A brief description of each step in the pipeline is as follows :
Input : Image of size m x n x 1 where, m = number of row pixels and n = number of column pixels. Each pre-trained model has a different input size (same as original papers)
Image Augmentation : The image is rotated by performing a random rotation sampled from the range [-60,60] and normalized to a scale of [0,1] using adaptive histogram equalization\citealt{Pizer1987}. The output label [angle, \(\theta\)] is similarly modified for each input image by adding the rotation angle to original label.
Feature Extraction : We use a pre-trained model to extract features from this image and get a 4-dimensional tensor from this step. These features are used to train a shallow network for doppler angle estimation.Model : Layer modules consisting of Batchnormaliztion, Fully Connected, Relu Activation and Dropout operations.
- Batchnormalization\citealt{pmlr-v37-ioffe15} : Batchnormalization operation deals with internal covariate shift at each layer by normalizing output of each intermediate layer and feeding this normalized output to the next layer as given in \ref{batch_normalization}.