110 likes | 124 Views
Explore the recognition of faces across various levels of watershed flooding using a face recognition method. Learn how different flooding levels impact the accuracy of face recognition through distance analysis and eigenvector calculations.
E N D
Overview • Question: At what level of flooding is a face recognizable? • Experiment: Use a face recognition method and apply it to the levels of flooding.
Recognition • I used a face recognition implementation from Drexel University. It is done in Matlab and the code is released for use or modification. • I made minor modifications to the code to suit the experiment.
Recognition Dataset Preparation • Obtain training set of faces • Create a vector for each image • Find the mean image • Find the difference between each image and the mean image • Calculate a set of eigenvectors that best describe the distribution of the differences
Recognition Recognition Procedure • Obtain the eigenface for test image using the steps listed previously • Find the dataset image with eigenvectors that best describe the test image by minimizing the distance between the two classes.
Experiment 1. Three face images are used as the dataset. 2. Create temporary images for each level of watershed flooding for one of the dataset images. 3. Use flooding images as test images for recognition algorithm. 4. Plot the distance values to visualize how well the recognition performed at different levels of flooding.
Experiment 1 2 3 Images used for the dataset and testing
Experiment Plot shows the distance between eigenvectors of image 1 at the current flood level and the original image. A smaller distance can be considered better recognition.
Experiment Distance plot for images 2 and 3. Results show that after level 40 the recognition improvement is proportional to the level increase.
Experiment Plot shows the distances from each dataset image to the flood levels of image 1. If the minimum distance were used, after level 65 recognition would be 100% Level 1 Level 25 Level 65 Level 256
Conclusion This is just a small experiment that I did based on what you mentioned in our last meeting. One thing I was curious about is how to define recognition. I don’t know exactly how trustworthy this measure is compared to a human. If I have time I might try to learn more about the actual eigenvector calculation and distance measure.