• Review of Region Filling and extractable features for recognition:
  • Area: sum all of the pixels belonging to a region
  • Centroid: sum all pixel locations in x and y then divide by area
  • Symmetry; check for equal pixel locations on each side of the centroid
  • Euler Number: number of "holes" in the object
• Convolution:
  • A mask is an arrangement of numbers which is used to filter an image
  • To apply a mask to an image, it is convolved with the image by multiplying the mask pixelwise with the image, summing the products, and placing this result as a pixel in a new image.
  • Into which pixel the result is placed needs to be determined and kept consistent.
  • 2D Convolution
  • You may need to pad the image at the edges in order to finish the convolution. This can be acheived by padding with zeros or with the edge pixel value.
  • In digital signal processing, the mask is actually flipped around when applying convolution. In the world of robotic vision, it is generally accepted that the mask is applied as is.
• Neural Nets

Convolution examples
16-311 General Robotics
Prof. Howie Choset

In examples below, red signal is the original, blue is the output from convolution, and green is the mask used in convolution.

Convolution with edge detector masks

Convolution with gaussian distribution

Edge detection on a signal that was convolved with a gaussian