22 October 2025 in IEEE Access
Image super-resolution (SR) is a fundamental yet challenging task in computer vision, aiming to reconstruct high-resolution (HR) images from low-resolution (LR) inputs. To address these limitations, this paper proposes a novel super-resolution framework that integrates several advanced architectural components to enhance image reconstruction quality. At its core, the framework introduces a Dense Residual Dense Block (DRDB), which incorporates residual dense connectivity and dual-pathway processing to effectively extract and fuse both low- and high-level features. This design improves feature representation and mitigates gradient vanishing in deeper networks. To enhance flexibility and visual consistency, Adaptive Instance Normalization (AdIN) is employed to dynamically align channel-wise feature statistics while preserving structural information. Additionally, a Channel Attention Block (CAB) with a Squeeze-and-Excitation (CAB-SE) mechanism is embedded to adaptively recalibrate channel-wise and spatial feature importance, allowing the network to focus on salient regions. Experimental results on standard SR benchmarks demonstrate that the proposed method outperforms existing approaches in terms of texture reconstruction, visual fidelity, and computational efficiency. Overall, this work presents a scalable and effective solution for deep super-resolution tasks, offering improved generalization, adaptability, and feature enhancement through synergistic architectural integration.
29 May 2026 in International Journal of Intelligent Transportation Systems Research
9 March 2026 in IEEE Access
This work presents the design of a compact wireless temperature and airflow sensing method using a complementary split ring resonator (CSRR)-loaded microstrip transmission line integrated with thermistors. The sensing principle relies on the temperature-dependent resistance of the thermistor, where varying airflow temperature modulate the heating and cooling dynamics, resulting in measurable changes in the slope and amplitude of the CSRR-loaded transmission line response. To establish a predictive relationship between these dynamic features and actual air temperature variation, a linear regression model is utilized and validated using experimentally obtained datasets. The model achieves a mean square error (MSE) of less than 0.15 dB and sensitivity of 0.2 dB/°Celcius over 18% frequency bandwidth with less than 0.05 dB sensitivity variation. Further validation confirms that such a sensing mechanism can be utilized to distinctively measure airflow rate by monitoring the time-dependent variation of the transmission response and mapping these dynamic features to actual flow rate measurements. The fabricated prototype is characterized under controlled hot-air flows in the range of 11.3 – 22.1 standard liters per minute (SLPM) at a constant source temperature of 100° C. With its low-profile architecture and compatibility with low-cost fabrication techniques, the proposed sensing mechanism offers a scalable, robust, and easily integrable solution for real-time airflow monitoring in industrial ventilation, distributed sensing, and internet of things (IoT) applications.
16 March 2026 in SSRN Electronic Journal
This paper presents a computationally-efficient neural architecture for real-time 3D object classification from sparse LiDAR point clouds, addressing critical challenges in autonomous vehicles under severe computational and sensor constraints. While state-of-the-art methods like PointNet++ and Point Transformers excel with dense clouds, they impose prohibitive costs for embedded deployment with low-resolution sensors. Our Fully Convolutional Neural Network (FCNN) achieves comparable performance with order-of-magnitude computational reduction by directly processing sparse 3D point clouds from 16-channel VLP-16 LiDAR (~30,000 points vs >100,000 from high-end sensors) without 2D projections. The architecture incorporates learned surface normals and density encodings compensating for sparsity, residual skip connections with dilated convolutions for multi-scale features, and dynamically-weighted focal loss handling severe class imbalance. Evaluated on 2,591 manually-annotated urban point clouds, our method achieves 97.04% accuracy and 0.9707 F1-score with sub-100ms latency, approximately 5× faster than PointNet++ with only 2.7% accuracy trade-off. Mean 3D IoU reaches 0.79-0.80 across six categories, with strong minority class performance (F1=0.88 for cart class, 137 samples). These results demonstrate that resource-aware architectures effectively bridge computational efficiency and classification performance for cost-sensitive autonomous systems.