Graph-cut-segmentation / graph_cut_segmentation.py
priyadip's picture
Upload 4 files
332f1d0 verified
"""
========================================================================
Graph Cut Image Segmentation Pipeline
CSL7360: Computer Vision — Assignment 2
========================================================================
This module implements a complete Graph Cut segmentation pipeline:
1. Interactive annotation (scribbles) via OpenCV GUI
2. Foreground/Background modeling using GMMs
3. Graph construction with unary (data) and pairwise (smoothness) terms
4. Min-Cut / Max-Flow optimization using PyMaxflow
5. Iterative refinement of GMM models and graph cuts
6. Artifact mitigation: morphological cleaning, boundary smoothing
7. Visualization and comparison of results
========================================================================
"""
import numpy as np
import cv2
import maxflow
import os
import argparse
from sklearn.mixture import GaussianMixture
import matplotlib
# Use non-interactive backend when saving; switch to TkAgg for GUI
matplotlib.use("TkAgg")
import matplotlib.pyplot as plt
# =====================================================================
# Section 1: Interactive Annotation Tool
# =====================================================================
class ScribbleAnnotator:
"""
Interactive GUI for collecting foreground/background scribbles.
Left mouse button → Foreground (Green)
Right mouse button → Background (Red)
Press 'q' or Enter → Finish annotation
Press 'r' → Reset scribbles
"""
def __init__(self, image: np.ndarray):
self.image = image.copy()
self.display = image.copy()
self.fg_mask = np.zeros(image.shape[:2], dtype=np.uint8) # foreground
self.bg_mask = np.zeros(image.shape[:2], dtype=np.uint8) # background
self.drawing = False
self.mode = None # 'fg' or 'bg'
self.brush_size = 5
def _mouse_callback(self, event, x, y, flags, param):
"""Handle mouse events for drawing scribbles."""
if event == cv2.EVENT_LBUTTONDOWN:
self.drawing = True
self.mode = "fg"
elif event == cv2.EVENT_RBUTTONDOWN:
self.drawing = True
self.mode = "bg"
elif event == cv2.EVENT_MOUSEMOVE and self.drawing:
if self.mode == "fg":
cv2.circle(self.fg_mask, (x, y), self.brush_size, 1, -1)
cv2.circle(self.display, (x, y), self.brush_size, (0, 255, 0), -1)
elif self.mode == "bg":
cv2.circle(self.bg_mask, (x, y), self.brush_size, 1, -1)
cv2.circle(self.display, (x, y), self.brush_size, (0, 0, 255), -1)
elif event in (cv2.EVENT_LBUTTONUP, cv2.EVENT_RBUTTONUP):
self.drawing = False
def run(self) -> tuple:
"""
Launch annotation window. Returns (fg_mask, bg_mask) as binary arrays.
"""
win = "Annotate: LEFT=FG(green), RIGHT=BG(red), q=done, r=reset"
cv2.namedWindow(win, cv2.WINDOW_NORMAL)
cv2.setMouseCallback(win, self._mouse_callback)
while True:
cv2.imshow(win, self.display)
key = cv2.waitKey(1) & 0xFF
if key in (ord("q"), 13): # q or Enter
break
elif key == ord("r"):
self.display = self.image.copy()
self.fg_mask[:] = 0
self.bg_mask[:] = 0
cv2.destroyAllWindows()
return self.fg_mask, self.bg_mask
def load_annotations_from_file(image_shape, fg_path, bg_path):
"""
Load pre-saved annotation masks from disk (for non-interactive / headless mode).
Masks should be single-channel images where nonzero = annotated.
"""
h, w = image_shape[:2]
fg_mask = np.zeros((h, w), dtype=np.uint8)
bg_mask = np.zeros((h, w), dtype=np.uint8)
if os.path.exists(fg_path):
fg_img = cv2.imread(fg_path, cv2.IMREAD_GRAYSCALE)
if fg_img is not None:
fg_mask = (cv2.resize(fg_img, (w, h)) > 127).astype(np.uint8)
if os.path.exists(bg_path):
bg_img = cv2.imread(bg_path, cv2.IMREAD_GRAYSCALE)
if bg_img is not None:
bg_mask = (cv2.resize(bg_img, (w, h)) > 127).astype(np.uint8)
return fg_mask, bg_mask
def generate_auto_annotations(image: np.ndarray):
"""
Automatically generate rough foreground/background scribbles.
Foreground: center region of the image.
Background: border region of the image.
This is useful for headless / automated runs.
"""
h, w = image.shape[:2]
fg_mask = np.zeros((h, w), dtype=np.uint8)
bg_mask = np.zeros((h, w), dtype=np.uint8)
# Foreground: small central cross (like actual scribbles),
# kept tight so only object pixels are included
cy, cx = h // 2, w // 2
rh, rw = h // 10, w // 10 # 10% radius instead of 20%
t = max(h // 30, 4) # scribble thickness
# Horizontal bar
fg_mask[cy - t:cy + t, cx - rw:cx + rw] = 1
# Vertical bar
fg_mask[cy - rh:cy + rh, cx - t:cx + t] = 1
# Background: border strips (10% from each edge)
bh, bw = max(h // 10, 5), max(w // 10, 5)
bg_mask[:bh, :] = 1
bg_mask[-bh:, :] = 1
bg_mask[:, :bw] = 1
bg_mask[:, -bw:] = 1
return fg_mask, bg_mask
# =====================================================================
# Section 2: Foreground / Background Modeling (GMM)
# =====================================================================
class PixelGMMModel:
"""
Gaussian Mixture Model for foreground or background pixel distribution.
Fits a GMM to the color values of annotated/labelled pixels and
returns log-likelihood scores for any query pixel.
"""
def __init__(self, n_components: int = 5):
self.n_components = n_components
self.gmm = GaussianMixture(
n_components=n_components,
covariance_type="full",
max_iter=200,
random_state=42,
)
self.fitted = False
def fit(self, pixels: np.ndarray):
"""
Fit GMM to pixel samples. pixels: (N, 3) array of BGR values.
"""
if len(pixels) < self.n_components:
# Fall back to fewer components if too few samples
self.gmm = GaussianMixture(
n_components=max(1, len(pixels)),
covariance_type="full",
max_iter=200,
random_state=42,
)
self.gmm.fit(pixels)
self.fitted = True
def score_pixels(self, pixels: np.ndarray) -> np.ndarray:
"""
Return per-sample log-likelihood. pixels: (N, 3).
Higher = more likely to belong to this model.
"""
if not self.fitted:
return np.zeros(len(pixels))
return self.gmm.score_samples(pixels)
def build_gmm_models(image: np.ndarray, fg_mask: np.ndarray, bg_mask: np.ndarray,
n_components: int = 5):
"""
Build foreground and background GMMs from annotated pixels.
Returns (fg_model, bg_model).
"""
fg_pixels = image[fg_mask == 1].reshape(-1, 3).astype(np.float64)
bg_pixels = image[bg_mask == 1].reshape(-1, 3).astype(np.float64)
fg_model = PixelGMMModel(n_components)
bg_model = PixelGMMModel(n_components)
if len(fg_pixels) > 0:
fg_model.fit(fg_pixels)
if len(bg_pixels) > 0:
bg_model.fit(bg_pixels)
return fg_model, bg_model
# =====================================================================
# Section 3: Energy Formulation & Graph Construction
# =====================================================================
def compute_unary_costs(image: np.ndarray, fg_model: PixelGMMModel,
bg_model: PixelGMMModel,
fg_mask: np.ndarray, bg_mask: np.ndarray,
hard_constraint_weight: float = 1e9) -> tuple:
"""
Compute unary (data) costs for each pixel.
E_data(x_p) = -log P(I_p | label)
For annotated pixels, we assign a very high cost to the opposite label
(hard constraints).
Returns:
fg_cost: (H, W) — cost of assigning pixel to foreground (source)
bg_cost: (H, W) — cost of assigning pixel to background (sink)
"""
h, w = image.shape[:2]
pixels = image.reshape(-1, 3).astype(np.float64)
# Log-likelihoods from GMMs
fg_ll = fg_model.score_pixels(pixels).reshape(h, w)
bg_ll = bg_model.score_pixels(pixels).reshape(h, w)
# Convert to costs: cost = -log_likelihood (lower likelihood → higher cost)
# We negate because score_samples returns log-probability
# Cost of labeling as foreground = negative log-prob under foreground model
# We want: if pixel looks like BG, cost of labeling it FG should be high
# So: cost_fg = -log P(pixel | FG) ... but score_samples already gives log P
# Therefore: cost_to_be_sink(bg) = -fg_ll (pixel not matching FG → high bg cost? No.)
#
# Standard formulation:
# source capacity (weight for cutting source edge = assigning to BG) = -log P(I|BG)
# sink capacity (weight for cutting sink edge = assigning to FG) = -log P(I|FG)
#
# Wait — let's be precise:
# If pixel is connected to Source (FG) and Sink (BG),
# cutting the source edge → pixel goes to BG → cost should be high if pixel is FG-like
# So source_cap = -log P(I|FG) is WRONG for that.
#
# Correct:
# source_cap (edge from S to pixel) = -log P(I_p | BG) → high when pixel unlikely BG
# sink_cap (edge from pixel to T) = -log P(I_p | FG) → high when pixel unlikely FG
#
# Cutting source edge means pixel goes to sink (BG).
# So source_cap should be the "penalty for going BG" = how unlikely it is under BG = -bg_ll
source_cap = -bg_ll # penalty for assigning to background
sink_cap = -fg_ll # penalty for assigning to foreground
# Shift to ensure non-negative costs
min_val = min(source_cap.min(), sink_cap.min())
if min_val < 0:
source_cap -= min_val
sink_cap -= min_val
# Hard constraints for annotated pixels
source_cap[fg_mask == 1] = hard_constraint_weight
sink_cap[fg_mask == 1] = 0
source_cap[bg_mask == 1] = 0
sink_cap[bg_mask == 1] = hard_constraint_weight
return source_cap, sink_cap
def compute_pairwise_costs(image: np.ndarray, beta: float = None,
gamma: float = 50.0) -> tuple:
"""
Compute pairwise (smoothness) costs between neighboring pixels.
E_smooth(x_p, x_q) = gamma * exp(-beta * ||I_p - I_q||^2) if x_p ≠ x_q
= 0 if x_p == x_q
beta = 1 / (2 * <||I_p - I_q||^2>) (average over all neighbor pairs)
We compute weights for 4-connected neighbors (right, down).
Returns:
right_weights: (H, W) — smoothness weight for horizontal edges
down_weights: (H, W) — smoothness weight for vertical edges
"""
img = image.astype(np.float64)
h, w = img.shape[:2]
# Compute differences for right and down neighbors
diff_right = img[:, 1:, :] - img[:, :-1, :] # (H, W-1, 3)
diff_down = img[1:, :, :] - img[:-1, :, :] # (H-1, W, 3)
dist_right = np.sum(diff_right ** 2, axis=2) # (H, W-1)
dist_down = np.sum(diff_down ** 2, axis=2) # (H-1, W)
# Compute beta from average squared color distance
if beta is None:
total_sum = dist_right.sum() + dist_down.sum()
total_count = dist_right.size + dist_down.size
avg_dist = total_sum / total_count if total_count > 0 else 1.0
beta = 1.0 / (2.0 * avg_dist) if avg_dist > 0 else 0.0
# Smoothness weights
right_weights = gamma * np.exp(-beta * dist_right)
down_weights = gamma * np.exp(-beta * dist_down)
return right_weights, down_weights, beta
def build_graph_and_cut(source_cap: np.ndarray, sink_cap: np.ndarray,
right_weights: np.ndarray, down_weights: np.ndarray) -> np.ndarray:
"""
Construct the graph using PyMaxflow and solve the min-cut / max-flow.
Graph structure:
- Source node S represents Foreground
- Sink node T represents Background
- Each pixel is a node
- Terminal edges: S→pixel (source_cap), pixel→T (sink_cap)
- Neighbor edges: between adjacent pixels (pairwise smoothness)
The min-cut partitions pixels into S-set (foreground) and T-set (background).
Returns:
labels: (H, W) binary mask — 1 = foreground, 0 = background
"""
h, w = source_cap.shape
# Create graph
g = maxflow.Graph[float](h * w, h * w * 2)
g.add_nodes(h * w)
# Add terminal edges (unary / data costs)
for i in range(h):
for j in range(w):
idx = i * w + j
g.add_tedge(idx, source_cap[i, j], sink_cap[i, j])
# Add pairwise (smoothness) edges — 4-connected neighborhood
# Right neighbors
for i in range(h):
for j in range(w - 1):
idx1 = i * w + j
idx2 = i * w + (j + 1)
weight = right_weights[i, j]
g.add_edge(idx1, idx2, weight, weight)
# Down neighbors
for i in range(h - 1):
for j in range(w):
idx1 = i * w + j
idx2 = (i + 1) * w + j
weight = down_weights[i, j]
g.add_edge(idx1, idx2, weight, weight)
# Solve min-cut / max-flow
flow = g.maxflow()
print(f" Max-flow value: {flow:.2f}")
# Extract labels: 0 = source side (FG), 1 = sink side (BG) in PyMaxflow
segments = np.array([g.get_segment(idx) for idx in range(h * w)])
labels = segments.reshape(h, w)
# In PyMaxflow: segment 0 = source side = foreground
# segment 1 = sink side = background
# We want 1 = foreground, 0 = background
labels = 1 - labels
return labels
# =====================================================================
# Section 4: Iterative Graph Cut Optimization
# =====================================================================
def iterative_graph_cut(image: np.ndarray, fg_mask: np.ndarray, bg_mask: np.ndarray,
n_iterations: int = 3, n_components: int = 5,
gamma: float = 50.0) -> tuple:
"""
Perform iterative graph cut segmentation:
1. Build initial GMMs from user scribbles.
2. Construct graph and compute min-cut.
3. Update GMMs using newly labelled pixels.
4. Repeat for n_iterations.
Returns:
final_mask: (H, W) binary segmentation
all_masks: list of masks at each iteration (for comparison)
energies: list of energy values per iteration
"""
h, w = image.shape[:2]
current_fg_mask = fg_mask.copy()
current_bg_mask = bg_mask.copy()
all_masks = []
energies = []
for it in range(n_iterations):
print(f" Iteration {it + 1}/{n_iterations}")
# Step 1: Build / Update GMMs
fg_model, bg_model = build_gmm_models(image, current_fg_mask, current_bg_mask,
n_components)
# Step 2: Compute unary costs
source_cap, sink_cap = compute_unary_costs(image, fg_model, bg_model,
fg_mask, bg_mask)
# Step 3: Compute pairwise costs
right_w, down_w, beta = compute_pairwise_costs(image, gamma=gamma)
# Step 4: Build graph and solve min-cut
labels = build_graph_and_cut(source_cap, sink_cap, right_w, down_w)
all_masks.append(labels.copy())
# Compute energy for monitoring convergence
pixels = image.reshape(-1, 3).astype(np.float64)
fg_ll = fg_model.score_pixels(pixels).reshape(h, w)
bg_ll = bg_model.score_pixels(pixels).reshape(h, w)
data_energy = -np.sum(fg_ll[labels == 1]) - np.sum(bg_ll[labels == 0])
# Smoothness energy (count boundary edges)
smooth_energy = 0
diff_h = (labels[:, 1:] != labels[:, :-1]).astype(float)
diff_v = (labels[1:, :] != labels[:-1, :]).astype(float)
smooth_energy = np.sum(diff_h * right_w) + np.sum(diff_v * down_w)
total_energy = data_energy + smooth_energy
energies.append(total_energy)
print(f" Energy: {total_energy:.2f} (data={data_energy:.2f}, smooth={smooth_energy:.2f})")
# Step 5: Update masks for next iteration
current_fg_mask = labels.copy()
current_bg_mask = (1 - labels).copy()
# Preserve hard constraints from user annotations
current_fg_mask[fg_mask == 1] = 1
current_bg_mask[bg_mask == 1] = 1
# Return the mask from the lowest-energy iteration (not necessarily the last)
best_iter = int(np.argmin(energies))
print(f" Best iteration: {best_iter + 1} (energy={energies[best_iter]:.2f})")
return all_masks[best_iter], all_masks, energies
# =====================================================================
# Section 5: Artifact Mitigation & Refinement
# =====================================================================
def remove_small_regions(mask: np.ndarray, min_area: int = 500) -> np.ndarray:
"""
Remove small isolated foreground and background regions using
connected component analysis and morphological operations.
"""
cleaned = mask.copy().astype(np.uint8)
# Remove small foreground regions
num_labels, labels, stats, _ = cv2.connectedComponentsWithStats(cleaned, connectivity=8)
for i in range(1, num_labels):
if stats[i, cv2.CC_STAT_AREA] < min_area:
cleaned[labels == i] = 0
# Remove small background holes (invert, clean, invert back)
inv = 1 - cleaned
num_labels, labels, stats, _ = cv2.connectedComponentsWithStats(inv, connectivity=8)
for i in range(1, num_labels):
if stats[i, cv2.CC_STAT_AREA] < min_area:
cleaned[labels == i] = 1
return cleaned
def smooth_boundaries(mask: np.ndarray, ksize: int = 5) -> np.ndarray:
"""
Smooth jagged segmentation boundaries using morphological closing
followed by Gaussian blur and re-thresholding.
"""
m = mask.astype(np.uint8) * 255
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (ksize, ksize))
# Close small gaps
m = cv2.morphologyEx(m, cv2.MORPH_CLOSE, kernel, iterations=1)
# Open to remove thin protrusions
m = cv2.morphologyEx(m, cv2.MORPH_OPEN, kernel, iterations=1)
# Gaussian blur + threshold for smooth boundary
m = cv2.GaussianBlur(m, (ksize * 2 + 1, ksize * 2 + 1), 0)
m = (m > 127).astype(np.uint8)
return m
def ensure_intensity_consistency(mask: np.ndarray, image: np.ndarray,
threshold: float = 30.0) -> np.ndarray:
"""
Intensity Consistency: re-label pixels near the boundary whose color is
significantly closer to the opposite region's mean color.
For each foreground pixel within a border band, if its color distance to
the background mean is smaller than to the foreground mean, flip it to
background (and vice versa). This corrects visually incoherent pixels
that slipped through the graph cut due to weak data terms.
"""
refined = mask.copy().astype(np.uint8)
img_f = image.astype(np.float32)
fg_pixels = img_f[refined == 1]
bg_pixels = img_f[refined == 0]
if len(fg_pixels) == 0 or len(bg_pixels) == 0:
return refined
fg_mean = fg_pixels.mean(axis=0) # mean FG color (BGR)
bg_mean = bg_pixels.mean(axis=0) # mean BG color (BGR)
# Build a narrow band around the boundary (dilate XOR original)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7, 7))
dilated = cv2.dilate(refined, kernel, iterations=1)
eroded = cv2.erode(refined, kernel, iterations=1)
band = (dilated - eroded).astype(bool) # True = boundary pixels
for label, correct_mean, wrong_mean in [(1, fg_mean, bg_mean),
(0, bg_mean, fg_mean)]:
region_band = band & (refined == label)
coords = np.argwhere(region_band)
for (r, c) in coords:
color = img_f[r, c]
d_correct = float(np.linalg.norm(color - correct_mean))
d_wrong = float(np.linalg.norm(color - wrong_mean))
# Flip only when the pixel is clearly closer to the opposite mean
if d_wrong < d_correct - threshold:
refined[r, c] = 1 - label
return refined
def refine_segmentation(mask: np.ndarray, image: np.ndarray,
min_area: int = None, smooth_ksize: int = 3) -> np.ndarray:
"""
Full refinement pipeline:
1. Remove small isolated regions (morphological noise removal)
2. Smooth jagged boundaries
3. Intensity consistency correction near boundaries
min_area defaults to 0.1% of image pixels to scale with image size.
smooth_ksize reduced to 3 to avoid distorting fine structures.
"""
print(" Refining segmentation...")
if min_area is None:
min_area = max(50, int(mask.size * 0.001)) # 0.1% of pixels
refined = remove_small_regions(mask, min_area)
refined = smooth_boundaries(refined, smooth_ksize)
refined = ensure_intensity_consistency(refined, image)
return refined
# =====================================================================
# Section 6: Naive Segmentation (for comparison)
# =====================================================================
def naive_thresholding_segmentation(image: np.ndarray) -> np.ndarray:
"""
Simple Otsu thresholding as a naive baseline for comparison.
Returns raw Otsu mask; label alignment to graph cut is done after
graph cut is computed (see align_naive_to_graphcut).
"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
_, mask = cv2.threshold(gray, 0, 1, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
return mask
def align_naive_to_graphcut(naive_mask: np.ndarray,
reference_mask: np.ndarray) -> np.ndarray:
"""
Align a naive mask's label convention to match the graph cut reference.
Checks whether the mask or its inverse has more overlap with the reference,
and returns whichever agrees more. This handles cases where Otsu/K-Means
assign FG=1 to the bright region while graph cut assigns FG=1 to the object.
"""
overlap_normal = np.sum(naive_mask == reference_mask)
overlap_inverted = np.sum((1 - naive_mask) == reference_mask)
if overlap_inverted > overlap_normal:
return 1 - naive_mask
return naive_mask
def naive_kmeans_segmentation(image: np.ndarray, k: int = 2) -> np.ndarray:
"""
K-Means clustering as another naive baseline.
"""
pixels = image.reshape(-1, 3).astype(np.float32)
criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 100, 0.2)
_, labels, centers = cv2.kmeans(pixels, k, None, criteria, 10,
cv2.KMEANS_RANDOM_CENTERS)
# Assign the darker cluster as background
labels = labels.reshape(image.shape[:2])
if centers[0].mean() > centers[1].mean():
labels = 1 - labels
return labels.astype(np.uint8)
# =====================================================================
# Section 7: Visualization
# =====================================================================
def create_overlay(image: np.ndarray, mask: np.ndarray,
color: tuple = (0, 255, 0), alpha: float = 0.4) -> np.ndarray:
"""
Overlay a segmentation mask on the original image.
"""
overlay = image.copy()
colored = np.zeros_like(image)
colored[:] = color
region = mask.astype(bool)
overlay[region] = cv2.addWeighted(image[region], 1 - alpha,
colored[region], alpha, 0)
# Draw contours
contours, _ = cv2.findContours(mask.astype(np.uint8), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
cv2.drawContours(overlay, contours, -1, color, 2)
return overlay
def visualize_results(image: np.ndarray, fg_mask: np.ndarray, bg_mask: np.ndarray,
raw_mask: np.ndarray, refined_mask: np.ndarray,
naive_mask: np.ndarray, naive_kmeans_mask: np.ndarray,
all_iter_masks: list, energies: list,
output_dir: str, img_name: str):
"""
Generate comprehensive visualization of all results and save to disk.
"""
# --- Figure 1: Main comparison (2×4 grid) ---
fig, axes = plt.subplots(2, 4, figsize=(24, 12))
fig.suptitle(f"Graph Cut Segmentation — {img_name}", fontsize=16, fontweight="bold")
# Original + scribbles
scribble_vis = image.copy()
scribble_vis[fg_mask == 1] = [0, 255, 0]
scribble_vis[bg_mask == 1] = [0, 0, 255]
axes[0, 0].imshow(cv2.cvtColor(scribble_vis, cv2.COLOR_BGR2RGB))
axes[0, 0].set_title("Input + Annotations")
axes[0, 0].axis("off")
# Naive segmentation — Otsu
axes[0, 1].imshow(naive_mask, cmap="gray")
axes[0, 1].set_title("Naive: Otsu Thresholding")
axes[0, 1].axis("off")
# Naive segmentation — K-Means
axes[0, 2].imshow(naive_kmeans_mask, cmap="gray")
axes[0, 2].set_title("Naive: K-Means (k=2)")
axes[0, 2].axis("off")
# Raw graph cut
axes[0, 3].imshow(raw_mask, cmap="gray")
axes[0, 3].set_title("Raw Graph Cut")
axes[0, 3].axis("off")
# Refined mask
axes[1, 0].imshow(refined_mask, cmap="gray")
axes[1, 0].set_title("Refined Graph Cut")
axes[1, 0].axis("off")
# Overlay on original
overlay = create_overlay(image, refined_mask)
axes[1, 1].imshow(cv2.cvtColor(overlay, cv2.COLOR_BGR2RGB))
axes[1, 1].set_title("Overlay on Original")
axes[1, 1].axis("off")
# Extracted foreground
extracted = image.copy()
extracted[refined_mask == 0] = [255, 255, 255]
axes[1, 2].imshow(cv2.cvtColor(extracted, cv2.COLOR_BGR2RGB))
axes[1, 2].set_title("Extracted Foreground")
axes[1, 2].axis("off")
# Side-by-side comparison: best naive vs graph cut (with gap)
h_cmp = naive_mask.shape[0]
gap = np.full((h_cmp, 10), 255, dtype=np.uint8) # white divider
compare = np.hstack([
naive_mask.astype(np.uint8) * 255,
gap,
refined_mask.astype(np.uint8) * 255
])
axes[1, 3].imshow(compare, cmap="gray")
axes[1, 3].set_title("Otsu vs Graph Cut (side-by-side)")
axes[1, 3].axis("off")
plt.tight_layout()
fig.savefig(os.path.join(output_dir, f"{img_name}_results.png"), dpi=150,
bbox_inches="tight")
plt.close(fig)
# --- Figure 2: Iteration progression ---
if len(all_iter_masks) > 1:
n = len(all_iter_masks)
fig2, axes2 = plt.subplots(1, n + 1, figsize=(5 * (n + 1), 5))
fig2.suptitle(f"Iterative Refinement — {img_name}", fontsize=14)
for i, m in enumerate(all_iter_masks):
axes2[i].imshow(m, cmap="gray")
axes2[i].set_title(f"Iteration {i + 1}")
axes2[i].axis("off")
axes2[n].imshow(refined_mask, cmap="gray")
axes2[n].set_title("After Post-Processing")
axes2[n].axis("off")
plt.tight_layout()
fig2.savefig(os.path.join(output_dir, f"{img_name}_iterations.png"), dpi=150,
bbox_inches="tight")
plt.close(fig2)
# --- Figure 3: Energy convergence ---
if len(energies) > 1:
fig3, ax3 = plt.subplots(figsize=(8, 5))
ax3.plot(range(1, len(energies) + 1), energies, "bo-", linewidth=2, markersize=8)
ax3.set_xlabel("Iteration", fontsize=12)
ax3.set_ylabel("Total Energy", fontsize=12)
ax3.set_title(f"Energy Convergence — {img_name}", fontsize=14)
ax3.grid(True, alpha=0.3)
fig3.savefig(os.path.join(output_dir, f"{img_name}_energy.png"), dpi=150,
bbox_inches="tight")
plt.close(fig3)
print(f" Visualizations saved to {output_dir}/")
# =====================================================================
# Section 8: Full Pipeline
# =====================================================================
def run_pipeline(image_path: str, output_dir: str = "outputs",
n_iterations: int = 3, n_components: int = 5,
gamma: float = 50.0, interactive: bool = True,
fg_anno_path: str = None, bg_anno_path: str = None,
auto_annotate: bool = False,
max_dim: int = 400):
"""
Run the complete Graph Cut segmentation pipeline on a single image.
Parameters:
image_path: Path to input image
output_dir: Directory to save results
n_iterations: Number of graph-cut iterations
n_components: GMM components
gamma: Smoothness weight
interactive: If True, open GUI for scribble annotation
fg_anno_path: Path to pre-made foreground annotation mask
bg_anno_path: Path to pre-made background annotation mask
auto_annotate: If True, generate automatic center/border annotations
max_dim: Resize image so largest dimension ≤ max_dim (for speed)
"""
os.makedirs(output_dir, exist_ok=True)
img_name = os.path.splitext(os.path.basename(image_path))[0]
print(f"\n{'='*60}")
print(f"Processing: {image_path}")
print(f"{'='*60}")
# Load image
image = cv2.imread(image_path)
if image is None:
print(f"ERROR: Could not load image '{image_path}'")
return
# Resize for tractability
h, w = image.shape[:2]
if max(h, w) > max_dim:
scale = max_dim / max(h, w)
image = cv2.resize(image, None, fx=scale, fy=scale, interpolation=cv2.INTER_AREA)
print(f" Resized from ({w},{h}) to {image.shape[1]}x{image.shape[0]}")
# Step 1: Obtain annotations
print("Step 1: Obtaining annotations...")
if interactive:
annotator = ScribbleAnnotator(image)
fg_mask, bg_mask = annotator.run()
elif fg_anno_path and bg_anno_path:
fg_mask, bg_mask = load_annotations_from_file(image.shape, fg_anno_path, bg_anno_path)
elif auto_annotate:
fg_mask, bg_mask = generate_auto_annotations(image)
else:
print(" No annotation source specified. Using auto-annotation.")
fg_mask, bg_mask = generate_auto_annotations(image)
fg_count = fg_mask.sum()
bg_count = bg_mask.sum()
print(f" Foreground scribble pixels: {fg_count}")
print(f" Background scribble pixels: {bg_count}")
if fg_count == 0 or bg_count == 0:
print(" WARNING: Need both FG and BG annotations. Using auto-annotation.")
fg_mask, bg_mask = generate_auto_annotations(image)
# Step 2: Naive segmentation (baseline — both Otsu and K-Means)
print("Step 2: Computing naive baseline segmentation...")
naive_mask = naive_thresholding_segmentation(image)
naive_kmeans_mask = naive_kmeans_segmentation(image)
# Step 3: Iterative graph cut
print("Step 3: Running iterative graph cut segmentation...")
raw_mask, all_masks, energies = iterative_graph_cut(
image, fg_mask, bg_mask,
n_iterations=n_iterations,
n_components=n_components,
gamma=gamma,
)
# Step 4: Refine segmentation
print("Step 4: Refining segmentation (artifact mitigation)...")
refined_mask = refine_segmentation(raw_mask, image)
# Align naive masks to graph cut label convention (FG=1 must mean the same thing)
naive_mask = align_naive_to_graphcut(naive_mask, refined_mask)
naive_kmeans_mask = align_naive_to_graphcut(naive_kmeans_mask, refined_mask)
# Step 5: Save outputs
print("Step 5: Saving results...")
cv2.imwrite(os.path.join(output_dir, f"{img_name}_raw_mask.png"),
(raw_mask * 255).astype(np.uint8))
cv2.imwrite(os.path.join(output_dir, f"{img_name}_refined_mask.png"),
(refined_mask * 255).astype(np.uint8))
overlay = create_overlay(image, refined_mask)
cv2.imwrite(os.path.join(output_dir, f"{img_name}_overlay.png"), overlay)
# Step 6: Visualize
print("Step 6: Generating visualizations...")
visualize_results(image, fg_mask, bg_mask, raw_mask, refined_mask,
naive_mask, naive_kmeans_mask, all_masks, energies, output_dir, img_name)
print(f" Done: {img_name}")
return refined_mask
# =====================================================================
# Section 9: Entry Point
# =====================================================================
def main():
parser = argparse.ArgumentParser(
description="Graph Cut Image Segmentation Pipeline — CSL7360 Assignment 2",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog="""
Examples:
# Interactive annotation (opens GUI window)
python graph_cut_segmentation.py --images img1.jpg img2.jpg img3.jpg
# Automatic annotation (headless, no GUI)
python graph_cut_segmentation.py --images img1.jpg --auto
# Custom parameters
python graph_cut_segmentation.py --images img1.jpg --iterations 5 --gamma 80 --gmm-components 7
""",
)
parser.add_argument("--images", nargs="+", required=True,
help="Paths to input images (at least 3 recommended)")
parser.add_argument("--output", default="outputs",
help="Output directory (default: outputs)")
parser.add_argument("--iterations", type=int, default=3,
help="Number of iterative optimization steps (default: 3)")
parser.add_argument("--gmm-components", type=int, default=5,
help="Number of GMM components per model (default: 5)")
parser.add_argument("--gamma", type=float, default=50.0,
help="Smoothness weight gamma (default: 50.0)")
parser.add_argument("--max-dim", type=int, default=400,
help="Max image dimension for processing (default: 400)")
parser.add_argument("--auto", action="store_true",
help="Use automatic center/border annotations (no GUI)")
parser.add_argument("--no-interactive", action="store_true",
help="Disable interactive GUI (use --auto or provide masks)")
args = parser.parse_args()
interactive = not (args.auto or args.no_interactive)
for img_path in args.images:
run_pipeline(
image_path=img_path,
output_dir=args.output,
n_iterations=args.iterations,
n_components=args.gmm_components,
gamma=args.gamma,
interactive=interactive,
auto_annotate=args.auto,
max_dim=args.max_dim,
)
print(f"\nAll results saved in '{args.output}/' directory.")
print("Pipeline complete.")
if __name__ == "__main__":
main()