通過自洽生成對抗網路進行無監督圖像去噪以檢測面板缺陷

Abstract

在TFT-LCD面板的製造過程中，灰塵、顆粒或設備參數的變化可能使面板產生瑕疵，導致面板亮度差、對比度低。為了找出有缺陷的面板，自動光學檢測（AOI）技術常被用來拍攝圖像並應用基於機器學習的圖像識別方法來檢測和分類缺陷。然而，當相機用於捕捉面板的圖像時，摩爾波紋會扭曲缺陷的外觀，從而難以準確分類瑕疵面板，特別是雲紋缺陷(Mura defect)。因此，在捕獲面板圖像後去除摩爾波紋成為一個關鍵問題。 TFT-LCD 領域中有關摩爾圖案去除的現有文獻通常仰賴成對的數據集(一張帶有摩爾波紋的圖片，有其不帶有摩爾波紋的乾淨版本)來進行訓練，而這在實際製造環境中很難獲得。在本研究中，我們提出了一種無需成對數據即可解決 TFT-LCD 面板摩爾圖案去除問題的新方法。我們的方法利用自洽生成對抗網絡（SC-GAN）和 U-Net 作為生成器，無需成對數據即可實現摩爾波紋的去除。具體來說，我們的 SC-GAN 模型採用對抗性訓練框架，並結合各種損失函數來指導訓練過程。實驗結果透過峰值信噪比（PSNR）和結構相似指數度量（SSIM）兩個指標來將我們提出的模型和其他經典監督去噪模型進行比較。實驗表明，我們提出的模型可以有效地去除缺陷圖像中的摩爾波紋。

During the manufacturing of TFT-LCD panels, dust, particles or variability of the equipment parameters may result in defects, which cause the panels to have poor brightness and low contrast. To find out the defective panels, Automated Optical Inspection (AOI) techniques are commonly employed to capture images and applying machine learning-based image recognition methods to detect and classify defects. However, when the camera is used to capture images of the panels, Moire pattern can distort the appearance of defects, making it difficult to accurately determine defect types, particularly mura defects. Therefore, removing the Moire pattern after capturing the panel images becomes a critical problems. Previous studies on Moiré pattern removal in the TFT-LCD field often relies on paired datasets (an image with Moiré pattern, paired with an image without Moiré) to train the model, which are difficult to obtain in practical manufacturing settings. This study propose a novel method to address the issue of Moiré pattern removal in TFT-LCD panels without the need for paired data. Our approach leverages a Self-Consistent Generative Adversarial Network (SC-GAN) with U-Net as generator to achieve Moiré pattern removal without paired data. Specifically, our SC-GAN model employs an adversarial training framework and incorporates various loss functions to guide the training process. The peak-signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) were used to compare with our proposed model and other classical supervised denoising models. The experiments show that our proposed model can effectively remove the Moire pattern from defect images.