Day #154 - CNN - Class Notes

Key Summary

• 540 million years to trace evolutions of vision
• Human vision is trained for 540 million years
• Hierarchy of layers in our vision are involved in processing

What Computers see

• Images are numbers
• Pixes represented by 2D array of numbers
• RGB (3D Array)
• Computer vision Tasks
• Regression (Output takes a continuous value)
• Classification (Single Class label)
• Detect presence of features in particular image

Manual Feature Extraction

• Domain Knowledge
• Define features
• Detect features and classify

Image Challenges

• Occlusion
• Viewpoint variation
• Scale variation
• Deformation
• Background Clutter
• Intra Class variation
• Illumination Conditions

Neural Networks

• Learn directly from Image data
• Low Level (Edge / Dark Spots)
• Mid Level (eyes, Ears, Nose)
• High Level (Facial Structures)

Fully Connected Neural Network

• Multiple Hidden Layers
• Input 2D Image (Vector of pixel values)
• All spatial information will be lost
• Connect neuron in hidden layer to all neurons in input layer
• Slide patch window across the image, this considers spatial structure
• Apply set of weights to extract local features
• Multiple filters and multiple set of weights
• Patchy Operation known as convolution

Feature Extraction and Convolution

• Convolution preserves spatial relationship between pixels
• Elementwise multiplication between patch and filters
• Different filters for Sharpening, Edge
• Use multiple filters to extract different features

CNNs for Classification

• Convolution - Apply filter with learned weights to generate feature maps
• Non-Linearity - Often Relu (Image data highly non-linear)
• Pooling - Downsampling for each feature map
• Train model to learn weights
• Each Neuron sees patch of inputs
• Apply matrix of weights for elementwise multiplication
• depth = number of filters
• Relu - Pixel by pixel operation that replaces all negative values by zero (Non-Linear operation)
• Pooling - Reduce dimensionality preserve spatial invariance (Downsampling operations)
• Layer operations to learn hierarchy of features
• Feature Learning Pipeline + Performing Classification

Imagenet CNN

• 14 million Images
• 21,841 categories
• Deeper Network vs How deep we can go

Architecture for Applications

• New architecture beyond Feature Learning
• Semantic Segmentation (Fully Convolutional Network) - Downsampling and Upsampling operations, Driving Scene Segmentation, Encoder-Decoder
• Object Detection - Region Proposals / Classify them, Really long time to compute
• Image Captioning - Generate Semantic Content - Remove Fully Connected layer and replace them with RNN
• CNN feature Layer + RNN (Trained to predict words that describe the image)
• CAM (Class Activation Map)

Happy Mastering DL!!!