Today we have three main categories:

• Text based Image Retrieval (TBIR)
  • Images are annotated manually
  • We can apply any of the text retrieval approaches
  • Important to note that words are not enough to carecterise an image
  • It’s hard to manually tag big collections
• Based on content (Content-Based Image Retrieval - CBIR)
  • Based on its own content and not “tags”
  • The content is extracted from the image (colour, forms, and etc)
  • User searches with an image and the system returns similar images
  • The content of the image is the “features” of the image
• Based on semantic (Semantic-Based Retrieval - SBIR)
  • Normally first you extract low-level info
    • And extract the semantic information from some regions

Colours

Channels

Note: We can use the 2 most significant bits on each channel and merge them into one variable (which will give 64 possible colours for RGB)

RGB: additive colours (common in monitors, based on light) CMYK: subtractive colours (common in printers, based on pigments) HSV: Hue, Saturation, and Value; H = colour, S = quantity of gray; V = colour intensity

Feature Vectors

• Examples: Colours histogram, Colours moments

Image descriptors

• A tuple of functions that extracts the feature vectors and another function that computes the distance of two feature vectors

colour coherence vector (CCV)

• https://tarekmamdouh.wordpress.com/2013/08/19/image-retrieval-color-coherence-vector/

Border Interior Pixel Classification (BIC)

• Classifies a pixel either as border or “interior”. (vizinhança 4)
• It normally uses dLog to compute the difference between two vectors

Textura

• Texturas finas: grandes variações
• Texturas ásperas: mais homogêneas
• Permite identificar projeções 3D

Gray Level Co-occurence Matrix

• Haralick Descriptors (1974)

Local Binary Pattern

• https://en.wikipedia.org/wiki/Local_binary_patterns
• Tenta encontrar textura por meios de histogramas

Descritores de forma

High level descriptors

• Objects, people, animals, and so on

Global features

• Colour histograms
• Textures

Local features

describes different keypoints of the image

• Scale Invariant Feature Transform (SIFT)
• LEARND Invariant Feature Transform (LIFT)

They don’t change based on the colour or the point-of-view, you can use it to match the keypoints in two images.