Middle East Technical University Research Information System
Not a member yet
104897 research outputs found
Sort by
Optimized KiU-Net: A Convolutional Autoencoder for Retinal Vessel Segmentation in Medical Images
No full textMedical image segmentation plays a central role in enhancing diagnosis, surgical planning, and treatment strategies, and in this work, the focus is on segmenting retinal vessels from color fundus images. Precise vessel extraction is essential because vessel morphology reflects several ophthalmic conditions. Wide neural networks such as U-Net and KiU-Net have improved retinal vessel segmentation; however, thin and low-contrast vessel regions remain difficult to capture. U-Net follows an undercomplete design that limits its ability to retain fine structures, whereas KiU-Net combines undercomplete and overcomplete paths to provide better detail extraction but still suffers from accuracy limitations and increased computational cost. We present an Optimized KiU Net model that improves the segmentation of thin and low-contrast retinal vessels while keeping the model lightweight. The design refines convolution channel selection and increases encoder depth, and the final feature fusion uses a single concatenation step, which supports faster convergence with fewer parameters. On the RITE dataset, the model achieves an F1 score of 79.80 and an IoU of 66.30, outperforming U-Net, KiU-Net, and other similar sized architectures. Compared to KiU-Net, the gains are about four points in F1 and six points in IoU with fewer parameters. Additional evaluation on the GlaS dataset shows F1 and IoU scores of 82.21 and 71.03, demonstrating that the method remains effective when compared with existing approaches of comparable scale
Structural and photoelectrical properties of an MDMO-PPV/n-Si hybrid heterojunction photodiode for broadband detection
No full textOrganic–inorganic hybrid photodiodes are attractive for low-cost, broadband photodetection, yet the role of the polymer/Si interface on power- and wavelength-dependent performance is still not fully clarified. In this study, we fabricate a simple Al/MDMO-PPV/n-Si/Al heterojunction photodiode using a solution-processed MDMO-PPV interlayer and systematically correlate its structural, optical and electrical characteristics. UV–Vis spectroscopy reveals a pronounced absorption band at 478nm, and Tauc analysis yields a direct optical band gap of 2.28eV. XRD and SEM measurements indicate a predominantly amorphous, porous polymer film with short-range order, providing interconnected pathways for charge transport. Current–voltage measurements under light illumination from 20 to 100 mWcm-2 show clear rectification with negligible dark current. Thermionic-emission and Cheung analyses give ideality factors of 3.0–3.5 and zero-bias barrier heights of 0.69–0.74eV, while the photocurrent follows a power law Iph∝P1.42, evidencing trap-assisted photoconduction. The device operates over 300–1050 nm, where the responsivity (R) increases from 0.15 to about 0.42AW-1, the photosensitivity (K) rises to ≈4−13, and the specific detectivity (D) improves from ∼1.1×1010to∼1.8×1010Jones. Time-resolved measurements at 0V bias yield rise and fall times of 0.70 s and 0.94 s, respectively. These results demonstrate that an MDMO-PPV interlayer can efficiently extend n-Si photodiodes into the visible–NIR region and provide bias-free broadband detection, offering a scalable and low-cost platform for next-generation optoelectronic and sensing applications
Multilingual Domain Adaptation for Speech Recognition Using LLMs
No full textWe present a practical pipeline for multilingual domain adaptation in automatic speech recognition (ASR) that combines the Whisper model with large language models (LLMs). Using Aya-23-8B, Common Voice transcripts in 22 languages are automatically classified into the Law and Healthcare domains, producing high-quality domain labels at a fraction of the manual cost. These labels drive parameterefficient (LoRA) fine-tuning of Whisper and deliver consistent relative Word Error Rate (WER) reductions of up to 14.3% for languages that contribute at least 800 in-domain utterances. A data-volume analysis reveals a clear breakpoint: gains become reliably large once that 800-utterance threshold is crossed, while monolingual tuning still rescues performance in truly low-resource settings. The workflow therefore shifts the key success factor from expensive hand labelling to scalable data acquisition, and can be replicated in new domains with minimal human intervention
Random Data and Its Cryptographic Applications
No full textRandom sequence construction represents fundamental components of modern cryptographic systems. The quantitative assessment of randomness relies upon rigorous statistical testing methodologies, establishing statistical randomness evaluation as a critical prerequisite for cryptographic algorithm security validation. Concurrently, data compression technologies have emerged as essential enablers of efficient information transmission within contemporary digital communication infrastructures. This research investigates statistical testing frameworks in cryptographic applications, with particular emphasis on compression-based evaluation methods, notably the Lempel-Ziv complexity test. We present empirical findings from our analysis of a novel bit-level pattern recognition algorithm, validated against data sequences generated through the Advanced Encryption Standard (AES). Furthermore, we introduce a compression methodology derived from this bit-level pattern detection approach, demonstrating its potential applications in cryptographic randomness assessment
Blockchain-Based Carbon Footprint Management
No full textThis paper introduces a novel approach to managing carbon footprints using blockchain technology to integrate these footprints intrinsically into the attributes of products, akin to their price. In contrast to conventional methods that treat carbon footprints as distinct, tradeable units, our model incorporates them directly into the product life cycle, thus maintaining the connection between environmental impact and product consumption. By closely examining blockchain’s functionality, this study illustrates how it can improve transparency and security in transactions while influencing market dynamics by making carbon accountability an integral part of product ownership. This method may aid in establishing a more sustainable economic model by better incorporating environmental costs into transactions, potentially fostering progress in environmental finance and sustainability initiatives
SAMChat: Introducing Chain of Thought Reasoning and GRPO to a Multimodal Small Language Model for Small Scale Remote Sensing
No full textRemarkable capabilities in understanding and generating text-image content have been demonstrated by recent advancements in multimodal large language models (MLLMs). However, their effectiveness in specialized domains-particularly those requiring resource-efficient and domain-specific adaptations- has remained limited. In this work, a lightweight multimodal language model termed SAMChat is introduced, specifically adapted to analyze remote sensing imagery in secluded areas, including challenging missile launch sites. A new dataset, SAMData, was compiled by verifying hundreds of aerial images through expert review, and subtle military installations were highlighted via detailed captions. Supervised fine-tuning on a 2B-parameter open-source MLLM with chain-of-thought (CoT) reasoning annotations was performed, enabling more accurate and interpretable explanations. Additionally, Group Relative Policy Optimization (GRPO) was leveraged to enhance the model's ability to detect critical domain-specific cues-such as defensive layouts and key military structures-while minimizing false positives on civilian scenes. Through empirical evaluations, it has been shown that SAMChat significantly outperforms both larger, general-purpose multimodal models and existing remote sensing-adapted approaches on open-ended captioning and classification metrics. Over 80% recall and 98% precision were achieved on the newly proposed SAMData benchmark, underscoring the potency of targeted fine-tuning and reinforcement learning in specialized real-world applications. Code and dataset will be available upon acceptance
Integrating solutions for enabling the sustainable development of energy, water and environment systems
No full textThe world has reached 1.24 degrees C of global warming above pre-industrial levels when averaged over the last decade, necessitating time-saving options for the mitigation of climate change. Within a long-term scientific pursuit for integrating solutions, the 19th Conference on Sustainable Development of Energy, Water and Environment Systems was organised in Rome, Italy, in 2024 alongside the regional 4th Latin American and 2nd Asia Pacific Conferences in Chile and Australia, respectively. The thematic coverage of this review-style editorial represents the key findings of the 22 original articles in this special issue, synthesized into six themes. Demand flexibility within decarbonisation processes includes end-use electrification scenarios in renewable energy communities, probabilistic load coordination, and optimising district heating and/or cooling systems. Securing cross-sectoral benefits with renewable energy extends to locating data centres based on community infrastructure needs for clean energy access and clean drinking water, better interconnector capacity, solutions for maritime transport, and tackling environmental pollution and electronic waste. Advances in green hydrogen supply and value chains span a co-simulation framework with solar energy and repurposing for green hydrogen stations, among others. For improving energy storage and waste heat utilisation, new battery energy storage systems, an energy targeting approach linking industry and buildings, and the utilisation of waste heat from water electrolysis are explored. Technological advances and enhanced heat transfer are then supported by findings for parabolic trough solar collectors with nanofluids, building thermal energy load prediction, and optimising heat exchanger design. These advances take place alongside those for marine energy by optimising large-scale, multi-gigawatt offshore wind farms, also considering aspects of public acceptance and options for energy system restructuring. The integrated solutions that are represented within these advances and their broader synthesis provide ample opportunities for mitigating climate change, providing benefits for improving livelihoods, and securing a safer climate future on this shared planet
Criterion shifts change the pattern of output interference
No full textOutput interference in recognition refers to a decrease in performance over the course of a test. The goal of the current study was to determine whether experimentally shifting the decision criterion changes the form of output interference and to identify a process account of any interaction. In two experiments, we manipulated the decision criterion via changes in the base rate of the old items at test (80%, 50%, 20%). Experiment 1 implemented this manipulation within-subjects and failed to induce criterion shifts. In contrast, when the base rate was manipulated between-subjects in Experiment 2, decision criteria differed across conditions. Qualitative patterns suggested that liberal criteria attenuated the hit rate (HR) decline and increased the false alarm rate (FAR) across blocks, whereas conservative criteria yielded steeper HR declines with relatively stable FAR. To further examine this effect, Experiment 3 employed longer test lists and a larger sample. The criterion was manipulated only via prior information about the base rates, while the actual base rate was 50% in all conditions. Experiment 3 revealed a significant interaction between response bias and output interference in HR and FAR. When we used an independent data set (Layher et al., Journal of Experimental Psychology: Learning, Memory, and Cognition, 46(11), 2075–2105, 2020), we demonstrated the same patterns. To account for these findings, we conducted simulations with the Retrieving Effectively from Memory (Shiffrin & Steyvers, Psychonomic Bulletin & Review, 4(2), 145-166, 1997) model. The results were best captured by a learning-during-test mechanism in which every test item is encoded as a new memory trace
Performance boost in QLEDs using octanethiol-capped core/shell/shell quantum dots
No full textQuantum dots attract significant attention as one of the most promising colloidal nanocrystals with unique optical properties and potential applications for the next generation of display technology. In this paper, we evaluate the performance of CdZnSeS-based alloyed-shell quantum dots (QDs) for electroluminescence devices upon additional shell growth and ligand exchange. This includes core/shell (C/S) and core/shell/shell (C/S/S) QDs, whose latter includes an additional ZnS shell and octanethiol (OT) ligands. We present detailed characterizations of QDs using transmission electron microscopy, XRD, and various spectroscopic techniques and demonstrate their QD light emitting (QLEDs). We find the photoluminescence quantum yield of C/S/S QDs increased from 68.8% to 88.7% compared to C/S QDs whereas the emission linewidth narrows from 22.2 nm to 20.8 nm. QLEDs fabricated with C/S/S QDs exhibit a higher peak external quantum efficiency (EQE) of 4.1% and maximum luminance of 85 000 cd m-2, compared to 2.3% EQE and 67 000 cd m-2for C/S QLEDs. In this respect, the OT-assisted shell growth significantly improves the optical property of QDs and performance of QLEDs, likely attributed to the enhanced charge balance and increased radiative recombination rate
A comparative machine learning framework for earthquake damage mapping in Turkey: incorporating MARS and ensemble models in the 2023 Kahramanmaraş earthquake using high-resolution Pléiades imagery
No full textEarthquakes cause severe damage to buildings and infrastructure, posing risks to humans and disrupting urban systems. Traditional ground-based assessments are slow and often impractical in dense urban areas. This study evaluates several machine learning algorithms for detecting damaged buildings after the 6 February 2023 earthquakes in southeastern Turkey, focusing on the city center of Hatay, one of the most affected regions. High-resolution post-event Pléiades imagery and building footprints were used to create training and test datasets covering more than 8,000 structures, and textural features extracted from the imagery served as predictors. We tested artificial neural network (ANN), random forest (RF), support vector machine (SVM) with different kernels, ensemble learners, and multivariate adaptive regression splines (MARS). To our knowledge, this is the first application of MARS to post-earthquake damage mapping using very high-resolution optical data. Model performance was assessed using 10-fold cross-validation, overall accuracy (OA), and F1 score. RF reached high training accuracy (∼0.95) but showed limited generalization. SVM with an RBF kernel behaved similarly, while polynomial SVM (degree 2) performed the worst. MARS produced moderate but stable results across folds. ANN and linear SVM showed comparable performance, with slightly higher stability for the latter. The ensemble model yielded the best test results (OA = 0.59, F1 = 0.66), offering a balance of accuracy, robustness, and computational efficiency. Although accuracies remained modest, the workflow can generate a usable damage map within 1–2 h after acquiring the post-event Pléiades image, making it suitable for rapid first-pass screening in operational settings