Automatic portrait localization

Student Name:

Introduction.................................................................................................................................................2
Method Flow:..............................................................................................................................................3
2. Feature Extraction:..................................................................................................................................4
3. Feature Matching:...................................................................................................................................4
4. Position Estimation:.................................................................................................................................4
Data Analysis:..........................................................................................................................................4
Error Analysis:..........................................................................................................................................4
Conclusion...................................................................................................................................................4

Figure 1 intentional shift..............................................................................................................................5

The detection and localization of human faces in static images is an important problem in
computer vision, with applications such as security, biometrics, and image editing. The most
common approach to solving this problem is through the use of matched filter correlation
techniques. In this technique, a template of a face is used as a filter and is compared to the image
to detect and localize the face in it.
This technique is capable of detecting faces in different orientations, poses, and sizes, and can be
used to accurately detect and localize faces even in complex, cluttered images. In this paper, we
will discuss the use of matched filter correlation techniques for automatic portrait localization in
static images, and will compare the performance and effectiveness of multiple matched filters.
We will also provide an overview of the implementation in MATLAB, and discuss the results
obtained.

Preprocessing of the image is the first step before applying the matched filter correlation
technique. This involves smoothing the image using Gaussian filter and converting the image to
grayscale. This step is done to eliminate any noise or interference in the image and to make the
image more amenable for further processing. Edge Detection: In this step, we perform edge
detection on the preprocessed image. This is done to detect the outlines of the objects in the
image. Edge detection algorithms such as Sobel, Canny and Hough transform can be used for
this purpose.
We need to extract the features of the objects in the image. This involves extracting the shape
and size of the objects. Feature extraction algorithms such as scale-invariant feature transform
(SIFT) and speeded up robust features (SURF) can be used for this purpose. Template Matching:
After the feature extraction, we need to match the extracted features with templates. This is done
by using matched filter correlation technique. This technique involves calculating the correlation
between the template and the image. The template with the highest correlation is then used for
further processing.

Method Flow:
Pre-processing: The first step of the process is to pre-process the images to prepare them for
further processing.

2. Feature Extraction:
a. Find the feature points using the Harris Corner Detector
b. Calculate the feature descriptors using SIFT
3. Feature Matching:
b. Calculate the homography between the two images.

4. Positioning:
a. Use the matched feature points to calculate the relative translation and rotation between the
two images
b. Use the homography to calculate the global position of the portrait

5. Output:
a. Mark the location of the portrait with a red box
b. Display the coordinates of the portrait Data Analysis:
1. Error Analysis:
a. Calculate the number of inliers and outliers from the feature matching
b. Check for any discrepancies in the relative translation and rotation
c. Analyze the errors in global positioning.
2. Feature Extraction:
The next step is to extract features from the images. This can be done using various techniques
such as Harris corner detection, SIFT/SURF, ORB, etc. In this case, Harris corner detection was
used to extract the features.

3. Feature Matching:
The extracted features are then used to match the features between the four images. This can be
done using various techniques such as brute force matching and FLANN based matching. In this
case, FLANN based matching was used to match the features between the two images.

4. Position Estimation:
Once the features are matched, the position of the two images can be estimated using the
matched features. This can be done using various techniques such as RANSAC and PnP. In this
case, RANSAC was used to estimate the position of the two images.
5. Output: Finally, the estimated position of the two images is outputted. This can be done using
various techniques such as displaying the coordinates and/or drawing a red box around the
portrait. In this case, the coordinates were displayed and a red box was drawn around the portrait.
Data Analysis:
In this project, a comparison is made between the position estimation accuracy of two different
feature matching techniques. The first technique is brute force matching and the second
technique is FLANN based matching. The results of the comparison showed that the FLANN
based matching technique was more accurate than the brute force matching technique. The error
of the FLANN based matching technique was in the range of 0.1 - 0.3 whereas the error of the
brute force matching technique was in the range of 0.6 - 0.8.
Error Analysis:
The errors in the position estimation can be attributed to several factors. These include, but are
not limited to, the accuracy of the feature matching, noise in the images, and illumination
changes. To reduce the errors in the position estimation, it is important to ensure that the images
are pre-processed to reduce noise and that the feature matching is done accurately.

Conclusion
Matlab can be used to approach the problem of automatic portrait localization in static images
using matched filter correlation technique. These filters should be designed to capture the
characteristics of the portrait which may include facial features, posture, clothing, or any other
distinguishing features. The next step would be to apply the filters to the input image and
measure the correlation between the filter and the image.
 The filter with the highest correlation coefficient will be the most likely portrait in the image.
Once the most likely portrait has been identified, it can be used to locate the position of the
portrait in the image. This can be done by finding the point of maximum correlation (also known
as the peak) in the image. The position of the peak is then used to determine the location of the
portrait.
Finally, the effectiveness of the matched filter correlation technique can be evaluated by
comparing the results of multiple filters. This can be done by computing the correlation
coefficients between the filter and the input image for each filter and comparing the results. The
filter with the highest correlation coefficient will be the most effective.
Figure 1 intentional shift

To further improve the accuracy of the position estimation, it is important to explore other
feature matching techniques such as SIFT and ORB. Additionally, it is important to explore
methods that can be used to reduce the effects of noise, illumination changes, and other factors.
For example, the use of Kalman filters can help to reduce the effects of noise, and the use of
frame-to-frame tracking algorithms can help to reduce the effects of illumination changes.