Computer Vision

Human Vision

Retina captures image and tuns it into a electrical signal. It then travels up the optic nerve and gets processed in the visual cortex. Frontal node adds other information

Computer Vision and how it captures Image

A camera or sensor captures an image and turns it into pixels. Image is processed into features, and algorithm processes signal. Algorithm output determines decisions. 

What are Decisions based on for CV

Based on algorithms, pattern matching and trained models

Human Vision Decision making based on

On past experiences, context and instinct

What was used after Multilayer/Feed Forward Models

Modern Convolutional Neural Network

What are challenges in CV?

Image Quality, Variability(Different angles, lighting

conditions, scale, and occlusions), Motion, Scale, Bias

Feedforward old method of CV 

flatten images, looses spacial awarness

Train a NN

1. Forward pass

2. Calculate the error

3. Propagate the error backwards

4. Update the weights

5. Repeat until some stopping criterion

Batches

Smaller Subset of data that gets passed through. An epoch is still only complete once all data has been through model

Learning Rate

It dictates how much we adjust the weights according to the error

signal

Convolution Neural Networks

Input and output layers (just like in MLPs)

• Convolutional layers

• Pooling layers

• Dense (fully-connected) layers (these are what we saw in MLPs)

• Pooling Pooling

Filter/Kernals

feature detection, builds ups. It chooses the best features by itself. 3D filters for RGB

Padding

Extra pixels at the edge usually 0 so that edges are treated equally to middle pixels.

Max Pooling

Take the maximum value, for a given window

Stride length

the amount the filter moves by. Less is more accurate but more computationally expensive.

Pooling layers meaning

Reduce Dimensionality, prevents overfitting and translation invariance

Dense/Fully-Connected Layers

Connect all layers together

• The same as our standard MLPs/feed-forward neural networks

• Flatten the feature maps from the last pooling layer, then feed

through dense layers

• Connects all neurons from the previous layer to the output

node(s) to produce a final decision/classification.