Onur – Python, Machine learning, Deep Learning

The project was a desktop-based digital forensics software suite designed to assist law enforcement and forensic professionals in analyzing and verifying images and videos. It provided advanced image processing filters and video analysis tools to detect tampering, enhance image quality, and extract critical information from digital media. The main features included real-time video processing, image enhancement filters, a user-friendly GUI, and multithreading support for smooth and efficient operation on desktop environments.

• Implemented image processing filters for desktop application to enhance and analyze images, improving the accuracy of media verification.
• Developed a video processing pipeline on top of GStreamer for desktop, allowing for seamless streaming of videos and extraction of frames for image filtering.
• Improved the desktop GUI by introducing new filter windows and optimizing the layout for easy navigation and usability in forensic analysis.
• Utilized Qt for event-driven programming to ensure that the desktop application had a responsive, interactive, and smooth user interface.
• Optimized performance using multithreading, ensuring efficient use of system resources and preventing UI freezes during resource-intensive tasks.

Implementation of identity document detection within an e-commerce application, aimed at verifying the legality of customers to purchase specific products.

• Implemented a low-power deep learning model.
• Fine-tuned algorithm parameters for precise detection of specific documents.
• Corrected the viewing angle of ID documents for optimal Optical Character Recognition.
• Successfully detected and classified passports, work permits, and driver's licenses, considering both country-wise and state-wise variations (European Union and United States).
• Published the model using Amazon Web Services.

SaaS platform to analyze the liveness of a user to overcome spoofing attacks. Combining 3 different aspects of liveness properties to understand whether the user is a real person:

1. head pose estimation directs the user to turn his/her head in different directions in a specific order, then detects his/her head movements;

2. eye-blink detection to count how many times the user blinks his/her eyes in a given time range;

3. hand landmark detection and classification to classify hand signs of users.

• Designed head pose model without using facial key points unlike traditional methods;
• avoided facial key points to reduce calculation cost and design a compact mechanism;
• implemented new feature(ratio) to calculate distances of vertical eye and horizontal eye;
• used two-stage approach as detecting hand landmarks and classification of them;
• trained SVM to classify different hand signs.

This software helps to understand the current emotions of users besides providing age & gender info. Our customer can evaluate how well communication between its agent, and its customer is based on provided data that was extracted.

• Designed model to extract age & gender with multi-head architecture;
• used the same detection model with the Face Recognition system;
• built age & gender feature extractor to be in the same model to reduce the calculation cost;
• refactored emotion estimation model to cover 7 different moods of a person;
• developed algorithms on ResNet architecture to overcome the overfitting issues.

End-to-end face recognition system which detects face and extracts specific features of individuals using deep learning methods. It ensures the detection of complex face samples that contain different scales, poses, occlusion, and blur. Then, it finds distinguishable characteristics of the detected face. Extracted features are used to calculate similarities between each other.

• Designed detection algorithm to be robust to hard samples;
• made extra five-face landmark annotation to detect blurry faces;
• used light-weight Mobile net architecture for detection to optimize memory allocation;
• refactored loss function of recognition model to enhance separation of features from each other;
• redesigned backend architecture for using deep learning models on either GPU or CPU.

Implementation of action recognition module with the help of deep learning to classify different motions in a compressed video. The developed module can analyze human-human, and human-object interactions to define actions.

• Compressed raw videos with MPEG-4 compression method;
• extracted compressed components of a video to work on them
• used these components (namely I-frame, motion vectors and residuals) to understand motion in a video;
• implemented different deep learning models to work with three compressed video components;
• developed an end-to-end module to classify more than a hundred action classes.

Application of U-shaped deep learning method to segmental aerial images taken by UAV. It helps to analyze the partition of the earth's surface for different purposes such as monitoring green areas, changes in place of buildings, etc.

• Annotated raw images to be ground truth for training and test purposes;
• designed encoder-decoder architecture for deep learning model;
• implemented encoder network as VGG16, which is in charge of producing feature maps;
• used upsampling-convolution operations to make a decoder network that maps low-resolution image to the intended segmented image;
• made a deep learning model that segments images under different classes such as cars, buildings, roads, rails, etc.