Contour Detection

• How do Image Processing libraries like OpenCV perform tasks such as contour detection?
• The grid (filter) on the left adds up to zero.

• When a Spatial Filter is applied to any 9 cells of an image, it shows how much the area around the center coordinate has changed.

  • In other words, the filter acts as a weight.

• If all the values are the same, the sum will be zero. If there is a significant difference between the values in the top, bottom, left, and right cells compared to the center cell, the sum after weighting (60 in the above figure) will be a large value.

• http://www.mis.med.akita-u.ac.jp/~kata/image/sobelprew.html

  • First-order difference filters like Sobel
  • Second-order difference filters like Laplacian

• If contour enhancement is too strong, it may also emphasize noise.

  • Therefore, blurring with a Gaussian filter is often done before contour detection.

• For more details, refer to the video below: https://www.youtube.com/watch?time_continue=259&v=OpcFn_H2V-Q&feature=emb_logo

---

• CNN is a method that automatically learns this filter.
• It automatically generates multiple filters, and the number of these filters is the number of Convolutional Layers.
  • One filter may be able to detect dog ears.
  • Another filter may detect dog eyes.

#image#udacity_intro_to_deep_learning_with_pytorch