Face recognition verification using fisherfaces algorithm
I am working on Face recognition as a part of my project. I am using the fisherfaces program provided in OpenCV website for the same. But the program code provided there does not have verification and the recognition always results in a subject in the database even though it is a false subject. The fisherfaces reconstruction does not perform well. The Similarity error for both the true test inputs (true test input - picked from the database) and false test inputs (false test input - NOT picked from the database) are quite close to each other and some even overlap with the range of values from true test inputs.
I performed an experiment on AT & T database which has 40 subjects.
I selected 20 out of them for database and rest 20 i used to provide false test input where the ideal behavior of false test input is not to get recognized to one of the images in database. But the normalized euclidean distance of the false test inputs are quite near and some even overlap with the normalized euclidean distance of the true test inputs. Please let me know if there is any method to set an appropriate threshold and hence make the algorithm work well.
Results for false_test_inputs
Identity: 0. Similarity error: 0.357048 Identity: 0. Similarity error: 0.364141 Identity: 0. Similarity error: 0.368228 Identity: 0. Similarity error: 0.322856 Identity: 0. Similarity error: 0.298744 Identity: 0. Similarity error: 0.3978 Identity: 0. Similarity error: 0.354158 Identity: 8. Similarity error: 0.417423 Identity: 0. Similarity error: 0.42364
Results for true_test_inputs
Identity: 8. Similarity error: 0.291803 Identity: 8. Similarity error: 0.381701 Identity: 8. Similarity error: 0.455005 Identity: 8. Similarity error: 0.378799 Identity: 8. Similarity error: 0.480244 Identity: 8. Similarity error: 0.456196 Identity: 8. Similarity error: 0.419657 Identity: 8. Similarity error: 0.361412
I can mail the complete list if in case it is needed. I am not posting the similarity errors of all 100 true and false test samples as it will not be clear in viewing them in the forum.
I can attach a plot of the samples.At_T.png