I am wondering if anyone has suggestions for metrics to use to quantify the edge strength and contrast of a target from its background in single channel images. I am trying to find the effectiveness of filtering methods in increasing object visibility of FLIR A310 thermal images for an eventual object detection algorithm.
I have tried using metrics suggested in this paper which were specifically created for FLIR A310 images:
M1: (mean_target - mean_background)/(mean_target + mean_background)
M2: (mean_target - mean_background)/((mean_target + mean_background) * 2 * min)
M3: (mean_target - mean_background)/(mean_background)
M4: (stdev_target - stdev_background)/(stdev_target + stdev_background)
(as well as a few others that used total image means and stdevs)
But they all (other than M4 which had too many false positives) show the filters as performing worse than the unfiltered image, which is against intuition since some of the filters (anisotropic, bilateral, Haar discrete wavelet transform) are widely accepted as effective at improving structure visibility, and upon visual inspection targets are more visible in filtered images.
I have looked at the SSIM , and adapted edge aware WSSI (Weighted Structural Similarity Index) but they don't work work in this scenario since we expect a change and structure and want to know if that change in structure results in stronger edges and higher contrast.
Here are some example images:
(Unfiltered, Rolling Guidance Filter)
(Bilateral, Anisotropic)
Where the white blob is the target. Any metric suggested are greatly appreciated. Thanks!
