The Information & Mutual Information Ratio for Counting Image Features and Their Matches
Presented at the 8th Iran Workshop on Communication and Information Theory(May 2020), Sharif University of Technology, Tehran, Iran.
Arxiv version: https://arxiv.org/abs/2005.06739
The IEEE Xplore webpage: https://ieeexplore.ieee.org/document/9163458
A. K. Mirabadi and S. Rini, "The Information & Mutual Information Ratio for Counting Image Features and Their Matches," 2020 Iran Workshop on Communication and Information Theory (IWCIT), Tehran, Iran, 2020, pp. 1-6, doi: 10.1109/IWCIT50667.2020.9163458.
-
Image joint histogram or 2D histogram, imhist2D.m:
The joint histogram of two given images can be easily computed through h_ij=imhist2D(X1,X2), where X1 is the first frame and X2 is the second frame. -
IR: the estimated count of image feature points, ir.m:
To estimate the number of existing feature points (count) in a given image, it is enough to call count=ir(h_i), where h_i is the image histogram. -
MIR: the estimated count of joint feature points of two frames, mir.m:
To estimate the number of existing joint feature points (jcount) in two given frames, it is enough to call jcount=mir(h_ij,h_i,h_j), where h_i, h_j, and h_ij are the 1st frame, 2nd frame, and joint(2D) histograms, respectively. -
The lower bound on IR (LIR); MATLAB function, lir.m:
This is a lower bound on the IR estimator(feature) and depends on image entropy and size. (Apply the same instruction as the ir function.) -
The lower bound on MIR (LMIR), lmir.m:
This is a lower bound on the MIR estimator(feature) and depends on image mutual information and size. (Apply the same instruction as the mir function.) -
Contiguous intensity merging for 1D histogram, bincom.m:
In a simple sentence, this is a thresholding function. The thresholded histogram(out_hist) can be computed through out_hist=bincom(h_i,d), where h_i is the histogram of a given image and d is the feature distance or threshold level. -
Contiguous intensity merging for 2D histogram, bincom2D.m:
This function is the 2D version of bincom.m. It gets the 2D histogram (h_ij) and feature distance(d) as its inputs and returns the thresholded 2D histogram(out_2Dhist) through out_2Dhist=bincom2D(h_ij,d). -
Image entropy and image mutual information, ShEntropy.m & mutualInf.m
The IR-based Optimization method, Koptimizer.m
The number of extracted feature points from an image can be boosted by this algorithm (How? I invite you to read the paper). The function Koptimizer gets an image and a feature distance (usually a number ranging within 6 to 9, d=8 is preferred). Thus, let image X and call K_o=Koptimizer(X,d), and then you will have a coefficient K_o which should be multiplied by the image (X): Y=K_o * X.
Now, you have an image that guarantees feature extraction algorithms can extract more features when using it than when using the original image. The improvement rate is roughly 15% to 35% on average, of course, depended on the utilized feature extraction algorithm.
The paper mainly discusses handcrafted feature points, but you can use this optimizing method for ML or DL algorithms as well (Why? because features are features :))
Good Luck!