Image denoising algorithm based on double branching for critical dimension scanning electron microscopy

BackgroundAs high-numerical-aperture extreme ultraviolet lithography technology continues to advance, photoresist thinning has become essential for meeting stringent technical requirements. However, the shrinkage observed in thin photoresists during scanning electron microscopy (SEM) measurements presents a significant challenge to measurement accuracy.AimTo mitigate the effects of photoresist shrinkage, SEM images are typically acquired using lower voltages, reduced currents, and fewer scanning frames. However, these adjustments introduce additional noise and diminish measurement accuracy. To mitigate these limitations, a double-branch neural network is proposed.ApproachThe neural network consists of three primary components: an edge enhancement branch, a denoising branch, and a feature fusion module. The denoising branch employs Swin-Conv. blocks to extract both local and global features for noise removal, whereas the edge enhancement branch applies Sobel convolution to capture edge features. The feature fusion module integrates features from both branches to enhance denoising performance.ResultsCompared with traditional image denoising algorithms, this method effectively mitigates edge blurring caused by excessive smoothing. In addition, it preserves more edge details than other deep learning–based denoising algorithms.ConclusionsResults from practical application experiments indicate that the algorithm indirectly improves the measurement accuracy of critical dimension scanning electron microscopy.