Beruflich Dokumente
Kultur Dokumente
• The first convolution Layer has 32 filters and a kernel size of 5, the
activation function used is ReLU.
ANNOTATIONS
• The second convolution Layer has 64 filters and a kernel size of 3
and the activation function used is ReLU.
• Both the max pooling layers perform down-sampling with strides of
2 and kernel size of 2.
• The data is flattened to a 1-D vector for the fully connected layer
(fc1). Fully connected layer fc1 has 1024 neurons.
• The output of fc1 is passed to a 10-way softmax, which produces a
probability distribution over ten labels.
• A Dropout rate of 0.75 was used, the learning rate was kept at
0.001, batch size of 128 images.
• Accuracy of 98.43% was achieved on a test set of 10000 images.
Methodology of IPPG
• ROI Detection
• Facial region is identified from each frame of the video using the built
in implementation of the Viola-Jones face detection algorithm which
uses various Haar- like features.
• The pixel intensity values in the RGB channels were averaged over the
ROI to generate a 3 x N matrix containing the intensity contours
where N is the number of frames in the video.
Methodology of IPPG
• Independent component analysis technique is applied on the
intensity contour signals to obtain the underlying source signals.
• The JADE (Joint Approximate Diagonalization of Eigenmatrices)
implementation is used to find the demixing matrix W.
• --(2)
• Where I_k = (r,p), source and peak pair from which HR is estimated in
k-th window and w_k is an array of frequency locations
corresponding to the N highest peaks in the three source signals
Methodology of IPPG
• Assuming conditional independence of w_k with respect to I_k, we
get:
--(3)
• Which can be obtained from (1)
• Assuming Markovian independence and applying chain rule we get:
--(4)
Methodology of IPPG
• This is the probability of the peak value pair changing from I_k-1 in (k-
1)th window to I_k in the kth window.
• It is defined as:
• Where lambda is a hyperparameter.
• Combinig 3 and 4 we get an expression:
•
• The HR sequence is obtained by maximising that expression:
•
• This optimization is solved using dynamic programming is a way
similar to viterbi decoding as proposed by Raseena et al.
References