File size: 5,611 Bytes
0f9608b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
import torch
import numpy as np
from PIL import Image
import cv2
from torchvision import transforms
from experiments.otsu_segmenter import otsu_threshold
from experiments.SegNet.efficient_b0_backbone.architecture import SegNetEfficientNet, NUM_CLASSES, DEVICE

def ensemble_segmentation(image_path, model_path="segnet_efficientnet_voc.pth", boundary_weight=0.3):
    """
    Ensemble segmentation combining Otsu thresholding and SegNet
    
    Args:
        image_path: Path to input image
        model_path: Path to SegNet model weights
        boundary_weight: Weight for boundary refinement (0-1)
        
    Returns:
        original_image: Original input image (PIL)
        ensemble_result: Ensemble segmentation result (PIL)
        method_comparison: Visualization of all methods side by side (PIL)
    """
    # 1. Load the image
    image = Image.open(image_path).convert('RGB')
    original = image.copy()
    image_np = np.array(image)
    
    # 2. Run Otsu thresholding for boundary detection
    # Convert to grayscale and apply Gaussian blur
    gray = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
    gray = cv2.cvtColor(gray, cv2.COLOR_BGR2GRAY)
    blurred = cv2.GaussianBlur(gray, (5, 5), 0)
    otsu_threshold_value, otsu_mask = otsu_threshold(blurred)
    
    # 3. Run SegNet for semantic segmentation
    model = SegNetEfficientNet(NUM_CLASSES).to(DEVICE)
    model.load_state_dict(torch.load(model_path, map_location=DEVICE))
    model.eval()
    
    transform = transforms.Compose([
    transforms.Resize((360, 480)),  # Or larger if needed
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])
    
    input_tensor = transform(image).unsqueeze(0).to(DEVICE)
    
    with torch.no_grad():
        output = model(input_tensor)
        segnet_pred = torch.argmax(output, dim=1).squeeze(0).cpu().numpy()
    
    # 4. Create edge map from Otsu result
    edges = cv2.Canny(otsu_mask, 50, 150)
    
    # Resize to match SegNet output size
    edges_resized = cv2.resize(edges, (segnet_pred.shape[1], segnet_pred.shape[0]), 
                               interpolation=cv2.INTER_NEAREST)
    
    # 5. Ensemble: Use Otsu edges to refine SegNet boundaries
    # Create a distance transform from the edges
    dist_transform = cv2.distanceTransform(255 - edges_resized, cv2.DIST_L2, 5)
    dist_transform = dist_transform / dist_transform.max()  # Normalize to 0-1
    
    # Areas close to edges get more influence from Otsu
    edge_weight_map = np.exp(-dist_transform * 5) * boundary_weight
    
    # Create binary mask from SegNet (foreground = any class other than background)
    segnet_binary = (segnet_pred > 0).astype(np.uint8) * 255
    
    # Resize Otsu mask to match SegNet output
    otsu_resized = cv2.resize(otsu_mask, (segnet_pred.shape[1], segnet_pred.shape[0]), 
                              interpolation=cv2.INTER_NEAREST)
    
    # Combine: Use SegNet classes but refine boundaries with Otsu
    # For boundary regions, adjust the segmentation based on Otsu
    refined_binary = segnet_binary.copy()
    boundary_region = edge_weight_map > 0.1
    refined_binary[boundary_region] = (
        (1 - edge_weight_map[boundary_region]) * segnet_binary[boundary_region] + 
        edge_weight_map[boundary_region] * otsu_resized[boundary_region]
    ).astype(np.uint8)
    
    # Apply the refined binary mask to the original SegNet prediction
    ensemble_result = segnet_pred.copy()
    # Where the refined binary is 0, set to background class (0)
    ensemble_result[refined_binary < 128] = 0
    
    # 6. Visualize results
    from matplotlib import cm
    import matplotlib.pyplot as plt
    import io
    
    # Convert semantic maps to color visualizations
    colormap = cm.get_cmap('nipy_spectral')
    
    segnet_colored = colormap(segnet_pred / (NUM_CLASSES - 1))
    segnet_colored = (segnet_colored[:, :, :3] * 255).astype(np.uint8)
    
    ensemble_colored = colormap(ensemble_result / (NUM_CLASSES - 1))
    ensemble_colored = (ensemble_colored[:, :, :3] * 255).astype(np.uint8)
    
    # Create side-by-side comparison
    fig, axes = plt.subplots(1, 4, figsize=(16, 4))
    
    # Resize original image to match the segmentation size
    original_resized = original.resize((segnet_pred.shape[1], segnet_pred.shape[0]))
    
    axes[0].imshow(original_resized)
    axes[0].set_title("Original Image")
    axes[0].axis('off')
    
    axes[1].imshow(otsu_mask, cmap='gray')
    axes[1].set_title(f"Otsu (t={otsu_threshold_value})")
    axes[1].axis('off')
    
    axes[2].imshow(segnet_colored)
    axes[2].set_title("SegNet Prediction")
    axes[2].axis('off')
    
    axes[3].imshow(ensemble_colored)
    axes[3].set_title("Ensemble Result")
    axes[3].axis('off')
    
    plt.tight_layout()
    
    # Convert the plot to an image
    buf = io.BytesIO()
    plt.savefig(buf, format='png')
    buf.seek(0)
    comparison_image = Image.open(buf)
    plt.close(fig)
    
    # Return results
    ensemble_pil = Image.fromarray(ensemble_colored)
    ensemble_pil = ensemble_pil.resize(original.size, Image.NEAREST)
    
    return original, ensemble_pil, comparison_image

# Add this function to your app.py
def generate_ensemble_segmentation(image_path, boundary_weight=0.3):
    """Wrapper for Gradio interface"""
    original, ensemble_result, comparison = ensemble_segmentation(
        image_path, 
        model_path="saved_models/segnet_efficientnet_camvid.pth",
        boundary_weight=boundary_weight
    )
    return original, ensemble_result, comparison