LEGENDFTW/image-filtering-explorer

1"""metrics.py — Image quality and noise assessment."""2 3import numpy as np4import cv25from skimage.metrics import structural_similarity as ssim_fn6from skimage.metrics import peak_signal_noise_ratio as psnr_fn7 8 9def compute_metrics(original: np.ndarray, filtered: np.ndarray) -> dict:10    """11    Compute image quality metrics between the original (noisy) input12    and the filtered output.13    """14    orig = original.astype(np.float32)15    filt = filtered.astype(np.float32)16 17    # PSNR — higher is better (more similar images)18    # We clamp to avoid log(0); identical images → inf, we cap at 60 dB19    mse = np.mean((orig - filt) ** 2)20    if mse == 0:21        psnr = 60.022    else:23        psnr = float(psnr_fn(original, filtered, data_range=255))24        psnr = min(psnr, 60.0)25 26    # SSIM — structural similarity, 0–127    ssim_val = float(28        ssim_fn(original, filtered, data_range=255, channel_axis=-1)29    )30 31    # Mean absolute difference32    mean_diff = float(np.mean(np.abs(orig - filt)))33 34    # Noise reduction estimate: compare std-dev of high-frequency residual35    # (Laplacian response) before and after36    lp_in  = _laplacian_std(original)37    lp_out = _laplacian_std(filtered)38    if lp_in > 0:39        noise_reduction = max(0.0, (lp_in - lp_out) / lp_in * 100)40    else:41        noise_reduction = 0.042 43    return {44        "psnr": psnr,45        "ssim": ssim_val,46        "mean_diff": mean_diff,47        "noise_reduction": noise_reduction,48        "mse": mse,49    }50 51 52def compute_noise_profile(image: np.ndarray) -> np.ndarray:53    """54    Return a 2-D map of estimated local noise magnitude via55    Laplacian high-frequency response (grayscale, float32).56    """57    gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)58    lap  = cv2.Laplacian(gray.astype(np.float32), cv2.CV_32F)59    # Smooth the absolute response for a "heatmap" look60    profile = cv2.GaussianBlur(np.abs(lap), (15, 15), 5)61    return profile62 63 64def _laplacian_std(image: np.ndarray) -> float:65    """Standard deviation of the Laplacian — proxy for noise level."""66    gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)67    lap  = cv2.Laplacian(gray.astype(np.float64), cv2.CV_64F)68    return float(lap.std())69