SEMNR: Why I Stopped Trusting "Clean" Images and Treated Metrics as Guardrails

The article discusses the author's experience in building a semiconductor SEM denoising model, where they realized that standard evaluation metrics like PSNR can be misleading and encourage over-smoothing, leading to the loss of critical details. The author developed a metric profile to balance different aspects of image quality, including fidelity, structure, edges, and texture.

Details

The author worked on a semiconductor SEM denoising project as part of an Applied Materials & Extra-Tech bootcamp. They realized that in scientific and industrial imaging, the goal is not just to remove noise, but to preserve critical evidence like microscopic edges, surface textures, and defects. Standard evaluation metrics like PSNR, which reward smoothness, can lead to models that blur these important details. The author built a 'SEMNR' model and developed a metric profile that included PSNR, SSIM, FSIM, DISTS, and CNR, each with a specific job description. They learned to read disagreements between these metrics as debugging signals, such as over-smoothing (PSNR up, FSIM down) or texture drift (SSIM stable, DISTS down). The goal was to maximize the overall 'area' of the metric profile, not just chase a single high score. This approach helped the author build a model that preserves analytical utility, even if it doesn't look as 'clean' to the human eye.