CheckParam.cpp is a program that I wrote to check the efficiency of an edge detection algorithm. The edge detection algorithm (which I did not write) inputs a PGM image and a set of 3 parameters and outputs an image file of the edges.
CheckParam.cpp inputs an image and the three parameters and outputs the edge file as well as a score from 0-100 of how efficienty the edge detection worked with that set of parameters. The three parameters are the diameter of the filter, and 2 coefficients for the gaussian function: the center and the surround.
The best edge detection output images have a high contrast between edges and the surrounding pixels. An ideal image will have only pixels that are perfectly either black or white. Therefore, I think the best way to test the image is to look at every given pixel and measure:
- The strength of the color
- Compare it to the surrounding pixels The reason I am not measuring the strength of color alone is because, while it will indicate a potentially correct distribution of color, it does not do enough to actually check whether you've found edges.
Theoretically, if only edges are marked on an image, there should only be lines of white pixels. This means that an ideal pixel will be perfectly white with two neighboring pixels also perfectly white. Every other pixel that touches it should be perfectly black. In order to do this measurement, the algorithm will iterate across the image, checking each pixel and giving it a score. To score a pixel, I will compare the saturation of the color (so a perfect score will be either a pixel being either entirely black or white, with a score of 0 being perfect grey), and then if it is a white pixel (so an edge) by seeing how many neighboring pixels are also edge pixels. Ideally, some of the edge pixels will touch other edge pixels, but if more than half of the neighboring pixels are also white then it failed to properly select an edge. That means it just found a range.
This algorithm will do a good job selecting edge pixels for most images assuming the image compression is at least somewhat functional, however it will do a less than ideal job when multiple edges converge at one location or if the image compression is so broken that the resulting image is in no way reminiscent of the original. Say if you were to perform an edge detection on an asterisk, the center pixel would receive a lower score even if the algorithm was perfect as it would be marked down for having so many edges touching. Since this should not happen too much in a typical image I think this corner case will not be triggered too much, but if you have an image with tons of overlapping edges it may not receive an ideal score.