Skip to content

Latest commit

 

History

History
34 lines (26 loc) · 1.37 KB

File metadata and controls

34 lines (26 loc) · 1.37 KB

Features

CNNs compare images piece by piece. The pieces that it looks for are called features. By finding rough feature matches in roughly the same positions in two images, CNNs get a lot better at seeing similarity than whole-image matching schemes.

Convolution

  • CNN will try to match the feature everywhere and in every possible position.
  • In calculating the match to a feature across the whole image, we call it a filter
  • to calculate the match of a feature to a patch of the image, simply multiply each pixel in the feature by the value of the corresponding pixel in the image
  • then add up the values and divide by the total number of pixels
  • to complete convolution, repeat the process, lining up the feature with every possible image patch
  • values close to 1 show strong matches, close to -1 indicate photographic negative of our feature, 0 means no match

Pooling

  • way to take large images and shrink them down while preserving the most important information
  • stepping a small window across an image and taking the maximum value from the window as each step.

Fully Connected Layers

  • take high level filtered images and translates them into votes
  • we have decide between X categories
  • some values are better than others in classifying a particular class, these get larger votes than others
  • these votes are expressed as weights or connection strengths