Machine Learning Developer in Skopje, Macedonia

A multimodal approach for affective analysis was developed that exploits features from video, audio, electrocardiogram (ECG), and electrodermal activity (EDA) combining two regression techniques, namely boosted regression trees and linear regression. Moreover, a novel regularization approach for the linear regression was proposed in order to exploit the temporal correlation of the affective dimensions. The final prediction is obtained using a decision level fusion of the regressors individually trained on the different groups of features. The promising results obtained on the benchmark dataset show the efficacy and effectiveness of the proposed approach.

In this project Discrete Wavelet Transformation implementation of a bandpass filter for denoising an ECG signal was made. New improved version of a DWT bandpass filter was realized using a circular buffer. Time performance analysis and comparison of obtained solution vs. existing solutions was done.

In this work, a novel two-phase approach is proposed for robust super-resolution in the presence of registration errors and outliers. In the first phase, a machine learning method is used to create a weight matrix for every LR (low resolution) image indicating the presence of registration errors. In the second phase, super-resolution is performed using all of the LR images and the associated weight matrices, creating an image that is free of error artifacts.

In this work, the problem of face de-identification in an image was tackled. The first step towards a solution to this problem is the design of a successful generic face detection algorithm, which will detect all of the faces in the image or video, regardless of the pose. If the face detection algorithm fails to detect even one face, the effect of the de-identification algorithm could be neutralized. That is why a novel face detection algorithm was proposed for face detection and pose estimation. The algorithm uses an ensemble of three linear SVM classifiers. The first, second, and third SVM classifier estimates the pitch, yaw, and roll angle of the face and logistic regression is used to combine the results and output a final decision. Second, the results of face detection and a simple space-variant de-identification algorithm are used to show the benefits of simultaneous face detection and face de-identification.