Mücahit Çalışan

Predicting the glass transition temperature (Tg) of polymers from structures remains difficult, and whether descriptors and molecular-language embeddings carry complementary information is unclear. CDEF-Tg (Compact Descriptor–Embedding Fusion for Glass Transition Temperature Prediction) concatenates 205 RDKit descriptors (RDKit205) with a frozen 384-dimensional ChemBERTa embedding (ChemBERTa384) and maps the 589-dimensional vector to Tg with an ExtraTreesRegressor. It was evaluated on 10,065 identity-resolved records of a polyimide-oriented dataset under three leakage-safe protocols: IID Group-Safe (IID, independent and identically distributed), Scaffold OOD (OOD, out-of-distribution), and Similarity OOD (T = 0.90). The model was selected on validation evidence and frozen before the final-test sets were opened. In the locked final test, CDEF-Tg reached a macro mean absolute error (MAE) of 48.84 °C, versus 50.17 °C (RDKit205 alone) and 54.33 °C (ChemBERTa384 alone). Group-aware bootstrap analyses separated CDEF-Tg from RDKit205 alone in MAE only under Similarity OOD and showed no MAE advantage over the comparator RDEF-Tg. Additional validation-only controls showed a lower macro MAE for a Morgan-fingerprint partner than for CDEF-Tg (49.11 versus 53.15 °C), and ChemBERTa384 lowered LightGBM error; the ExtraTrees ordering was stable across random states (split seeds were not varied). The results suggest a modest, protocol-dependent benefit within this dataset; the additional analyses provide descriptive evidence.

Electroluminescence (EL) imaging reveals internal defects in photovoltaic (PV) cells, but supervised inspection depends on costly defect annotation. This study presents a structure-aware anomaly detector built on a frozen DINOv2 ViT-S/14 backbone. A global branch scores whole-image deviation from normal training images. A local branch stores eight prototypes of normal appearance for each patch position of the cell and yields a patch anomaly map. Reference memories are fitted on defect-free images only; labeled anomaly validation images are used for model selection and threshold tuning, and bounding boxes are used only for evaluation. On a content-deduplicated PVEL-AD test set of 5143 images, the fused detector reached an AUROC of 0.8876 and an average precision (AP) of 0.9253 with 27.7 MiB of reference memory. Adapted PaDiM and PatchCore baselines, evaluated on the same partitions but with different backbones and input sizes, reached AUROCs of 0.7959 and 0.7907. Box-derived localization gave a Pointing Game hit rate of 0.6404 and a top-1% patch hit rate of 0.9122. Transferring the PVEL-AD calibration unchanged to the ELPV dataset failed: all calibrated scores saturated (AUROC 0.5000), and post-hoc raw scores ranked only weakly and differently for each cell type. Calibration transfer and domain sensitivity therefore remain open problems. The results support frozen foundation-model features as a compact option for PV EL inspection; energy, cost, and field benefits were not measured.

Detecting and identifying objects outside the direct line of sight is important for security, search and rescue, and autonomous systems, and acoustic sensing offers a low-cost and environmentally robust alternative to optical non-line-of-sight (NLOS) methods. Most acoustic NLOS studies address a single object. This study addresses the classification of paired object combinations from secondary NLOS acoustic reflections. Linear chirp signals reflected from a wall were recorded with an eight-loudspeaker, eight-microphone array in a single controlled room (120 recordings, four balanced classes). A fixed-size, 513-dimensional short-time Fourier-transform (STFT) spectral summary was classified with Random Forest, Decision Tree, k-Nearest Neighbors, and Multilayer Perceptron models using recording-level stratified five-fold cross-validation repeated over five seeds. Random Forest and the Multilayer Perceptron reached accuracies of 0.9850 ± 0.0207 and 0.9600 ± 0.0231, respectively, against a chance level of 0.25. Comparisons with MFCC, Welch, and compact spectral-descriptor representations, single-microphone and multi-channel strategies, and preprocessing ablations, together with label-permutation and nuisance-baseline controls, indicate that the separation is carried by the spectral envelope rather than by gross amplitude, and that the eight-channel array gave no systematic advantage over a single microphone in this dataset. The results are limited to one room, one reflective surface, and recording-level evaluation, and they do not demonstrate generalization to unseen sessions or rooms.

Non-Line-of-Sight (NLOS) sensing has emerged as a critical research domain for detecting and classifying hidden subjects in obstructed environments, with particular relevance to search-and-rescue operations. Whereas traditional modalities such as Light Detection and Ranging (LiDAR), radar, and cameras are constrained by direct visibility, acoustic sensing provides a cost-effective and robust alternative by exploiting diffraction and reflection to propagate around obstacles. This study investigates an acoustic NLOS sensing approach for reliable “liveness” detection in cluttered scenarios, addressing two major gaps in the literature: (i) the underexplored classification of human liveness versus inanimate objects, and (ii) the impact of multichannel data representation strategies on model performance. We systematically compare raw 8-channel signals, arithmetic mean fusion, and a novel Convolutional-Transformer Autoencoder (CT-AE) fusion across Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Random Forest classifiers. Experimental results reveal that raw signals yield promising accuracy for deep learning models (up to 0.972 with 2D-CNN), whereas channel averaging degrades performance in certain architectures (0.795 with 1D-CNN). The proposed CT-AE fusion introduces a “fusion paradox.” While deep learning models suffer performance drops, Random Forest achieves state-of-the-art accuracy (0.912) by leveraging denoised latent features. These findings highlight the importance of aligning representation strategies with model inductive biases and provide actionable insights for designing robust acoustic NLOS detection systems. Beyond methodological contributions, this study advances life-critical applications by enabling responders to distinguish between survivors and debris in disaster scenarios.

The healthy operation of solar energy panels used for electricity production from solar energy is important in terms of the efficiency of the energy produced. In this study, the types of damage that affect energy production from solar energy panels, the effects of these damages on energy production and the methods of detecting the damages are discussed in detail. Damages in solar energy panels occur due to environmental, mechanical, thermal and electrical factors. These damages directly affect energy production negatively and reduce efficiency.The types of damages were examined under four main headings in the study. These are mechanical damages (micro cracks, delamination), thermal damages (hot spot formation, overheating), electrical damages (short circuit, disconnections) and environmental damages (dusting, effect of UV rays). Early detection of damages is important to prevent energy losses and extend the life of the panels. When the damage detection methods were examined, they were grouped under three main headings: Traditional methods, advanced technology-based methods and other methods.The findings obtained in the study were compiled and the difficulties encountered in the damage detection processes, as well as the inadequacies of traditional methods and the advantages and disadvantages of advanced technological methods based on artificial intelligence were revealed, and solution suggestions for damage detection processes were presented. This study provides an important guide for the development and application of damage detection methods.

Cell defects in photovoltaic (PV) modules directly affect production yield, making early and reliable detection critical. Electroluminescence (EL) imaging is widely used for this purpose since it reveals defects such as microcracks and inactive regions; however, manual assessment is time-consuming and subjective. In this study, EL images from the public ELPV benchmark dataset are automatically classified into four defect-severity levels (flawless, low, medium, high). The proposed method augments a ResNet50 backbone with the Squeeze-and-Excitation (SE) block, which adaptively recalibrates channel-wise features. The model is compared against ResNet50, MobileNetV2 and EfficientNetB0 baselines using ImageNet pretrained weights and stratified 5-fold cross-validation. The proposed SE-ResNet50 outperforms all baselines with 0.7609 accuracy and 0.559 macro-F1. Furthermore, an ordinal error-magnitude analysis shows that the SE block yields the smallest and closest errors, reducing the mean absolute level error to 0.443. Nevertheless, the contribution of SE remains limited for the minority 'medium' class, indicating that class imbalance is the principal bottleneck.

In this study, a model based on Hybrid CNN-NCP and Self-Attention is proposed to recognize human activities from sensor data. The proposed model aims to improve the accuracy of human activity recognition (HAR) systems by learning spatial and temporal dependencies more effectively. The CNN layer, one of the main components of the model, extracts low-level spatial patterns from sensor data. These features are then transferred to the NCP layer, where long-term dependencies in time series data are learnt using dynamic feedback mechanisms. In the final stage, the Self-Attention mechanism improves the decision-making process of the model by focusing on critical information in different time periods. In this study, accelerometer and gyroscope sensor data are used to classify 12 different movement activities that are frequently performed in daily life. The performance of the model was evaluated by Group K-Fold cross-validation method and an average accuracy rate of 90.38% was obtained. The proposed hybrid model provides an innovative solution in HAR systems by providing higher performance compared to traditional deep learning approaches.

Fish is important for human health due to its high nutritional value. However, it is prone to spoilage due to its structural characteristics. Traditional freshness assessment methods, such as visual inspection, are subjective and prone to inconsistency. This study proposes a novel, cost-effective hybrid methodology for automated three-level fish freshness classification (Day 1, Day 2, Day 3) by integrating single-wavelength laser reflectance data with deep learning-based image features. A comprehensive dataset was created by collecting visual and laser data from 130 mackerel specimens over three consecutive days under controlled conditions. Image features were extracted using four pre-trained CNN architectures and fused with laser features to form a unified representation. The combined features were classified using SVM, MLP, and RF algorithms. The experimental results demonstrated that the proposed multimodal approach significantly outperformed single-modality methods, achieving average classification accuracy of 88.44%. This work presents an original contribution by demonstrating, for the first time, the effectiveness of combining low-cost laser sensing and deep visual features for freshness prediction, with potential for real-time mobile deployment.

Bees, which have a critical role in the sustainability of our ecosystem, are of vital importance in maintaining ecological balance and biodiversity. More than one-third of the world’s food production is dependent on pollinators, and one of the most important of these is the honey bee. As the decline of bee populations threatens ecological balance and biodiversity, the importance of computer vision techniques for the automatic detection of these species is increasing. In this study, two different deep learning models based on classical convolutional neural network (CNN) and Residual Neural Network (ResNet) architectures are proposed and their effectiveness in classifying honey bee species is examined. Model1 is based on the CNN structure and Model2 uses CNN-ResNet integration. In the experiments on the BeeImage dataset containing images of six different honey bee species, it was determined that Model1 achieved 94.91%, and Model2 achieved 93.63% with the K-fold Cross-Validation technique. The results show that deep learning models can classify bee species in a fast and low-cost way. It is seen that these technologies have significant potential in biological research. The high success rates obtained in the study show that the developed deep learning models can contribute to the protection of biodiversity.Cite this article as: M. Çalışan and N. Olgun, “Evaluation of classification performance of honey bee species with deep learning models,” Electrica, 25, 0157, 2025. doi: 10.5152/electr.2025.24157.

This paper systematically explores the applications of image processing techniques in machined surface analysis, a critical area in industries like manufacturing, aerospace, automotive, and healthcare. It examines the integration of image processing in traditional Computer Numerical Control (CNC) machining and micromachining, focusing on its role in tool wear analysis, workpiece detection, automatic CNC programming, and defect inspection. With AI and machine learning advancements, these technologies enhance defect detection, surface texture analysis, predictive maintenance, and quality optimization. The paper also discusses future advancements in high resolutions, 3D imaging, augmented reality, and Industry 4.0, highlighting their impact on productivity, precision, and challenges such as data privacy. In conclusion, image processing remains vital to improving manufacturing efficiency and quality control.