setiastrosuitepro 1.6.7__py3-none-any.whl

setiastro/saspro/autostretch.py

@@ -0,0 +1,198 @@
+# pro/autostretch.py
+import numpy as np
+
+_MAX_STATS_PIXELS = 1_000_000
+_DEFAULT_SIGMA = 3
+_U8_MAX  = 4095  # 12-bit output for better gradations than 255
+_U16_MAX = 65535
+
+# ---------- helpers (generic N-level pipeline) ----------
+def _to_uN(a: np.ndarray, maxv: int) -> np.ndarray:
+    """Convert to uint8/uint16 [0..maxv] for cheap hist/LUT work."""
+    tgt_dtype = np.uint16 if maxv > _U8_MAX else np.uint8
+    if a.dtype == tgt_dtype:
+        return a
+    if np.issubdtype(a.dtype, np.integer):
+        info = np.iinfo(a.dtype)
+        if info.max <= 0:
+            return np.zeros_like(a, dtype=tgt_dtype)
+        scaled = np.clip(a.astype(np.float32), 0, info.max) * (maxv / float(info.max))
+        return (scaled + 0.5).astype(tgt_dtype)
+    # float-ish
+    af = np.clip(a.astype(np.float32), 0.0, 1.0)
+    return (af * maxv + 0.5).astype(tgt_dtype)
+
+def _choose_stride(h: int, w: int, max_pixels: int) -> tuple[int, int]:
+    n = h * w
+    if n <= max_pixels:
+        return 1, 1
+    s = max(1, int(np.sqrt(n / float(max_pixels))))
+    return s, s
+
+def _compute_lut_from_sample(sample_uN: np.ndarray, target: float, sigma: float, maxv: int,
+                             qfloor: float = 0.001) -> np.ndarray:
+    """Return a 0..maxv -> [0..1] LUT using the same math as _fast_channel_autostretch_uN."""
+    hist = np.bincount(sample_uN.ravel(), minlength=maxv + 1)
+    total = int(hist.sum()) or 1
+
+    cdf = np.cumsum(hist)
+    med = int(np.searchsorted(cdf, (total + 1) // 2))
+
+    bins      = np.arange(maxv + 1, dtype=np.float64)
+    hist_low  = hist[:med + 1]
+    total_low = int(hist_low.sum()) or 1
+    mean_low  = float((hist_low * bins[:med + 1]).sum() / total_low)
+    var_low   = float((hist_low * (bins[:med + 1] - mean_low)**2).sum() / total_low)
+    std_low   = float(np.sqrt(max(1e-12, var_low)))
+
+    floor_idx = int(np.searchsorted(cdf, int(qfloor * total)))
+    bp = int(max(floor_idx, med - sigma * std_low))
+    bp = int(np.clip(bp, 0, maxv - 1))
+
+    return _build_lut_generic(bp, target, med, maxv)
+
+def _stats_from_hist_generic(uN: np.ndarray, maxv: int) -> tuple[float, float, int, int]:
+    """(median, std, minv, maxv) using a histogram of levels [0..maxv]."""
+    hist = np.bincount(uN.ravel(), minlength=maxv + 1)
+    total = int(hist.sum())
+    if total == 0:
+        return 0.0, 1.0, 0, 0
+
+    # min / max (first/last nonzero bin)
+    minv = int(np.argmax(hist > 0))
+    maxv_found = int(maxv - np.argmax(hist[::-1] > 0))
+
+    cdf = np.cumsum(hist)
+    med_idx = int(np.searchsorted(cdf, (total + 1) // 2))
+
+    bins = np.arange(hist.size, dtype=np.float64)
+    s1 = float((hist * bins).sum())
+    s2 = float((hist * (bins * bins)).sum())
+    mean = s1 / total
+    var = max(0.0, s2 / total - mean * mean)
+    std = float(np.sqrt(var))
+    return float(med_idx), std, minv, maxv_found
+
+def _build_lut_generic(bp: int, target_median: float, med_uN: float, maxv: int) -> np.ndarray:
+    denom_bp = max(1, maxv - int(bp))
+    median_rescaled = (med_uN - bp) / float(denom_bp)
+    median_rescaled = float(np.clip(median_rescaled, 1e-9, 1.0))
+
+    x = np.arange(maxv + 1, dtype=np.float32)
+
+    # ✅ KEY FIX: clamp r to [0,1] so x<=bp -> r=0 (maps to 0), x>=max -> r=1
+    r = np.clip((x - bp) / float(denom_bp), 0.0, 1.0)
+
+    denom = median_rescaled * (target_median + r - 1.0) - target_median * r
+    denom = np.where(np.abs(denom) < 1e-12, 1e-12, denom)
+    out = ((median_rescaled - 1.0) * target_median * r) / denom
+    return np.clip(out, 0.0, 1.0).astype(np.float32)
+
+def _fast_channel_autostretch_uN(ch_uN: np.ndarray, target: float, sigma: float, maxv: int,
+                                 qfloor: float = 0.001) -> np.ndarray:
+    """
+    qfloor: low-percentile floor (e.g. 0.1%) so BP doesn't peg to minv=0.
+    """
+    # Subsample
+    h, w = ch_uN.shape
+    sy, sx = _choose_stride(h, w, _MAX_STATS_PIXELS)
+    sample = ch_uN[::sy, ::sx]
+
+    # Histogram on the sample
+    hist = np.bincount(sample.ravel(), minlength=maxv + 1)
+    total = int(hist.sum()) or 1
+
+    # CDF + median index
+    cdf = np.cumsum(hist)
+    med = int(np.searchsorted(cdf, (total + 1)//2))
+
+    # Robust std: only use bins <= median (avoids bright-tail inflation)
+    bins      = np.arange(maxv + 1, dtype=np.float64)
+    hist_low  = hist[:med + 1]
+    total_low = int(hist_low.sum()) or 1
+    mean_low  = float((hist_low * bins[:med + 1]).sum() / total_low)
+    var_low   = float((hist_low * (bins[:med + 1] - mean_low)**2).sum() / total_low)
+    std_low   = float(np.sqrt(max(1e-12, var_low)))
+
+    # Percentile floor (avoid pegging to 0)
+    floor_idx = int(np.searchsorted(cdf, int(qfloor * total)))
+
+    # Black point driven by sigma, but never below the floor
+    bp = int(max(floor_idx, med - sigma * std_low))
+    bp = int(np.clip(bp, 0, maxv - 1))
+
+    # Build LUT and map
+    lut = _build_lut_generic(bp, target, med, maxv)
+    return lut[ch_uN]
+
+# ---------- public API ----------
+def autostretch(
+    img: np.ndarray,
+    target_median: float = 0.25,
+    linked: bool = False,
+    sigma: float = _DEFAULT_SIGMA,
+    *,
+    use_16bit: bool | None = None,
+) -> np.ndarray:
+    """
+    High-quality autostretch that can operate in 16-bit (HQ, default) or 8-bit (fast) mode.
+
+    • 16-bit mode: smooth gradients, minimal posterization (recommended).
+    • 8-bit mode: slightly faster on very large images, lower fidelity.
+
+    If use_16bit is None, we try to read QSettings("display/autostretch_16bit") and
+    default to True on failure (no Qt in context).
+    """
+    if img is None:
+        return None
+
+    # Optional auto-read from QSettings if caller didn’t pass a flag.
+    if use_16bit is None:
+        try:
+            from PyQt6.QtCore import QSettings
+            use_16bit = QSettings().value("display/autostretch_16bit", True, type=bool)
+        except Exception:
+            use_16bit = True
+
+    maxv = _U16_MAX if use_16bit else _U8_MAX
+    a = np.asarray(img)
+
+    # MONO (or pseudo-mono)
+    if a.ndim == 2 or (a.ndim == 3 and a.shape[2] == 1):
+        u = _to_uN(a.squeeze(), maxv)
+        out = _fast_channel_autostretch_uN(u, target_median, sigma, maxv)
+        return out.astype(np.float32, copy=False)
+
+    # color
+    u = _to_uN(a, maxv)
+    C = u.shape[2]
+
+    if linked:
+        # sample fewer pixels for stats
+        h, w, _ = u.shape
+        sy, sx = _choose_stride(h, w, max(1, _MAX_STATS_PIXELS // 3))
+        sample = u[::sy, ::sx]
+
+        # Rec.709-ish luminance, integer dtype preserved
+        # (weights sum to 1; cast back to u.dtype for the LUT builder)
+        lum = (0.2126 * sample[..., 0] + 0.7152 * sample[..., 1] + 0.0722 * sample[..., 2]).astype(u.dtype)
+
+        lut = _compute_lut_from_sample(lum, target_median, sigma, maxv)
+
+        out = np.empty_like(u, dtype=np.float32)
+        # Vectorized LUT application: apply to all RGB channels at once
+        # lut is 1D array of float32; u[..., :3] selects RGB indices
+        out[..., :3] = lut[u[..., :3]]       
+        
+        if C > 3:  # pass-through non-RGB channels
+            out[..., 3:] = u[..., 3:] / float(maxv)
+        return out
+    else:
+        out = np.empty_like(u, dtype=np.float32)
+        for c in range(min(3, C)):
+            out[..., c] = _fast_channel_autostretch_uN(u[..., c], target_median, sigma, maxv)
+        if C > 3:
+            out[..., 3:] = u[..., 3:] / float(maxv)
+        return out
+    
+    print(f"med={med} std_low={std_low:.2f} floor={floor_idx} σ={sigma} → bp={bp}")