is typically slower and resource-intensive. One-stage detectors, such as SSD (Liu et al., 2016) and RetinaNet (Lin et al., 2017), prioritize speed but once had relatively lower precision.
Recent advancements in one-stage detectors have blurred these distinctions, producing high accuracy without sacrificing operational speed. Models like YOLOv5, YOLOv7 (Wang, Bochkovskiy, & Liao, 2023), and YOLOv9 (Wang, Yeh, & Mark Liao, 2024) offer unmatched performance, particularly in real-time applications, including the vital task of semiconductor defect detection.
Addressing Limitations in Static Detection Systems
Despite advancements, existing static detection systems for defects, particularly in contact ring seals (CRSs) used during electroplating in semiconductor manufacturing, still face considerable challenges. Current methodologies often utilize a fixed-region-of-interest (ROI) approach with predefined coordinates for defect detection. This practice has inherent weaknesses, including an inability to adapt to geometric variations or repositioned seals, ultimately yielding high false-negative rates, especially when CRSs undergo movement during the manufacturing process.
An Innovative Dynamic Detection System
To combat these issues, recent research has developed a dynamic detection system that employs artificial intelligence for enhanced defect identification on CRSs. This system integrates advanced segmentation techniques with traditional image processing methods to produce a robust and effective detection mechanism.
Key Contributions of the Dynamic System
-
Two-Stage Contour Extraction Method: This approach merges residual structures with U-Net for efficient segmentation, followed by a Hough transform to accurately detect circular shapes. This strategy optimally reduces image complexity while enhancing defect localization.
-
Innovative Sampling Method: To minimize interference from background noise, this method accurately crops regions with defects, significantly improving the ability to identify subtle anomalies.
- Adaptive Classifier: The system’s classifier dynamically adjusts according to the task’s complexity, allowing for flexible and precise defect detection in various operational scenarios.
It’s crucial to highlight that this system represents a comprehensive engineering integration rather than introducing new algorithms. By systematically merging existing deep learning architectures (such as ResUNet) with established computer vision techniques (like the Hough transform), the framework is specifically tailored for the intense demands of industrial semiconductor manufacturing environments.
The focus here is on addressing practical challenges through innovative integration methodologies, achieving better detection accuracy and expedited response times when compared to static systems. This advancement translates to increased operational efficiency and reduced manual intervention in semiconductor manufacturing processes, ultimately paving the way for a brighter future in the ever-evolving landscape of technology.