Another topic that attracts my attention lately is Generelized Category Discovery, which I consider a realistic scenario in today's vision systems, that deal typically with a set of known classes during training and are often faced with unknown classes at deployment.

This is a collaboration with Dr. Meltem Elitaş Hartmann, where the objective is to develop biomedical image analysis based, robust and precise cell tracking solutions in order to explore the effect of antibiotics on bacteria.

This project (Tübitak 1001) is a collaboration with Prof. Koray Kayabol from Gebze Technical University. It addresses the domain shift in common remote sensing image analysis tasks, such as scene classification, object detection and change detection. It aims to design and implement novel domain generalization solutions adapted specifically for remote sensing multi-modal data, coming from multispectral and synthetic aperture radar sensors. 

This is an international project (Expanding the Experience of Citizens through Extended Reality (X-CITE)) funded by the EU Horizon program, led by Finland, and consists of a large consortium with many private and public partners from Türkiye, Netherlands  and Belgium. My role is to design and develop computer vision solutions oriented towards crowd counting and density estimation from central urban areas of Istanbul and also from public transportation vehicles such as buses and subways.

This project is led by Dr. Mehmet Kuru from the History Department of Sabanci University, and the objective is to facilitate the research of historians with the now digitized Ottoman Archive, by providing them a visual content based retrieval system that will enable them to retrieve (all?) the documents in the archive containing a user provided query word via its digital handwritten image.

This is a collaboration with Dr. Hüseyin Özkan from the Electronics Engineering program of Sabanci University and Dr. Vijaya B. Kolachalama from the Department of Medicine, Boston University. The goal is to design and implement robust visual recognition models, that can handle the frequently encountered domain shift across 3D MRI devices, and acquisition protocols, in the context of Alzheimer's disease detection. We are also exploring novel augmentation methods based on mathematical morphology that can leverage the a priori shape information of the brain.

Neural networks have been criticized for their lack of explainability and interpretabilitly ever since their inception. Lately there have been a lot of efforts focused on resolving these issues through a variety of tools such as GradCAM, Shap values, Lime, etc. In the context of our studies we have employed and adapted these tools so as to be able to use them with the challenging data types (multi-spectral, multi-temporal, etc) commonly encountered in remote sensing.

Domain generalization tackles a problem even more challenging than domain adaptation, where no access to target label or data is possible. It often operates with multiple source domains, where the objective is to calculate features that capture the domain-invariant properties. Our recent studies on this topic have led to one of the first domain generalized object detection designs, with no covariate shift assumptions, with both single and two-stage detectors. This has been a collaboration with Dr. K. Seemakurthy (UK), Dr. Petra Bosilj (NL) and Dr. Charles Fox (UK).

The domain shift is omnipresent in remote sensing; e.g. geographically, temporally, across sensors, etc. Combined with the scarcity of labeled data in this area, it constitutes a significant practical obstacle preventing large scale deployment of remote sensing applications. We use domain adaptation/generalization through various methods to tackle this problem.

Bingöl is renounced for its honey, and has won the first place multiple times in international contests. In this project the focus is on its pollen-wise and genetic branding. My role is to design and develop a system for pollen localization and identification from microscopy images. This project is a collaboration between Sabanci, Gebze Technical and Bingöl Üniversities.

Estimating crowd density reliably and accurately has a large security oriented application potential. It tackles serious computer vision challenges in terms of viewpoint, scale, object density and occlusion. This work was conducted in the context of a bilaterla project between Türkiye and Qatar, and led by Gebze Technical University.

Robotic phenotyping requires computer vision methods that estimate the number of fruit or grains in an image. We are researching deep regression and detection based approaches towards this goal. This work was a collaboration with the University of Lincoln, UK.

Real time detection of weeds helps to apply weed-killers directly to their intended target, thus minimizing the contamination of crops and of soil. We have achieved high levels of performance in terms of both accuracy and execution speed using end-to-end convolutional neural networks. We have additionally shown that it is possible to transfer the knowledge gain with one type of crops to another. This work was a collaboration with the University of Lincoln, UK.

Satellites and aircraft acquire surface visual data at speeds far exceeding our capacity to label them. Consequently, methods and techniques for their management and retrieval have become a paramount necessity. In this project we developed new content description and ranking approaches specifically designed for remote sensing images. This work was supported by a Tübitak 1001 Grant (115E857).

The Balik Lake at the Delta of Kizilirmak river is of paramount environmental importance to Turkey, as it accommodates numerous endemic species. As such it requires constant monitoring. In this context, I and my colleagues from Samsun University work on the calculation of automatic water pollution maps from satellite images, in order to eliminate the time consuming need of in situ sample acquisition and laboratory analysis. My work on this topic was supported by a Tubitak grant (1001). The website is online: HERE

With the widespread availability of both high spectral and high spatial resolution remote sensing images, several new application areas have emerged, such as sophisticated compound object detection and retrieval. Mathematical morphology being shape oriented per se, is highly suitable for processing images of high spatial resolution. Then again, the high number of bands (tens and often hundreds) constitutes a serious problem, since many of the existing image analysis methods do not lend themselves to multivariate data. In this context, we investigate methods for best applying morphological operators to hyperspectral images and focus specifically on tree representations and attribute profiles (max-tree, alpha-tree, etc.). My work on this topic was supported by a career grant from Tubitak (3501).

We participated with colleagues from the South Brittany University (France) in the mitosis detection contest organized during the ICPR 2012 (Contest Website). The objective was to detect and recognize cells in mitosis in an effort to aid the automated prognosis of breast cancer. We worked both on color images (as seen below) as well as multispectral data. Out of the 129 participants only 17 submitted their results. Out of which we achieved the best 10th score using a purely automated solution; please note that most of the contestants relied on human assisted approaches.

Texture classification or categorization is an important sub-topic of texture analysis with hundreds of solutions being available to this end. The main challenge is designing scale, rotation and illumination invariant features. With this purpose, besides several other experimentations, I have extended the classical definition of morphological covariance, so that it can outperform other widely used texture features such as LBP and MR8 with several publicly available texture collections. For more information: http://dx.doi.org/10.1016/j.patcog.2012.06.004.

ImageCLEF has been organizing a plant identification track since 2011, where plant species are to be identified from various plant images: leaves, fruits, stems, branches, etc. During our collaboration with Sabanci University, I developed several morphology based content features and we participated in all occasions. We have achieved twice the 1st place (2012-13) and once the 2nd place (2016). For more information: PlantCLEF. Our work on this topic was supported by a Tubitak grant (1001).

The objective of CBIR, which constitutes the main application of my PhD, is to exploit images and more generally visual data using their visual content. Although some visual search engines have already reached the general public (e.g. like.com), their majority still index image files based on their filenames and textual meta data. Our ultimate aim is to carry out successfully the unsupervised annotation of a heterogeneous visual collection. Our work so far, has resulted in an engine called MIMAR (Morphological IMage Annotation and Retrieval), which using purely morphological features, is based on an architecture of keywords in order to implement content-based retrieval of still images. The system has led so far only to limited success using medium-sized image collections. Nonetheless, we have only begun to explore the potential of morphology for color indexation purposes.

Even though mathematical morphology provides a rich variety of tools and operators specializing particularly in the extraction of spatial structures, its application to color and more generally to multivariate images (e.g. multispectral, hyperspectral, etc) is highly problematic and still limited. Overcoming these limitations and extending morphology to color images has been the main topic of my PhD, and although we have introduced several novel approaches in this regard, our efforts will continue until we have obtained the ultimate color morphological solution.