Cover Letter

I am curious to learn new concepts and topics since my undergraduate and graduate studies. In my bachelor thesis, I worked on a project called Superpixel Based Classification of Hyperspectral Images. When I finished this thesis, the feeling of joy and pride made me do more projects. In light of these, I decided to continue my master's studies in Registration and Fusion of Infrared and Visual Images. I got my MSc degree in Electronics Engineering in 2019.

Since November 2020, I have worked as an algorithm designer in the data fusion of an integrated communication, navigation and identification system in aircraft. This system includes navigation, IFF, and Data Link systems. I participated in requirements and achievability analysis in DOORS and PLM environments. I designed the data fusion concept for navigation system using SysML. (Inertial navigation, terrain and vision based navigation).

I am very much eager to get to know new technologies and techniques in Navigation Systems, Machine Learning, Pattern Recognition, and Signal Processing. I deeply believe that working with an extremely distinguished engineer team with my ambitious, reliable, and diligent personality will bring success to your company, these will motivate me to work harder.

Feel free to message me if you have a problem that I can help you with.
----------------------------------------------------
I would like to express my interest in taking part in an engineering team that would strongly enrich my future studies and helps my prospective career in my further applications at ABC company.

I deeply believe that working with an extremely distinguished engineer team with my investigative, reliable, and diligent personality will bring success to ABC company, these would motivate me to work harder.

--------------------------------------------------
Master of Science Thesis:
The thesis called “Registration and Fusion of Infrared and Visual Images” is suggested by AC company which is a company of Armed Forces Foundation and realized under the supervision of Associate Professor XXX, ABC University.

Multi-sensor data usually provides complementary information of the region surveyed and image fusion which aims to
create a single image from such data offering better scene representation has played an important role in scene analysis in the areas of pattern recognition, medical imaging, remote sensing, and modern military. Specifically, infrared and visible has shown significant advantages in enhancing human visual perception, target detection, and recognition. One the one hand, infrared imaging characterizes the thermal radiation of an object, which can highlight targets efficiently as they often have relatively high temperatures. Also, Infrared imaging is less affected by illumination or weather changes and hence can work well in different conditions. On the other hand, visible images often have high resolutions and reach textures, which are a benefit to describing the details of targets. Therefore, it is desirable to combine these two types of information to increase the accuracy of target detection and recognition. An important prerequisite in the fusion of two images should have been aligned at a pixel level.

In this paper, we aim to solve the registration and fusion problem and to generate more accurate fusion and fast alignments between infrared and visible images. Unregistered thermal and visible image data sets are obtained from FLIR ADAS and TNO. Key points are selected and matched manually from unregistered infrared and visual images. After that, a geometric transformation model is obtained over a reference day image and the same transformation model is applied on the night image pair. Finally, registered images are fused different fusion methods; Principle Component Analysis (PCA), Laplace Pyramid (LP), Discrete Wavelet Transform (DWT), Non-Subsampled Contourlet Transform (NSCT) and Non-negative Matrix Factorization (NMF). The performance of fused images is evaluated using different quality metrics; mutual information (MI), image attribute-based metric QAB/F, NAB/F, LAB/F, image edge-based metric QE, average gradient (AG) and spatial frequency (SF).







Master of Science Project Scholarship:
The project called “Superpixel Based Classification of Hyperspectral Images” is supported by The Scientific and Technological Research Council of Turkey (TUBITAK), under the supervision of Associate Professor XXX.
used techniques; principal component analysis(PCA), simple linear iterative clustering(SLIC), quick shift(QS), gray level co-occurrence matrix(GLCM), histogram of oriented gradients(HOG), local histogram, support vector machine(SVM), k-means, k -nearest neighbor (KNN) in MATLAB.
The goal of this study is the classification of hyperspectral images which are Indian Pines, Salinas, Pavia Centre, and Pavia University acquired from A.V.I.R.I.S (Airborne Visible/Infrared Imaging Spectrometer) and R.O.S.I.S (Reflective Optics System Imaging Spectrometer) sensors. Dimension reduction is applied to hyperspectral images with PCA method. After this step, 3 bands of these images are segmented to its superpixels using SLIC and QS segmentation methods. Image features are extracted implementing GLCM, HOG, and local histogram methods to superpixels. Finally, features are classified using k-means, SVM, and KNN.

Bachelor of Science Thesis Project Scholarship:
Thesis called “Superpixel Based Classification of Hyperspectral Images” is supported by TUBITAK under the supervision of Associate Professor XXX.

used techniques; PCA, linear discriminant analysis (LDA), SLIC, QS, k-means, decision tree (DT), discriminant analysis classifier in MATLAB.
The aim of this thesis is to classify the hyperspectral images which are Indian Pines, Salinas, Pavia Centre, and Pavia University. To do that, firstly, finding effective band numbers and classification types of these images are aimed respectively to go through dimensional reduction and classification process. Secondly, images are segmented to its superpixels using SLIC and QS methods. Image features are defined as an averaging each one of the superpixels. Finally, features are clustered using k-means.

Course and Industrial Projects:

Local Binary Pattern and Morphology Profiles Based Classification of Hyperspectral Images

used techniques; morphologic profiles (MP - opening, closing), local binary pattern, support vector machines, random forest (RF) in MATLAB.

The target of this project is to increase the accuracy ratio of the classification using spatial features in hyperspectral images. To obtain spatial features, PCA should be applied to hyperspectral data. Then, Local Binary Pattern (LBP) and Morphological Profiles (MP) are priory methods used for the extraction of local attributes such as texture and shape from neighboring pixels. In the classification stage, 5% and %10 percent of each class are used to train classification algorithms. In each class, the rest of the data is used for the testing scheme. In terms of classification accuracy, training data is randomized with a reverse knife check algorithm. RF and SVM are classification algorithms using the tests; a radial based kernel for SVM. Once the appropriate kernel scale and box constraint parameter are found, other tests are done according to these parameters. In the random forest approach, a large number of decision trees are created. Every observation is fed into every decision tree. The most common outcome for each observation is used as the final output. A new observation is fed into all the trees and taking a majority vote for each classification model. An error estimate called OOB(Out-of-bag) finds the proper tree number. In the first test, 50-150-250-500 trees are tested to find optimum tree numbers according to OBB error. Final tests are done in this tree number.

• Handwritten Recognition Using MNIST Digit and Extended MNIST Letter Datasets, in PYTHON

used techniques; Traditional Approach; feature extraction: DWT, HOG, and classification: DT, RF, SVM, Deep Learning Approach; Convolution2D, Maxpooling2D, Dropout, Batch Normalization, and Multi-layer Perceptron, Softmax classifier
The main goal of this project is to determine the best model on MNIST Digit and Extended MNIST Letter Handwritten Character Recognition datasets using traditional and deep learning approaches. In this purpose, I have compared 27 deep learning model to see the effect of model optimizers (Adadelta, Adam, Stochastic Gradient Descend), learning rates (0.01, 0.1, 0.5, 1), Batch Normalization and Dropout(x= 0.2, 0.3, 0.5) layers and 9 traditional models according to see the effect of different feature extractors ( DWT, HOG) and classifiers (DT, SVM, RF).

For MNIST Digit Dataset, class label is 10. Best accuracy is 99.24 % which achieved using Adadelta optimizer model based on Convolution2D(x, kernel size (3,3)), Batch Normalization, Maxpooling2D, Dropout(0,2) combined with a Fully Connected Network and Neural Network (256), Batch Normalization and Dropout(0,2) layers via Softmax classifier (x= 32,64,128; y= 0.2). Totally, 1 Input Layer, 16 Hidden Layer, and 1 Output Layer have used for deep learning.

For Extended MNIST Letter Dataset, class label is 37. Best accuracy is 95.07 % which achieved using Adadelta optimizer model based on Convolution2D(x, kernel size (3,3)), Batch Normalization, Convolution 2D(x, kernel size (3,3)), Maxpooling2D, Dropout(0,3), Batch Normalization combined with a Fully Connected Network and Neural Network (512) layers, Dropout(0,3) layers via Softmax classifier (x= 32,64). Totally, 1 Input Layer, 13 Hidden Layer and 1 Output Layer has used for deep learning. Second best accuracy is 92.189 % used same model on MNIST Digit Dataset.