Skip to content

Object Detection using multiple image manipulation techniques

Notifications You must be signed in to change notification settings

jpthanga/Object_Detection

Repository files navigation

Detecting Objects

To run the application, run make command on part2. Then run ./detect <filename of image>.

This will create three files:

  1. edges.png - Contains the output of the sobel edge detector.
  2. detected.png - Contains Input Image with boundaries around the ICs
  3. detected.txt - Description of boundaries in the required format.

Algorithm:

  • The black and white image is take and sibel detector is applied to find the edge points.

  • Lines passing through these points are found using Hough Transform.

  • The corners on these lines are found using Harris corner detector.

  • All the rhombi formed with the corner points are considered.
  • The image between the image is clipped and compared in the Fourier space with test images(present in repository) of IC's.
  • The boundaries are decided based in the score from the comparison.

Assumptions:

  • The algorithm assumes that the IC board with be parallel to the edges of the image.
  • Only the vertical and horizontal lines are comsidered for corner detection.

Test Results:

  • The algorithm struggle on IC's with a dimension of less that 32 px is one or more directions.
  • Accuracy: The algorithm detects nearly all big square IC's and struggles a little on the rectangualr ones.

Future Improvements:

  • Non Maximal Suppression of boundaries need to be implemented, edges and lines are implemented.
  • More varied sample test images needs to be used for ideal accuracy.