Improve precision algorithm to detect facial expression using LBP - opencv

Question

I'm developping a simple algorithm to detect several facial expressions (happiness, sadness, anger...). I'm based on this paper to do that. I'm preprocessing before to apply LBP uniform operator dividing the normalized image into 6x6 regions as shown in the example below:
By applying uniform LBP 59 feats are extracted for each region, so finally I have 2124 feats by image (6x6x59). I think it's a too large number of feats when I have about 700 images to train a model. I have read that's not good to get a good precission. My question is how can I reduce the dimension of the feats or another technique to improve the precision of the algorithm.

Answer 1

A straightforward way to reduce feature dimensionality - and increase robustness at the same time - would be using rotation-invariant uniform patterns. For a circular neighbourhood of radius and formed by pixels, the texture descriptor represents each region through 10 features. Thus dimensionality is reduced from 2124 to 6 × 6 × 10 = 360.

Answer 2

PCA can help to reduce the size of descriptor without loosing important information. Just google "opencv pca example".
Another helpful thing is to add rotation invariance to your uniform lbp features. This will improve the precision as well as dramatically decrease size of descriptor from 59 to 10.
static cv::Mat rotate_table = (cv::Mat_<uchar>(1, 256) <<
0, 1, 1, 3, 1, 5, 3, 7, 1, 9, 5, 11, 3, 13, 7, 15, 1, 17, 9, 19, 5, 21, 11, 23,
3, 25, 13, 27, 7, 29, 15, 31, 1, 33, 17, 35, 9, 37, 19, 39, 5, 41, 21, 43, 11,
45, 23, 47, 3, 49, 25, 51, 13, 53, 27, 55, 7, 57, 29, 59, 15, 61, 31, 63, 1,
65, 33, 67, 17, 69, 35, 71, 9, 73, 37, 75, 19, 77, 39, 79, 5, 81, 41, 83, 21,
85, 43, 87, 11, 89, 45, 91, 23, 93, 47, 95, 3, 97, 49, 99, 25, 101, 51, 103,
13, 105, 53, 107, 27, 109, 55, 111, 7, 113, 57, 115, 29, 117, 59, 119, 15, 121,
61, 123, 31, 125, 63, 127, 1, 3, 65, 7, 33, 97, 67, 15, 17, 49, 69, 113, 35,
99, 71, 31, 9, 25, 73, 57, 37, 101, 75, 121, 19, 51, 77, 115, 39, 103, 79, 63,
5, 13, 81, 29, 41, 105, 83, 61, 21, 53, 85, 117, 43, 107, 87, 125, 11, 27, 89,
59, 45, 109, 91, 123, 23, 55, 93, 119, 47, 111, 95, 127, 3, 7, 97, 15, 49, 113,
99, 31, 25, 57, 101, 121, 51, 115, 103, 63, 13, 29, 105, 61, 53, 117, 107, 125,
27, 59, 109, 123, 55, 119, 111, 127, 7, 15, 113, 31, 57, 121, 115, 63, 29, 61,
117, 125, 59, 123, 119, 127, 15, 31, 121, 63, 61, 125, 123, 127, 31, 63, 125,
127, 63, 127, 127, 255
);
// the well known original uniform2 pattern
static cv::Mat uniform_table = (cv::Mat_<uchar>(1, 256) <<
0,1,2,3,4,58,5,6,7,58,58,58,8,58,9,10,11,58,58,58,58,58,58,58,12,58,58,58,13,58,
14,15,16,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,17,58,58,58,58,58,58,58,18,
58,58,58,19,58,20,21,22,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,
58,58,58,58,58,58,58,58,58,58,58,58,23,58,58,58,58,58,58,58,58,58,58,58,58,58,
58,58,24,58,58,58,58,58,58,58,25,58,58,58,26,58,27,28,29,30,58,31,58,58,58,32,58,
58,58,58,58,58,58,33,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,34,58,58,58,58,
58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,58,
58,35,36,37,58,38,58,58,58,39,58,58,58,58,58,58,58,40,58,58,58,58,58,58,58,58,58,
58,58,58,58,58,58,41,42,43,58,44,58,58,58,45,58,58,58,58,58,58,58,46,47,48,58,49,
58,58,58,50,51,52,58,53,54,55,56,57
);
static cv::Mat rotuni_table = (cv::Mat_<uchar>(1, 256) <<
0, 1, 1, 2, 1, 9, 2, 3, 1, 9, 9, 9, 2, 9, 3, 4, 1, 9, 9, 9, 9, 9, 9, 9, 2, 9, 9, 9,
3, 9, 4, 5, 1, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 2, 9, 9, 9, 9, 9, 9, 9,
3, 9, 9, 9, 4, 9, 5, 6, 1, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 2, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
3, 9, 9, 9, 9, 9, 9, 9, 4, 9, 9, 9, 5, 9, 6, 7, 1, 2, 9, 3, 9, 9, 9, 4, 9, 9, 9, 9,
9, 9, 9, 5, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 6, 9, 9, 9, 9, 9, 9, 9, 9,
9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 7, 2, 3, 9, 4,
9, 9, 9, 5, 9, 9, 9, 9, 9, 9, 9, 6, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 7,
3, 4, 9, 5, 9, 9, 9, 6, 9, 9, 9, 9, 9, 9, 9, 7, 4, 5, 9, 6, 9, 9, 9, 7, 5, 6, 9, 7,
6, 7, 7, 8
);
static void hist_patch_uniform(const Mat_<uchar> &fI, Mat &histo,
int histSize, bool norm, bool rotinv)
{
cv::Mat ufI, h, n;
if (rotinv) {
cv::Mat r8;
// rotation invariant transform
cv::LUT(fI, rotate_table, r8);
// uniformity for rotation invariant
cv::LUT(r8, rotuni_table, ufI);
// histSize is max 10 bins
} else {
cv::LUT(fI, uniform_table, ufI);
}
// the upper boundary is exclusive
float range[] = {0, (float)histSize};
const float *histRange = {range};
cv::calcHist(&ufI, 1, 0, Mat(), h, 1, &histSize, &histRange, true, false);
if (norm)
normalize(h, n);
else
n = h;
histo.push_back(n.reshape(1, 1));
}
The input is your CV_8U grey-scaled patch (one of those rects). The out is the rotation invariant, uniform, normalized reshaped histogram (1 line). Then you concat your patches histograms into the face descriptor. You will have 6*6*10 = 360. This is good by itself but with pca you can make it 300 or less without loosing important information and even improving the quality of detection because removed dimensions (let's say with variances less than 5%) not just occupy space but also contain mostly the noise (coming from, for example, gaussian noise from the sensor).
Then you can compare this concat histogram with the bank of faces or using svm (rbf kernel fits better). If you do it correctly, then predict for one face should not take more than 1-15ms (5 ms on my iphone7).
Hope this helps.