Handwritten annotations and ink contaminants compromise the visual quality and readability of documents. While deep learning has demonstrated impressive results in related tasks, such as text enhancement, it has yet to be applied to the prevalent issue of handwriting removal, primarily due to the absence of foundational datasets and robust methodologies. To address this limitation, we construct HW5K, a comprehensive dataset comprising over 5,000 pairs of high-resolution (2047×2537) document images affected by handwriting. This dataset encompasses a diverse range of handwriting styles, backgrounds, and document formats, thereby providing a robust foundation for advancing research in document image restoration. In conjunction with HW5K, we propose DocUnfold, a novel model specifically designed for high-quality restoration of contaminated document images. DocUnfold employs a multi-stage unfolding process to systematically extract and reconstruct multi-level image features. The model is featured with the DocShuffle module, which disrupts the spatial dependencies of degradation through random shuffling operations. It employs a multi-window partitioning mechanism alongside a corresponding windowed attention module to extract feature information after shuffling, enabling DocUnfold to effectively remove handwritten content while preserving the integrity of the original document. Extensive experiments demonstrate the efficacy of both our dataset and model, yielding impressive results in qualitative and quantitative evaluations. HW5K and DocUnfold collectively establish a new benchmark in the field of document handwriting removal. The link of code and data is https://github.com/CXH-Research/DocUnfold.

The current level of adoption of solar photovoltaic systems requires better methods of thermal fault monitoring, which will support sustainable energy production systems. The application of traditional methods in detecting hotspots is highly significant in analysing the thermal anomaly pattern, but the procedure is ineffective since the process of inspection is manual, time-consuming, and labour-intensive. However, thermal imaging devices offer scalable solutions that are limited in terms of managing complex thermal patterns and combining temperature data at various scales. This paper proposes a Mish-BiFPN Framework to detect thermal hotspots in solar PV systems. Unlike conventional BiFPN or Mish-based models, the proposed framework integrates a temperature-controlled Mish activation with CBAM-guided fusion at every pyramid level, combined with a RegNet backbone and thermal-preserving preprocessing to enhance sensitivity to subtle heat gradients. The offered architecture uses multi-scale feature extraction and attention layers in a way that the model can focus on temperature-sensitive and spatial features at different pyramid layers. It ensures thermal characteristics are retained through preprocessing and data augmentation. The framework further provides a basis to combine the real-time thermal performance monitoring with long-term energy planning that allows dynamically calibrating estimates of solar PV energy generation. This model is trained and validated on thermal hotspot datasets, achieving 12 percentage-point improvement in precision, a 6% improvement in recall, 9% in F1-score, and 21% in Intersection over Union in comparison to the YOLOv10 model. The model shows its potential to address complicated thermal patterns. Energy loss modeling demonstrates that single hotspot detection can prevent approximately 9 kWh per year of power losses per 450 W module. Therefore, this model is very appropriate for large-scale solar PV systems, with both accurate detection and energy conservation optimization.

Retinopathy of Prematurity (ROP) remains one of the leading causes of blindness in children globally. Global screening for ROP is limited by the shortage of trained ophthalmologists and low access to care in resource-limited settings. Vision-Language Models (VLMs) can be utilized to analyze medical images. However, VLMs also have limitations for ROP screening, including patient data privacy concerns, computational requirements for point-of-care devices, and the need for clinically interpretable predictions. In this study, we present TeleROP-VLF, a federated learning framework for VLM collaboration to train and perform analysis across multi-wearable retinal imaging devices while preserving sensitive patient data from leaving the hospital. TeleROP-VLF contributes three distinct innovations to the literature: 1) a lightweight VLM architecture that fits within wearable retina imaging devices through structured pruning and knowledge distillation; 2) a federated aggregation protocol with provable convergence; and 3) an explainable module that provides vessel geometry analysis with clinical meaning, aligned with established ROP diagnosis criteria including plus disease indicators, zone classification, and tortuosity. Using TeleROP-VLF in five neonatal intensive care units, we observe diagnostic performance comparable to expert ophthalmologists at 87.6% accuracy ultimately producing on-device inference that processes within 2.3 seconds on an autonomous edge computing platform while providing differential privacy guarantees and retaining a privacy budget of ϵ = 1.8. Coupled with clinical validation of explainable vascular geometric features, TeleROP-VLF aids clinician trust and facilitates early interventions to prevent vision loss for vulnerable premature infants.

The rapid adoption of electric vehicles (EVs) has heightened the need for advanced diagnostic frameworks to ensure the safety, reliability, and longevity of battery systems. Batteries, as the cornerstone of EV performance, are susceptible to capacity degradation, thermal instability, and anomalies that compromise efficiency and safety. Traditional diagnostic approaches, relying on physics-based models and heuristic thresholds, often fail to address the complexity and variability of real-world EV operations, necessitating scalable, efficient, and privacy-preserving solutions. This paper presents a Split Multi-Task Federated Learning (SMTFL) framework that leverages a Transformer-based architecture to address the dual challenges of battery health anomaly detection and capacity estimation. By combining the efficiency of split learning with the collaborative benefits of federated learning, the proposed system enables EVs to train and share partial model updates while retaining sensitive data locally. Experimental results on a recent large-scale EV dataset demonstrate the superior performance of the SMTFL framework, achieving an accuracy of 98.9% for battery health anomaly detection and a mean absolute error (MAE) of 0.32 Ah for capacity estimation, corresponding to a relative error of approximately 1.80%, significantly outperforming state-of-the-art approaches. By training a single model for both tasks and performing computations during charging sessions, the framework minimizes energy consumption and computational overhead, ensuring practicality for large-scale deployment in real-world EV systems.