Frequency-Specific Cross-Entropy Super-Resolution: Addressing High-Frequency Details and Reducing Artifacts

초록

Improving high-frequency details in super resolution has recently become a significant challenge. Traditional super resolution is typically approached as a regression task using L2 loss, which results in blurry images. To address this issue, we design cross-entropy super-resolution (CS) by transforming it into a classification task and applying cross entropy (CE) loss. CE loss is suitable for high frequencies by employing the one-hot encoding for the predicted probabilities. While CS enhances sharpness and details, unexpected artifacts may also appear in some regions. Therefore, since L2 loss focuses more on large signals like low frequencies, we propose frequency-specific cross-entropy super-resolution (FCS) to improve performance by training low frequencies with L2 loss and high frequencies with CE loss. We evaluate our method using the learned perceptual image patch similarity (LPIPS) metric, which is robust to distortions and correlates well with human perception. In all tested datasets, our FCS improves LPIPS by approximately 1.7 times compared to baseline.

키워드