High Accuracy Swin Transformers for Imagebased Wafer Map Defect Detection

Abstract

A wafer map depicts the location of each die on the wafer and indicates whether it is a Product, Secondary Silicon, or Reject. Detecting defects in Wafer Maps is crucial in order to ensure the integrity of the chips processed in the wafer, as any defect can cause anomalies thus decreasing the overall yield. With the current advances in anomaly detection using various Computer Vision Techniques, Transformer Architecture based Vision models are a prime candidate for identifying wafer defects. In this paper, the performance of Four such Transformer based models -BEiT (BERT Pre-Training of Image Transformers), FNet (Fourier Network), ViT (Vision Transformer) and Swin Transformer (Shifted Window based Transformer) in wafer map defect classification are discussed. Each of these models were individually trained, tested and evaluated with the "MixedWM38" dataset obtained from the online platform, Kaggle. During evaluation, it has been observed that the overall accuracy of the Swin Transformer Network algorithm is the highest, at 97.47%, followed closely by Vision Transformer at 96.77%. The average Recall of Swin Transformer is also 97.54%, which indicates an extremely low encounter of false negatives (24600 ppm) in contrast to true positives, making it less likely to expose defective products in the market.

AFCN: An attention‐directed feature‐fusion ConvNeXt network for low‐voltage apparatus assembly quality inspection

“…The comparative experiments include an in‐depth analysis of multiple networks, such as ResNet101 [30], DenseNet [39], EfficientNet‐V2 [40], Vision Transformer [25], MobileVit [41], Swin Transformer [27], and NoisyViT [42]. These networks or their variants achieved superior performance in many scenarios [24, 26, 28]. Among them, ResNet101, DenseNet, and EfficientNet‐V2 are typical convolutional neural networks; Vision Transformer, Mobile ViT, and Swin Transformer are typical transformer networks; and NoisyViT is a newly proposed network with satisfactory results in classification tasks.…”

“…Microsoft Research Asia proposed Swin Transformer [27], which achieved outstanding performance on various public datasets. Nafi et al [28] used Swin Transformer for quality inspection of wafer maps and obtained better performance than traditional CNNs. Tang et al [29] proposed a method for steel plate surface defects based on Swin Transformer, which could generate defect category information and its precise location.…”

“…The comparative experiments include an in-depth analysis of multiple networks, such as ResNet101 [30], DenseNet [39], EfficientNet-V2 [40], Vision Transformer [25], MobileVit [41], Swin Transformer [27], and NoisyViT [42]. These networks or their variants achieved superior performance in many scenarios [24,26,28] best accuracy. Therefore, the optimal training strategies of different networks are used to train the models to obtain their best indicators.…”

“…Nafi et al. [28] used Swin Transformer for quality inspection of wafer maps and obtained better performance than traditional CNNs. Tang et al.…”