letuano5/noise-detector — reverse-engineered prompt

Build me a Python project that recreates the noise detector from the ICIP 2002 paper “An Accurate Noise Detector for Image Restoration”.

I want it to work on 8 bit grayscale images with salt and pepper noise. It should let me run a single example, run a paper style reproduction, benchmark different noise levels, and tune parameters from YAML config files. It should include the PSM baseline, the proposed two stage detector with edge flag and verification logic, median based restoration, and an iterative median filter baseline.

Please make the outputs easy to inspect. For each run, save the resolved config, metrics as CSV and JSON, the original and noisy images, detected masks, restored images, and a comparison grid. Include common metrics like undetection, misdetection, precision, recall, f1, accuracy, SNR, PSNR, SSIM, and MSE.

Use built in skimage sample images by default, but also support local image paths. Add tests, clean command line entry points, and keep the code easy to extend with new detectors or metrics.