Perceptra 2-D Test Run – 1, 1st Jan 2015
The image below shows the 2-D creation of the JND edges for hand drawn digit #61. It has two errors in it. One is the two right side vertical edges that are one above the other at pixel 30 and 38. The other error is the missing bottom horizontal edge for pixel #38. These two errors occurred due to recognizing edges between pixels using the IDL to provide the JND between adjacent pixels. The values of the readings are:
4 14 13 Log Values = 7.60 14.47 14.07
16 16 16 15.21 15.21 15.21
14 12 1 14.47 13.63 0.00
There is no noticeable difference between 14 and 13 or 14 and 16 but there is between 13 and 16. This gives a horizontal edge between the 13 and 16 that sticks out in the middle of the 2x2 group of pixels. Similarly there is no noticeable difference between 16 and 14 and 14 and 12 but there is between 12 and 16.
The approach I am using is to create all these edges and then start combining them at the corners where they touch. Any two edges then produce a corner of a particular type with a length ratio. If the same area made up of corners is then found of a different size the ratios of the lengths would be the same. This would provide for recognition independent of size.
After this, areas that touch each other would be combined in a hierarchy.
However this process requires the formation of different types of things explicitly in the structure. Maybe the better approach is to just combine edges at all levels of complexity and let the tree sort out which ones form objects based of coincidence.
I’m not sure how this will result in recognition of a shape independent of size.