Introduction to Convolutional Neural Networks (CNNs).pptx
- 2. Basic Structure of CNN • Input Layer: Accepts input images as pixel data. • Convolutional Layer: Applies filters to extract features. • ReLU Layer: Introduces non-linearity to the network. • Pooling Layer: Reduces spatial dimensions of feature maps. • Fully Connected Layer: Final layer for classification.
- 3. Convolutional Layer • Filters/Kernels: Detect specific features in input images. • Stride: Controls the movement of filters across the input. • Padding: Adds pixels around the input to maintain dimensions. • Output: Produces feature maps indicating detected features.
- 4. Padding in CNN • Zero Padding: Adds zeros around the input image to preserve dimensions. • Valid Padding: No padding, reduces the size of output feature maps. • Role: Helps preserve edge information during convolution.
- 5. Pooling Layer • • Purpose: Reduces dimensionality and computation in the network. • • Max Pooling: Selects the maximum value from each pooling region. • • Average Pooling: Takes the average value from each pooling region. • • Impact: Retains important features while reducing overfitting.
- 6. Basic Mathematics of CNN (B&W Image) • • Convolution: Applies a filter matrix across the image to detect features. • • Example: Sliding a 3x3 filter over a grayscale image, producing a feature map. • • ReLU: Applies non-linearity after convolution. • • Pooling: Reduces the size of the resulting feature map.
- 7. Basic Mathematics of CNN (Colored Image) • • Convolution: Applies the same filter across each RGB channel. • • Result: Produces a combined feature map from all channels. • • Example: Sliding a filter across an RGB image and summing up feature maps. • • Pooling: Reduces the size of the resulting feature map while preserving important information.
- 8. Fully Connected Layer • • Purpose: Flattens the output and connects to a fully connected layer. • • Function: Combines features for final classification. • • Uses: Softmax or sigmoid activation functions for output.
- 9. LeNet-5 Architecture • • Designed for handwritten digit recognition (MNIST dataset). • • Structure: 2 convolutional layers, 2 subsampling layers, 2 fully connected layers. • • Key Feature: Simple and efficient, early CNN model.
- 10. AlexNet Architecture • • Winner of the ImageNet competition in 2012. • • Structure: 5 convolutional layers, 3 fully connected layers. • • Features: Uses ReLU, dropout, and data augmentation. • • Impact: Revolutionized deep learning and computer vision.
- 11. VGG-16 Architecture • • Uses 16 layers (13 convolutional, 3 fully connected). • • Features: Smaller filters (3x3) with deeper networks. • • Strength: Achieves high accuracy with a simple structure.
- 12. ResNet Architecture • • Introduces Residual Learning to combat vanishing gradients. • • Structure: Skip connections or shortcuts between layers. • • Impact: Allows very deep networks (e.g., ResNet-50, ResNet-101).
- 13. Inception (GoogLeNet) Architecture • • Introduces Inception modules: parallel convolutional filters. • • Structure: Multiple filter sizes (1x1, 3x3, 5x5) in parallel. • • Impact: Efficient and scalable for large- scale image recognition.
- 14. Transfer Learning • • Concept: Uses a pre-trained model on a new but related task. • • Benefits: Speeds up training, requires less data, and improves performance. • • Example: Using a pre-trained model like ResNet for a new image classification task.
- 15. Object Localization • • Purpose: Identifies the location of objects within an image. • • Methods: Bounding box regression, Region Proposal Networks (RPNs). • • Applications: Object detection, image segmentation.
- 16. Landmark Detection • • Definition: Detects specific key points or landmarks within an image. • • Applications: Facial recognition, medical imaging (e.g., key anatomical points). • • Methods: CNNs used to detect and regress the position of landmarks.
- 17. Conclusion • • CNNs have revolutionized computer vision tasks. • • Architectures like LeNet, AlexNet, VGG, ResNet, and Inception paved the way for modern image processing. • • Transfer learning, object localization, and landmark detection expand the versatility of CNNs.