Evaluation of performance fruit detection algorithm and training the algorithm

I am trying to find a way to evaluate the performance of my fruit detection algorithm. The detection of the fruit is currently done with the circular Hough transform. How can I compare this result with the original RGB image and therefore check if the detected fruit are a true positive or a false positive (to calculate the accuracy)? And how is it possible to train the algorithm?

I have added the original RGB image, the black-white image after preprocessing and the parameters of the CHT algorithm.

[centers,radii, metric] = imfindcircles(BW_Canny,[Rmin Rmax],'Sensitivity',0.95,'EdgeThreshold',0.15); % 0.95 0.15
% d = imdistline; %To check manually the approximate radius of the fruit in the image
viscircles(centers, radii,'EdgeColor','b');

For example for image 41, there are 3 true positives and 1 false positive.

[yanqi liu]

yes，sir，may be compute IOU、MAP for image object detection

Answer 1

[yanqi liu]

in object detect application，may be use evaluateDetectionPrecision to compare detect location and real location，such as

https://ww2.mathworks.cn/help/releases/R2021b/vision/ref/fasterrcnnobjectdetector.html