Merge pull request 'feat(perception): terrain roughness estimator via Gabor + LBP (Issue #296)' (#305) from sl-perception/issue-296-terrain-rough into main

jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/_terrain_roughness.py

@@ -0,0 +1,219 @@
+"""
+_terrain_roughness.py — Terrain roughness estimation via Gabor energy + LBP variance
+                         (no ROS2 deps).
+
+Algorithm
+---------
+Both features are computed on the greyscale crop of the bottom *roi_frac* of a BGR
+D435i colour frame (the floor region).
+
+1. Gabor filter-bank energy (4 orientations × 2 spatial wavelengths, quadrature pairs):
+
+       E = mean over all (θ, λ) of  mean_pixels( |real_resp|² + |imag_resp|² )
+
+   A quadrature pair uses phase ψ=0 (real) and ψ=π/2 (imaginary), giving orientation-
+   and phase-invariant energy.  The mean across all filter pairs gives a single energy
+   scalar that is high for textured/edgy surfaces and near-zero for smooth ones.
+
+2. LBP variance (Local Binary Pattern, radius=1, 8-point, pure NumPy, no sklearn):
+
+       LBP(x,y) = Σ_k  s(g_k − g_c) · 2^k      (k = 0..7, g_c = centre pixel)
+
+   Computed as vectorised slice comparisons over the 8 cardinal/diagonal neighbours.
+   The metric is var(LBP image).  A constant surface gives LBP ≡ 0xFF (all bits set,
+   since every neighbour ties with the centre) → variance = 0.  Irregular textures
+   produce a dispersed LBP histogram → high variance.
+
+3. Normalised roughness blend:
+
+       roughness = clip( _W_GABOR · E / _GABOR_REF  +  _W_LBP · V / _LBP_REF ,  0, 1 )
+
+   Calibration: _GABOR_REF is tuned so that a coarse gravel / noise image maps to ~1.0;
+   _LBP_REF = 5000 (≈ variance of a fully random 8-bit LBP image).
+
+Public API
+----------
+  RoughnessResult(roughness, gabor_energy, lbp_variance)
+  estimate_roughness(bgr, roi_frac=0.40) -> RoughnessResult
+"""
+
+from __future__ import annotations
+
+from typing import NamedTuple, Optional
+
+import numpy as np
+
+
+# ── Tuning constants ──────────────────────────────────────────────────────────
+
+# Gabor filter-bank parameters
+_GABOR_N_ORIENTATIONS = 4          # θ ∈ {0°, 45°, 90°, 135°}
+_GABOR_WAVELENGTHS    = [5.0, 10.0]  # spatial wavelengths in pixels (medium + coarse)
+_GABOR_KSIZE          = 11           # kernel side length
+_GABOR_SIGMA          = 3.0          # Gaussian envelope std (pixels)
+_GABOR_GAMMA          = 0.5          # spatial aspect ratio
+
+# Normalization references
+_GABOR_REF = 500.0    # Gabor mean power at which we consider maximum roughness
+_LBP_REF   = 5000.0  # LBP variance at which we consider maximum roughness
+
+# Blend weights (must sum to 1)
+_W_GABOR = 0.5
+_W_LBP   = 0.5
+
+# Module-level cache for the kernel bank (built once per process)
+_gabor_kernels: Optional[list] = None
+
+
+# ── Data types ────────────────────────────────────────────────────────────────
+
+class RoughnessResult(NamedTuple):
+    """Terrain roughness estimate for a single frame."""
+    roughness:    float   # blended score in [0, 1]; 0=smooth, 1=rough
+    gabor_energy: float   # raw Gabor mean energy (≥ 0)
+    lbp_variance: float   # raw LBP image variance (≥ 0)
+
+
+# ── Internal helpers ──────────────────────────────────────────────────────────
+
+def _build_gabor_kernels() -> list:
+    """Build (real_kernel, imag_kernel) pairs for all (orientation, wavelength) combos."""
+    import cv2
+
+    kernels = []
+    for i in range(_GABOR_N_ORIENTATIONS):
+        theta = float(i) * np.pi / _GABOR_N_ORIENTATIONS
+        for lam in _GABOR_WAVELENGTHS:
+            real_k = cv2.getGaborKernel(
+                (_GABOR_KSIZE, _GABOR_KSIZE),
+                _GABOR_SIGMA, theta, lam, _GABOR_GAMMA, 0.0,
+                ktype=cv2.CV_64F,
+            ).astype(np.float32)
+            imag_k = cv2.getGaborKernel(
+                (_GABOR_KSIZE, _GABOR_KSIZE),
+                _GABOR_SIGMA, theta, lam, _GABOR_GAMMA, np.pi / 2.0,
+                ktype=cv2.CV_64F,
+            ).astype(np.float32)
+            kernels.append((real_k, imag_k))
+    return kernels
+
+
+def _get_gabor_kernels() -> list:
+    global _gabor_kernels
+    if _gabor_kernels is None:
+        _gabor_kernels = _build_gabor_kernels()
+    return _gabor_kernels
+
+
+def gabor_energy(grey: np.ndarray) -> float:
+    """
+    Mean quadrature Gabor energy across the filter bank.
+
+    Parameters
+    ----------
+    grey : (H, W) uint8 or float ndarray
+
+    Returns
+    -------
+    float — mean power (≥ 0); higher for textured surfaces
+    """
+    import cv2
+
+    img = np.asarray(grey, dtype=np.float32)
+    if img.size == 0:
+        return 0.0
+
+    # Subtract the mean so constant regions (DC) contribute zero energy.
+    # This is standard in Gabor texture analysis and ensures that a flat
+    # uniform surface always returns energy ≈ 0.
+    img = img - float(img.mean())
+
+    kernels = _get_gabor_kernels()
+    total   = 0.0
+    for real_k, imag_k in kernels:
+        r = cv2.filter2D(img, cv2.CV_32F, real_k)
+        i = cv2.filter2D(img, cv2.CV_32F, imag_k)
+        total += float(np.mean(r * r + i * i))
+
+    return total / len(kernels) if kernels else 0.0
+
+
+def lbp_variance(grey: np.ndarray) -> float:
+    """
+    Variance of the 8-point radius-1 LBP image (pure NumPy, no sklearn).
+
+    Parameters
+    ----------
+    grey : (H, W) uint8 ndarray
+
+    Returns
+    -------
+    float — variance of LBP values; 0 for flat surfaces, high for irregular textures
+    """
+    g = np.asarray(grey, dtype=np.int32)
+    h, w = g.shape
+    if h < 3 or w < 3:
+        return 0.0
+
+    # Centre pixel patch (interior only, 1-px border excluded)
+    c = g[1:h-1, 1:w-1]
+
+    # 8-neighbourhood comparisons using array slices (vectorised, no loops)
+    lbp = (
+        ((g[0:h-2, 0:w-2] >= c).view(np.uint8)      ) |
+        ((g[0:h-2, 1:w-1] >= c).view(np.uint8) << 1 ) |
+        ((g[0:h-2, 2:w  ] >= c).view(np.uint8) << 2 ) |
+        ((g[1:h-1, 0:w-2] >= c).view(np.uint8) << 3 ) |
+        ((g[1:h-1, 2:w  ] >= c).view(np.uint8) << 4 ) |
+        ((g[2:h  , 0:w-2] >= c).view(np.uint8) << 5 ) |
+        ((g[2:h  , 1:w-1] >= c).view(np.uint8) << 6 ) |
+        ((g[2:h  , 2:w  ] >= c).view(np.uint8) << 7 )
+    )
+
+    return float(lbp.var())
+
+
+# ── Public API ────────────────────────────────────────────────────────────────
+
+def estimate_roughness(
+    bgr:      np.ndarray,
+    roi_frac: float = 0.40,
+) -> RoughnessResult:
+    """
+    Estimate terrain roughness from the bottom roi_frac of a BGR image.
+
+    Parameters
+    ----------
+    bgr      : (H, W, 3) uint8 BGR ndarray (or greyscale (H, W))
+    roi_frac : fraction of image height to use as the floor ROI (from bottom)
+
+    Returns
+    -------
+    RoughnessResult(roughness, gabor_energy, lbp_variance)
+    roughness is in [0, 1] where 0=smooth, 1=rough.
+    """
+    import cv2
+
+    # Crop to bottom roi_frac
+    h = bgr.shape[0]
+    r0 = int(h * max(0.0, 1.0 - min(float(roi_frac), 1.0)))
+    roi = bgr[r0:, :]
+
+    if roi.size == 0:
+        return RoughnessResult(roughness=0.0, gabor_energy=0.0, lbp_variance=0.0)
+
+    # Convert to greyscale
+    if roi.ndim == 3:
+        grey = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
+    else:
+        grey = np.asarray(roi, dtype=np.uint8)
+
+    ge = gabor_energy(grey)
+    lv = lbp_variance(grey)
+
+    gabor_norm = min(ge / _GABOR_REF, 1.0)
+    lbp_norm   = min(lv / _LBP_REF,   1.0)
+
+    roughness = float(np.clip(_W_GABOR * gabor_norm + _W_LBP * lbp_norm, 0.0, 1.0))
+
+    return RoughnessResult(roughness=roughness, gabor_energy=ge, lbp_variance=lv)

jetson/ros2_ws/src/saltybot_bringup/saltybot_bringup/terrain_rough_node.py

@@ -0,0 +1,102 @@
+"""
+terrain_rough_node.py — D435i terrain roughness estimator (Issue #296).
+
+Subscribes to the RealSense colour stream, computes Gabor filter texture
+energy + LBP variance on the bottom-40% ROI (floor region), and publishes
+a normalised roughness score at 2 Hz.
+
+Intended use: speed adaptation — reduce maximum velocity when roughness is high.
+
+Subscribes (BEST_EFFORT):
+  /camera/color/image_raw     sensor_msgs/Image   BGR8
+
+Publishes:
+  /saltybot/terrain_roughness   std_msgs/Float32   roughness in [0,1] (0=smooth, 1=rough)
+
+Parameters
+----------
+roi_frac    float   0.40    Fraction of image height used as floor ROI (from bottom)
+publish_hz  float   2.0     Publication rate (Hz)
+"""
+
+from __future__ import annotations
+
+import rclpy
+from rclpy.node import Node
+from rclpy.qos import QoSProfile, ReliabilityPolicy, HistoryPolicy
+
+from cv_bridge import CvBridge
+
+from sensor_msgs.msg import Image
+from std_msgs.msg import Float32
+
+from ._terrain_roughness import estimate_roughness
+
+
+_SENSOR_QOS = QoSProfile(
+    reliability=ReliabilityPolicy.BEST_EFFORT,
+    history=HistoryPolicy.KEEP_LAST,
+    depth=4,
+)
+
+
+class TerrainRoughNode(Node):
+
+    def __init__(self) -> None:
+        super().__init__('terrain_rough_node')
+
+        self.declare_parameter('roi_frac',   0.40)
+        self.declare_parameter('publish_hz', 2.0)
+
+        self._roi_frac   = float(self.get_parameter('roi_frac').value)
+        publish_hz       = float(self.get_parameter('publish_hz').value)
+
+        self._bridge      = CvBridge()
+        self._latest_bgr  = None   # updated by subscription callback
+
+        self._sub = self.create_subscription(
+            Image, '/camera/color/image_raw', self._on_image, _SENSOR_QOS)
+        self._pub = self.create_publisher(
+            Float32, '/saltybot/terrain_roughness', 10)
+
+        period = 1.0 / max(publish_hz, 0.1)
+        self._timer = self.create_timer(period, self._on_timer)
+
+        self.get_logger().info(
+            f'terrain_rough_node ready — roi_frac={self._roi_frac} '
+            f'publish_hz={publish_hz}'
+        )
+
+    # ── Callbacks ─────────────────────────────────────────────────────────────
+
+    def _on_image(self, msg: Image) -> None:
+        try:
+            self._latest_bgr = self._bridge.imgmsg_to_cv2(
+                msg, desired_encoding='bgr8')
+        except Exception as exc:
+            self.get_logger().error(
+                f'cv_bridge: {exc}', throttle_duration_sec=5.0)
+
+    def _on_timer(self) -> None:
+        if self._latest_bgr is None:
+            return
+
+        result = estimate_roughness(self._latest_bgr, roi_frac=self._roi_frac)
+
+        msg = Float32()
+        msg.data = result.roughness
+        self._pub.publish(msg)
+
+
+def main(args=None) -> None:
+    rclpy.init(args=args)
+    node = TerrainRoughNode()
+    try:
+        rclpy.spin(node)
+    finally:
+        node.destroy_node()
+        rclpy.shutdown()
+
+
+if __name__ == '__main__':
+    main()

jetson/ros2_ws/src/saltybot_bringup/setup.py

@@ -43,6 +43,8 @@ setup(
             'color_segmenter         = saltybot_bringup.color_segment_node:main',
             # Motion blur detector (Issue #286)
             'blur_detector           = saltybot_bringup.blur_detect_node:main',
+            # Terrain roughness estimator (Issue #296)
+            'terrain_roughness       = saltybot_bringup.terrain_rough_node:main',
         ],
     },
 )

jetson/ros2_ws/src/saltybot_bringup/test/test_terrain_roughness.py

@@ -0,0 +1,321 @@
+"""
+test_terrain_roughness.py — Unit tests for terrain roughness helpers (no ROS2 required).
+
+Covers:
+  RoughnessResult:
+    - fields accessible by name
+    - roughness in [0, 1]
+
+  lbp_variance:
+    - returns float
+    - solid image returns 0.0
+    - noisy image returns positive value
+    - noisy > solid (strictly)
+    - small image (< 3×3) returns 0.0
+    - output is non-negative
+
+  gabor_energy:
+    - returns float
+    - solid image returns near-zero energy
+    - noisy image returns positive energy
+    - noisy > solid (strictly)
+    - empty image returns 0.0
+    - output is non-negative
+
+  estimate_roughness — output contract:
+    - returns RoughnessResult
+    - roughness in [0, 1] for all test images
+    - roughness == 0.0 for solid (constant) image (both metrics are 0)
+    - roughness > 0.0 for random noise image
+    - roi_frac=1.0 uses entire frame
+    - roi_frac=0.0 returns zero roughness (empty ROI)
+    - gabor_energy field matches standalone gabor_energy() on the ROI
+    - lbp_variance field matches standalone lbp_variance() on the ROI
+
+  estimate_roughness — ordering:
+    - random noise roughness > solid roughness (strict)
+    - checkerboard roughness > smooth gradient roughness
+    - rough texture > smooth texture
+
+  estimate_roughness — ROI:
+    - roughness changes when roi_frac changes (textured bottom, smooth top)
+    - bottom-only ROI gives higher roughness than top-only on bottom-textured image
+"""
+
+import sys
+import os
+
+import numpy as np
+import pytest
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+
+from saltybot_bringup._terrain_roughness import (
+    RoughnessResult,
+    gabor_energy,
+    lbp_variance,
+    estimate_roughness,
+)
+
+
+# ── Image factories ───────────────────────────────────────────────────────────
+
+def _solid(val=128, h=64, w=64) -> np.ndarray:
+    return np.full((h, w, 3), val, dtype=np.uint8)
+
+
+def _grey_solid(val=128, h=64, w=64) -> np.ndarray:
+    return np.full((h, w), val, dtype=np.uint8)
+
+
+def _noise(h=64, w=64, seed=42) -> np.ndarray:
+    rng = np.random.default_rng(seed)
+    data = rng.integers(0, 256, (h, w, 3), dtype=np.uint8)
+    return data
+
+
+def _grey_noise(h=64, w=64, seed=42) -> np.ndarray:
+    rng = np.random.default_rng(seed)
+    return rng.integers(0, 256, (h, w), dtype=np.uint8)
+
+
+def _checkerboard(h=64, w=64, tile=8) -> np.ndarray:
+    grey = np.zeros((h, w), dtype=np.uint8)
+    for r in range(h):
+        for c in range(w):
+            if ((r // tile) + (c // tile)) % 2 == 0:
+                grey[r, c] = 255
+    return np.stack([grey, grey, grey], axis=-1)
+
+
+def _gradient(h=64, w=64) -> np.ndarray:
+    """Smooth horizontal linear gradient."""
+    row = np.linspace(0, 255, w, dtype=np.uint8)
+    grey = np.tile(row, (h, 1))
+    return np.stack([grey, grey, grey], axis=-1)
+
+
+# ── RoughnessResult ───────────────────────────────────────────────────────────
+
+class TestRoughnessResult:
+
+    def test_fields_accessible(self):
+        r = RoughnessResult(roughness=0.5, gabor_energy=100.0, lbp_variance=200.0)
+        assert r.roughness    == pytest.approx(0.5)
+        assert r.gabor_energy == pytest.approx(100.0)
+        assert r.lbp_variance == pytest.approx(200.0)
+
+    def test_roughness_in_range(self):
+        for v in [0.0, 0.5, 1.0]:
+            r = RoughnessResult(roughness=v, gabor_energy=0.0, lbp_variance=0.0)
+            assert 0.0 <= r.roughness <= 1.0
+
+
+# ── lbp_variance ─────────────────────────────────────────────────────────────
+
+class TestLbpVariance:
+
+    def test_returns_float(self):
+        assert isinstance(lbp_variance(_grey_solid()), float)
+
+    def test_non_negative(self):
+        for img in [_grey_solid(), _grey_noise(), _grey_solid(0)]:
+            assert lbp_variance(img) >= 0.0
+
+    def test_solid_returns_zero(self):
+        """Constant image: all neighbours equal centre → all LBP bits set → zero variance."""
+        assert lbp_variance(_grey_solid(128)) == pytest.approx(0.0)
+
+    def test_solid_black_returns_zero(self):
+        assert lbp_variance(_grey_solid(0)) == pytest.approx(0.0)
+
+    def test_solid_white_returns_zero(self):
+        assert lbp_variance(_grey_solid(255)) == pytest.approx(0.0)
+
+    def test_noise_returns_positive(self):
+        assert lbp_variance(_grey_noise()) > 0.0
+
+    def test_noise_greater_than_solid(self):
+        assert lbp_variance(_grey_noise()) > lbp_variance(_grey_solid())
+
+    def test_small_image_returns_zero(self):
+        """Image smaller than 3×3 has no interior pixels."""
+        assert lbp_variance(np.zeros((2, 2), dtype=np.uint8)) == pytest.approx(0.0)
+
+    def test_checkerboard_high_variance(self):
+        """Alternating 0/255 checkerboard → LBP alternates 0x00/0xFF → very high variance."""
+        v = lbp_variance(np.tile(
+            np.array([[0, 255], [255, 0]], dtype=np.uint8), (16, 16)))
+        assert v > 1000.0, f'checkerboard LBP variance too low: {v:.1f}'
+
+    def test_noise_higher_than_tiled_gradient(self):
+        """Random 2D noise has varying neighbourhood patterns → higher LBP variance
+        than a horizontal gradient tiled into rows (which has a constant pattern)."""
+        assert lbp_variance(_grey_noise()) > lbp_variance(_grey_solid())
+
+
+# ── gabor_energy ──────────────────────────────────────────────────────────────
+
+class TestGaborEnergy:
+
+    def test_returns_float(self):
+        assert isinstance(gabor_energy(_grey_solid()), float)
+
+    def test_non_negative(self):
+        for img in [_grey_solid(), _grey_noise(), _grey_solid(0)]:
+            assert gabor_energy(img) >= 0.0
+
+    def test_solid_near_zero(self):
+        """Constant image has zero Gabor response after DC subtraction."""
+        e = gabor_energy(_grey_solid(128))
+        assert e < 1.0, f'solid image gabor energy should be ~0 after DC removal, got {e:.4f}'
+
+    def test_noise_positive(self):
+        assert gabor_energy(_grey_noise()) > 0.0
+
+    def test_noise_greater_than_solid(self):
+        assert gabor_energy(_grey_noise()) > gabor_energy(_grey_solid())
+
+    def test_empty_image_returns_zero(self):
+        assert gabor_energy(np.zeros((0, 0), dtype=np.uint8)) == pytest.approx(0.0)
+
+    def test_noise_higher_than_solid(self):
+        """Random noise has non-zero energy at our Gabor filter frequencies; solid has none."""
+        assert gabor_energy(_grey_noise()) > gabor_energy(_grey_solid())
+
+
+# ── estimate_roughness — output contract ──────────────────────────────────────
+
+class TestEstimateRoughnessContract:
+
+    def test_returns_roughness_result(self):
+        r = estimate_roughness(_solid())
+        assert isinstance(r, RoughnessResult)
+
+    def test_roughness_in_range_solid(self):
+        r = estimate_roughness(_solid())
+        assert 0.0 <= r.roughness <= 1.0
+
+    def test_roughness_in_range_noise(self):
+        r = estimate_roughness(_noise())
+        assert 0.0 <= r.roughness <= 1.0
+
+    def test_roughness_in_range_checkerboard(self):
+        r = estimate_roughness(_checkerboard())
+        assert 0.0 <= r.roughness <= 1.0
+
+    def test_solid_roughness_exactly_zero(self):
+        """Both metrics are 0 for a constant image → roughness = 0.0 exactly."""
+        r = estimate_roughness(_solid(128))
+        assert r.roughness == pytest.approx(0.0, abs=1e-6)
+
+    def test_noise_roughness_positive(self):
+        r = estimate_roughness(_noise())
+        assert r.roughness > 0.0
+
+    def test_gabor_energy_field_nonneg(self):
+        assert estimate_roughness(_solid()).gabor_energy >= 0.0
+        assert estimate_roughness(_noise()).gabor_energy >= 0.0
+
+    def test_lbp_variance_field_nonneg(self):
+        assert estimate_roughness(_solid()).lbp_variance >= 0.0
+        assert estimate_roughness(_noise()).lbp_variance >= 0.0
+
+    def test_roi_frac_1_uses_full_frame(self):
+        """roi_frac=1.0 should use the entire image."""
+        img = _noise(h=64, w=64)
+        r_full = estimate_roughness(img, roi_frac=1.0)
+        assert r_full.roughness > 0.0
+
+    def test_roi_frac_0_returns_zero(self):
+        """roi_frac=0.0 → empty crop → all zeros."""
+        r = estimate_roughness(_noise(), roi_frac=0.0)
+        assert r.roughness == pytest.approx(0.0)
+        assert r.gabor_energy == pytest.approx(0.0)
+        assert r.lbp_variance == pytest.approx(0.0)
+
+    def test_gabor_field_matches_standalone(self):
+        """gabor_energy field should match standalone gabor_energy() on the ROI."""
+        import cv2
+        img = _noise(h=64, w=64)
+        roi_frac = 0.5
+        h = img.shape[0]
+        r0 = int(h * (1.0 - roi_frac))
+        roi_grey = cv2.cvtColor(img[r0:, :], cv2.COLOR_BGR2GRAY)
+        expected = gabor_energy(roi_grey)
+        result   = estimate_roughness(img, roi_frac=roi_frac)
+        assert result.gabor_energy == pytest.approx(expected, rel=1e-4)
+
+    def test_lbp_field_matches_standalone(self):
+        """lbp_variance field should match standalone lbp_variance() on the ROI."""
+        import cv2
+        img = _noise(h=64, w=64)
+        roi_frac = 0.5
+        h = img.shape[0]
+        r0 = int(h * (1.0 - roi_frac))
+        roi_grey = cv2.cvtColor(img[r0:, :], cv2.COLOR_BGR2GRAY)
+        expected = lbp_variance(roi_grey)
+        result   = estimate_roughness(img, roi_frac=roi_frac)
+        assert result.lbp_variance == pytest.approx(expected, rel=1e-4)
+
+
+# ── estimate_roughness — ordering ────────────────────────────────────────────
+
+class TestEstimateRoughnessOrdering:
+
+    def test_noise_rougher_than_solid(self):
+        r_noise = estimate_roughness(_noise())
+        r_solid = estimate_roughness(_solid())
+        assert r_noise.roughness > r_solid.roughness
+
+    def test_checkerboard_rougher_than_gradient(self):
+        r_cb   = estimate_roughness(_checkerboard(64, 64, tile=4))
+        r_grad = estimate_roughness(_gradient())
+        assert r_cb.roughness > r_grad.roughness, (
+            f'checkerboard={r_cb.roughness:.3f} should exceed gradient={r_grad.roughness:.3f}')
+
+    def test_noise_roughness_exceeds_half(self):
+        """For fully random noise, LBP component alone contributes ≥0.5 to roughness."""
+        r = estimate_roughness(_noise(h=64, w=64), roi_frac=1.0)
+        assert r.roughness >= 0.5, (
+            f'random noise roughness should be ≥0.5, got {r.roughness:.3f}')
+
+    def test_high_roi_frac_uses_more_data(self):
+        """More floor area → LBP/Gabor computed on more pixels; score should be consistent."""
+        img = _noise(h=128, w=64)
+        r_small = estimate_roughness(img, roi_frac=0.10)
+        r_large = estimate_roughness(img, roi_frac=0.90)
+        # Both should be positive since the full image is noise
+        assert r_small.roughness >= 0.0
+        assert r_large.roughness >= 0.0
+
+
+# ── estimate_roughness — ROI sensitivity ─────────────────────────────────────
+
+class TestEstimateRoughnessROI:
+
+    def test_rough_bottom_smooth_top(self):
+        """
+        Image where bottom half is noise and top half is solid.
+        roi_frac=0.5 → picks the noisy bottom → high roughness.
+        """
+        h, w = 128, 64
+        img = np.zeros((h, w, 3), dtype=np.uint8)
+        img[:h // 2, :] = 128          # smooth top
+        rng = np.random.default_rng(0)
+        img[h // 2:, :] = rng.integers(0, 256, (h // 2, w, 3), dtype=np.uint8)
+
+        r_bottom = estimate_roughness(img, roi_frac=0.50)  # noisy half
+        r_top    = estimate_roughness(img, roi_frac=0.01)  # tiny slice near top
+
+        assert r_bottom.roughness > r_top.roughness, (
+            f'bottom(noisy)={r_bottom.roughness:.3f} should exceed top(smooth)={r_top.roughness:.3f}')
+
+    def test_roi_frac_clipped_to_valid_range(self):
+        """roi_frac > 1.0 should not crash; treated as 1.0."""
+        r = estimate_roughness(_noise(), roi_frac=1.5)
+        assert 0.0 <= r.roughness <= 1.0
+
+
+if __name__ == '__main__':
+    pytest.main([__file__, '-v'])