add patch evluation

patch_generation/patch_generation_3.py

@@ -0,0 +1,139 @@
+#%%
+import numpy as np
+import cv2
+
+def merge_intervals(intervals):
+    # Merge overlapping intervals.
+    if not intervals:
+        return []
+    intervals.sort(key=lambda x: x[0])
+    merged = [intervals[0]]
+    for current in intervals[1:]:
+        last = merged[-1]
+        if current[0] <= last[1]:
+            merged[-1] = (last[0], max(last[1], current[1]))
+        else:
+            merged.append(current)
+    return merged
+
+def get_allowed_intervals(neighbor_values, delta):
+    # Compute forbidden intervals for this channel from neighbor values.
+    forbidden = []
+    for c in neighbor_values:
+        low = max(0, c - delta)
+        high = min(1, c + delta)
+        forbidden.append((low, high))
+    # Merge overlapping forbidden intervals.
+    merged = merge_intervals(forbidden)
+    allowed = []
+    # Allowed region before the first forbidden interval.
+    if merged:
+        if merged[0][0] > 0:
+            allowed.append((0, merged[0][0]))
+        # Allowed regions between forbidden intervals.
+        for i in range(1, len(merged)):
+            allowed.append((merged[i-1][1], merged[i][0]))
+        # Allowed region after the last forbidden interval.
+        if merged[-1][1] < 1:
+            allowed.append((merged[-1][1], 1))
+    else:
+        allowed.append((0, 1))
+    return allowed
+
+def sample_from_allowed(allowed):
+    # Choose a random value from the allowed intervals with probability proportional to interval length.
+    lengths = [end - start for start, end in allowed]
+    total_length = sum(lengths)
+    if total_length <= 0:
+        return None
+    point = np.random.rand() * total_length
+    cumulative = 0
+    for (start, end), length in zip(allowed, lengths):
+        cumulative += length
+        if point <= cumulative:
+            return np.random.uniform(start, end)
+    return allowed[-1][0]
+
+def sample_channel(neighbor_values, delta):
+    # Compute allowed intervals and sample a value.
+    allowed = get_allowed_intervals(neighbor_values, delta)
+    if not allowed:
+        return None
+    return sample_from_allowed(allowed)
+
+def generate_valid_color(neighbors, r):
+    """
+    Generate a random color that satisfies, for each neighbor,
+    |R - R_neighbor| >= sqrt(r), and similarly for G and B.
+    neighbors: list of neighbor colors (each is an array of shape (3,))
+    r: threshold for squared difference.
+    """
+    delta = np.sqrt(r)
+    color = np.zeros(3)
+    for channel in range(3):
+        neighbor_channel = [n[channel] for n in neighbors]
+        sampled = sample_channel(neighbor_channel, delta)
+        # If the allowed intervals are empty, use a fallback.
+        if sampled is None:
+            sampled = np.random.rand()
+        color[channel] = sampled
+    return color
+
+def generate_color_pattern(h, w, d, r):
+    """
+    Generate a dense color pattern.
+    h, w: height and width of the output image in pixels.
+    d: patch size in pixels (each patch is d x d).
+    r: randomness regulation factor. For each neighbor the condition is:
+       (|channel_new - channel_neighbor|)^2 >= r.
+    """
+    pattern = np.zeros((h, w, 3), dtype=np.float32)
+    num_patches_x = h // d
+    num_patches_y = w // d
+
+    for i in range(num_patches_x):
+        for j in range(num_patches_y):
+            # Collect neighbor patch colors (using mean because patches have uniform color).
+            neighbors = []
+            if i > 0:
+                patch_above = pattern[(i-1)*d:i*d, j*d:(j+1)*d]
+                neighbors.append(np.mean(patch_above, axis=(0, 1)))
+            if j > 0:
+                patch_left = pattern[i*d:(i+1)*d, (j-1)*d:j*d]
+                neighbors.append(np.mean(patch_left, axis=(0, 1)))
+                
+            if len(neighbors) == 0:
+                # For the first patch use a random color.
+                new_color = np.random.rand(3)
+            else:
+                new_color = generate_valid_color(neighbors, r)
+                
+            pattern[i*d:(i+1)*d, j*d:(j+1)*d] = new_color
+            print('Patch ({}, {})'.format(i, j))
+    return pattern
+
+def main():
+    # Parameters for the pattern.
+    # For example, using 100x100 pixels image with patches of 3x3 pixels.
+    height = 200
+    width = 200
+    patch_size = 3
+    # Set r to a value that gives sufficient difference.
+    # Note: r is the threshold for squared difference per channel, and the condition is
+    # |c - c_neighbor| >= sqrt(r). Very high r may make it impossible to satisfy.
+    r = 0.5
+
+    pattern = generate_color_pattern(height, width, patch_size, r)
+    # Scale to 0-255 for display and saving.
+    pattern_display = (pattern * 255).astype(np.uint8)
+
+    # Display the pattern with OpenCV.
+    cv2.imshow('Pattern', pattern_display)
+    cv2.waitKey(0)
+    cv2.destroyAllWindows()
+
+    # Save the pattern as an image file.
+    cv2.imwrite('color_pattern.png', pattern_display)
+
+if __name__ == '__main__':
+    main()