Module 9: Image Segmentation

1. Learning Objectives

• Understand the principles of semantic and instance segmentation
• Explore deep learning architectures used in segmentation tasks
• Apply pre-trained segmentation models on real images

2. What is Image Segmentation?

Image segmentation is the process of partitioning an image into meaningful segments or regions. Unlike object detection, segmentation assigns a label to each pixel.

Semantic vs. Instance Segmentation

• Semantic Segmentation: Classifies each pixel into a category (e.g., person, road).
• Instance Segmentation: Differentiates between individual objects of the same category (e.g., person 1, person 2).

3. Key Models and Architectures

• Fully Convolutional Networks (FCN): Pioneered semantic segmentation using CNNs
• U-Net: Widely used in medical imaging with encoder-decoder structure
• DeepLab v3+: Combines atrous convolutions and spatial pyramid pooling
• Mask R-CNN: Extends Faster R-CNN for instance segmentation

4. Applications

• Autonomous driving (lane and road detection)
• Medical image analysis (tumor segmentation)
• Remote sensing (land use classification)
• Interactive photo editing

5. Lab: Run Semantic Segmentation using DeepLabV3

from torchvision import models
import torchvision.transforms as T
from PIL import Image
import matplotlib.pyplot as plt

# Load a pretrained DeepLabV3 model
model = models.segmentation.deeplabv3_resnet101(pretrained=True).eval()

# Load and preprocess image
img = Image.open("your_image.jpg")
transform = T.Compose([T.Resize(520), T.ToTensor()])
input_tensor = transform(img).unsqueeze(0)

# Run inference
with torch.no_grad():
    output = model(input_tensor)["out"][0]
segmentation = output.argmax(0)

# Display result
plt.imshow(segmentation)
plt.axis('off')
plt.show()

6. Assignment