Marcin Kowalski

Thermal infrared face analysis remains challenging compared to its visible spectrum counterpart due to reduced texture, sensor noise, and large variations across camera characteristics. These factors limit the generalization of existing landmark detectors, which exhibit substantial performance degradation when deployed on unseen datasets or acquisition conditions. In this study, we introduce a dense facial landmark detection framework that directly regresses a 478-point facial mesh from thermal images. The model is based on a ConvFormer-S36 backbone with a proposed lightweight Multi-Query Attention Pooling head, enabling the learning of stable thermal representations with reduced overfitting to dataset-specific statistics. Proposed model achieved stable vertical-axis regression behavior. We systematically analyzed failure modes across multiple architectures and identified distinct patterns that emerged under extreme poses, occlusions, or low-contrast conditions. We further proposed a lightweight, geometry-based failure detection mechanism that leverages the distinctive signature of errors to flag degraded predictions. We validated our framework by using several thermal infrared datasets for emotion recognition and landmarking. Quantitatively, our model achieved a normalized mean error of 0.045 on the external dataset after mapping the dense mesh to the 68-point standard format, whereas qualitative evaluations on external datasets demonstrated improved cross-dataset generalization under diverse imaging conditions, poses, and occlusions. Together, these results establish dense mesh-based thermal landmarking as a generalizable paradigm and highlight the importance of understanding the failure modes for deployment in real-world applications.

Skin diseases pose a major public health challenge, with skin cancer being the most prevalent malignancy. Early detection is critical for improving patient outcomes, yet traditional diagnostic methods rely on expert evaluation, which is limited by clinician variability, image quality, and accessibility constraints. Deep learning (DL)-based models offer automated diagnostic capabilities but require large-scale centralized data aggregation, which conflicts with privacy regulations and poses security risks. Federated Learning (FL) enables collaborative training across multiple hospitals without sharing raw patient data, yet it introduces communication overhead, security vulnerabilities, and interpretability challenges, hindering its real-world deployment. To address these issues, an Encrypted FedAvg-Based Explainable Federated Learning approach has been proposed utilizing a Lightweight Deep Learning Multi-Scale Convolutional Neural Network (LWMS-CNN) for efficient, privacy-enhanced, and interpretable skin cancer diagnosis. The proposed method integrates Homomorphic Encryption (HE) in a simulated federated system for secure model aggregation, privacy enhancements while maintaining strong diagnostic performance. Additionally, SHapley Additive exPlanations (SHAP) and GradCAM enhance interpretability, enabling clinicians to understand AI-driven predictions. Experimental evaluations on the HAM10000 dataset demonstrate that the proposed LWMS-CNN-FL model achieves 98.62% accuracy, with only a 0.3% tradeoff when encryption is applied, ensuring robust security without compromising diagnostic reliability. The model’s generalization ability is further confirmed through assessments on additional benchmark datasets, including ISIC 2019, where it achieved an accuracy of 96.22%, and PAD-UFES-20, where it reached 89.84%. By integrating lightweight deep learning, federated learning, encryption, and explainability, this study presents a scalable, privacy-enhanced, computationally efficient, and clinically interpretable AI-driven solution for secure and accurate early skin cancer detection in world healthcare applications.

Brain tumors present a significant global health challenge, and their early detection and accurate classification are crucial for effective treatment strategies. This study presents a novel approach combining a lightweight parallel depthwise separable convolutional neural network (PDSCNN) and a hybrid ridge regression extreme learning machine (RRELM) for accurately classifying four types of brain tumors (glioma, meningioma, no tumor, and pituitary) based on MRI images. The proposed approach enhances the visibility and clarity of tumor features in MRI images by employing contrast-limited adaptive histogram equalization (CLAHE). A lightweight PDSCNN is then employed to extract relevant tumor-specific patterns while minimizing computational complexity. A hybrid RRELM model is proposed, enhancing the traditional ELM for improved classification performance. The proposed framework is compared with various state-of-the-art models in terms of classification accuracy, model parameters, and layer sizes. The proposed framework achieved remarkable average precision, recall, and accuracy values of 99.35%, 99.30%, and 99.22%, respectively, through five-fold cross-validation. The PDSCNN-RRELM outperformed the extreme learning machine model with pseudoinverse (PELM) and exhibited superior performance. The introduction of ridge regression in the ELM framework led to significant enhancements in classification performance model parameters and layer sizes compared to those of the state-of-the-art models. Additionally, the interpretability of the framework was demonstrated using Shapley Additive Explanations (SHAP), providing insights into the decision-making process and increasing confidence in real-world diagnosis.

Crop diseases can significantly affect various aspects of crop cultivation, including crop yield, quality, production costs, and crop loss. The utilization of modern technologies such as image analysis via machine learning techniques enables early and precise detection of crop diseases, hence empowering farmers to effectively manage and avoid the occurrence of crop diseases. The proposed methodology involves the use of modified MobileNetV3Large model deployed on edge device for real-time monitoring of grape leaf disease while reducing computational memory demands and ensuring satisfactory classification performance. To enhance applicability of MobileNetV3Large, custom layers consisting of two dense layers were added, each followed by a dropout layer, helped mitigate overfitting and ensured that the model remains efficient. Comparisons among other models showed that the proposed model outperformed those with an average train and test accuracy of 99.66% and 99.42%, with a precision, recall, and F1 score of approximately 99.42%. The model was deployed on an edge device (Nvidia Jetson Nano) using a custom developed GUI app and predicted from both saved and real-time data with high confidence values. Grad-CAM visualization was used to identify and represent image areas that affect the convolutional neural network (CNN) classification decision-making process with high accuracy. This research contributes to the development of plant disease classification technologies for edge devices, which have the potential to enhance the ability of autonomous farming for farmers, agronomists, and researchers to monitor and mitigate plant diseases efficiently and effectively, with a positive impact on global food security.

The design of transparent conductive electrodes (TCEs) for optoelectronic devices requires a trade‐off between high conductivity and transmittivity, limiting their efficiency. This paper demonstrates the best ever achieved TCEs with the novel approach to fabricating TCEs that effectively alleviates this trade‐off: a monolithic high contrast grating integrated with metal (metalMHCG). The metalMHCG enables higher electrical conductivity than other TCEs, while providing transmissive and antireflective properties. It focuses on infrared spectrum TCEs, which are essential for sensing, thermal imaging, and automotive applications. However, due to elevated free carrier absorption, they are much more demanding than TCEs for the visible spectrum. It demonstrates a record 75% absolute transmittiance of unpolarized light, resulting in a record 108% transmittance relative to plain GaAs substrate. It achieves even larger absolute transmittance of polarized light, reaching 92% or 133% relative transmittance. Despite the record high transmittance, the sheet resistance of the metalMHCG is the best ever reported, several times lower than any other TCE, ranging from 0.5 to 1 Ω Sq⁻¹.