Please enable JavaScript.

Coggle requires JavaScript to display documents.

Computer Vision - Coggle Diagram

- - - - Physical Model: The Eye
      - Experimental Model: The Signals
      - Psychological Model: The Brain
  - - - Industrial Vision
        Vision and sensor systems used for control and operation in factories.
      - Robot Vision
        The vision a robot uses to understand and navigate its environment.
      - Vision in the Wild
        Vision systems that can be used in miscellaneous scenarios, such as ANPR, recognition, image interpretation.
    - - Image Acquisition Hardware
        What it says on the tin. Image quality is important and that means more than just megapixels.
      - Computer Vision Software
        The mushy brains behinds the eyes that interprets the signal.
    - - Robustness
        
        The vision system must be robust to changes in its environment, i.e. lighting, sensor noise, defocusing, camera position.
        
        What you want
      - Repeatability
        
        The system must work the same over and over, regardless of changes in the environment.
        
        What you want
      - Invariance
        
        Invariance to environmental factors helps achieve robustness and repeatability.
        
        Types of invariance include scale, rotation, illumination.
        
        What you design your system to be
      - Constraints
        
        Limitations enforced on hardware, software, and wetware to make the vision system work in a repeatable, robust way.
        
        What you apply to make it work.
    - - Software
        
        Hough Transform
        
        Intelligent usage of colour
      - Physical
        
        Environment
        Lighting, enclosure, mounting
        
        Acquisition Hardware
        Lenses (focal length, sharpness, aperture), Filters (polarising), Sensor (infra red)
      - "Apply as many hardware and wetware constraints as possible, and let the software take up the slack."
    - - RGB (Red, Green, Blue)
        
        Couples brightness (luminance) with each channel
        
        Illumination invariance is difficult
      - HSV (Hue, Saturation, Value)
        
        Hue encodes the pure colour as an angle
        
        Saturation is how vibrant the colour is
        
        Value encodes brightness
        Lighting invariance can be achieved using just H or H&S components.
- - - - Applications
        
        Understanding an analysis
        
        Speeding up algorithms
        
        Representation (infariance)
        
        Codign
        
        Recognition
  - - - Brightness histogram
      - Ordinary vs Cumulative
      - Intensity Mappings
        
        Copy
        
        Inversion
        
        Addition
        
        Multiplication
        
        Log
        
        Exponent
    - - Intensity Normalisation
      - Intensity Equalisation
        Improves the contrast
    - - Manual
      - Automatic
        
        Otsu Thresholding
        Maximises interclass variance
        
        Adaptive Thresholding
        
        Start with the mean
        
        Iteratively adjust to be the mean of the means of the two segments
  - - - Vertical
      - Horizontal
      - Vertical + Horizontal
      - Prewitt
      - Sobel
      - Canny
        
        Interpolation in Non-maximum Suppression
        
        Hysteresis
    - - Laplacian
      - Laplacian of Gaussian
      - Marr-Hildreth
        Zero Crossings of the LoG
  - - - Idea
        Use a known template and match to see if occurs inside image.
      - Terms
        
        I = Image
        
        T = Template
        
        Omega = Local search area
        
        Accumulator Space = Tracks number of matching pixels by brightness between template and omega.
      - Properties
        
        Optimal in occlusion
        
        Optimal in noise, but slow
      - Process
        
        We use the correlation operator rather than convolution as we are matching the template, not applying it..
        
        Sliding window approach
      - Template Matching with Fourier
        
        Motivation
        Template matching is slow so we use the FFT
        
        Cost
        2xFFT plus multiplication
        
        Process
        
        Start with the image
        
        Flip and pad the template
        
        Fourier transform of template
        
        Fourier transform of image
        
        Multiply transforms
        
        Extract location from the result of multiplication
    - - Linear Hough
      - Circular Hough
      - Conic Hough
      - Generalised Hough
- - - - Extraction
      - Matching
      - Segmentation
        Partitions the image into sets of pixels called partitions.
        
        Methods
        
        Global Binary Thresholding
        Divides grayscale images into two segments using a threshold value.
        
        Upsides
        
        Really fast
        
        Can work well in applications with lots of physical constraints such as high contrast objects, or lighting control.
        
        Downsides
        
        Requires a manually set static threshold
        
        Not robust to lighitng changes
        
        Otsu's Thresholding Method
        An automatic method to find the threshold for global binary thresholding.
        
        Assume there are two classes (i.e. background and foreground)
        
        The histogram must have two peaks
        
        Exhaustively search for the threshold betwen 0-255 that maximises interclass variance.
        
        Adaptive/Local Thresholding
        
        Mean Adaptive Thresholding
        
        K-Means Segmentation
        
        Advanced Techniques
      - Types
        
        Global
        Extracted from contents of entire image.
        
        Grid/Block
        Splits image into fixed blocks and extracted one feature from each.
        
        Region-based
        The image is split into kind of regions, as a feature is extracted from each. Differs from grid as regions are more general.
        
        Local
        Interest points are detected and a feature is extracted from the surrounding pixels of each.
      - Joint-Colour Histogram
        A 3D, 3-channel histogram of the colours quantised into bins and accumulated.
        
        Normalisation allows the histogram to be size invariant
        
        Rotation invariant
        
        Choice of colour space can make it invariant uniform to lighting change
        
        Howeverm vastly different images can have the same histogram
      - Connected Components
        
        Pixel Connectivity
        
        Connected Component
        
        Connected Component Labelling
        
        Two-Pass Algorithm (Union Find)
    - - Shape Decriptions
        
        Region Descriptions
        
        Moments
        
        Normalised Central Moments
        
        Hu Moments
        
        Zernike Moments
        
        Central Moments
        
        Standard Moments
        
        Simple Scalar Shape Features
        
        Compactness
        
        Centre of Mass
        
        Irregularity / Dispersion
        
        Perimeter
        
        Area
        
        Boundary Descriptions
        
        Chain Codes
        Can be made cale and rotation invariant
        Not very good for shape matching
        
        Multiple Shapes
        
        Region adjacency graphs
        
        Flexible Modelling
        
        Point distribution model
        
        Procrustes Analysis
        
        Generalised Procrustes analysis
        
        Active Shape Models (ASMs)
        
        Constrained Local Models (CLMs)
        
        Active Appearance Models (AAMs)