Fusing medical images from diverse modalities like MRI, PET, and SPECT helps improve diagnostic precision by combining their complementary features. However, existing deep learning-based fusion methods often suffer from limited detail preservation and inefficient attention modeling across spatial and channel dimensions. This study introduces an innovative framework called hierarchical attention fusion network (HAF-Net) for robust and high-quality medical image fusion. The proposed model incorporates a hierarchical feature aggregation (HFA) module to extract scale-adaptive features, and a residual attention convolution (RAC) block to enhance fine-grained details using gradient-aware spatial and frequency-domain information. Furthermore, a multispectral frequency-aware channel attention (MFCA) mechanism is introduced to capture discriminative features across multiple frequency bands, and a cross-interaction attention module (CIAM) is designed to jointly model spatial-channel relationships. An adaptive fusion weighting (AFW) strategy is employed to dynamically combine multi-scale features based on their contextual relevance. Extensive experiments on standard PET/MRI and SPECT/MRI datasets demonstrate that HAF-Net achieves superior performance compared to state-of-the-art fusion methods. The results validate the effectiveness of the proposed modules in preserving structural integrity and enhancing detail in fused medical images.

Accidents involving escalators in mass rapid transit (MRT) systems pose a serious risk to public safety, often resulting from clothing or footwear getting caught, or large items toppling during movement. Despite the availability of passive warnings, such as signage and audio announcements, these methods often go unnoticed by commuters and lack the ability to adapt to real-time risks. Existing computer vision solutions are either too computationally intensive for deployment on edge devices or lack sufficient accuracy for practical use. To address these challenges, this study proposes a real-time, lightweight object detection system using a pruned YOLOv7-Tiny model, optimized for deployment on the NVIDIA Jetson Nano edge computing platform. The system is designed to identify safety-critical items, such as general footwear, high heels, long skirts, suitcases, strollers, and shopping trolleys, in real-time. Upon detection, it issues visual and auditory alerts, and in cases involving large items, sends email notifications to station personnel. Model pruning significantly reduces computational overhead while maintaining high accuracy. Experimental results demonstrate that the system achieves a mean average precision (mAP) of 94.69%, outperforming conventional detection models while maintaining real-time performance. These results highlight the system’s potential for enhancing passenger safety and operational efficiency in resource-constrained public transit environments.

Bird excrement deposited on solar panels can lead to hotspots, significantly reducing the efficiency of solar power plants. This article presents a novel solution to this problem leveraging unmanned aerial vehicle (UAV) systems for the automated geolocation and removal of bird excrement across large-scale solar power facilities. First, a UAV executes a predefined flight path to capture sequential aerial images of the plant. These images are subsequently stitched to produce a high-definition orthomosaic of the entire facility. An advanced detection framework based on YOLOv7, enhanced with an attention module, is employed to accurately detect bird excrement by reducing background noise and highlighting key features. An additional prediction head is integrated to improve detection of smaller bird excrements. To compute precise geolocation of the detected excrement, the midpoint pixel coordinates of the excrement along with the azimuth angle and actual ground distance (AGD) relative to a ground control point (GCP) is used. This article further proposes a cleaning technique that employs a traveling salesman problem (TSP) approximation algorithm to efficiently optimize flight path of the cleaning UAV. Experimental results indicate the system achieves an average detection precision (AP) of 93.91% and GPS coordinate accuracy with an average error of 0.149 m, demonstrating the efficacy of the proposed method in both geolocation and removal of bird excrement from solar panels.

In the era of rapidly advancing technology, food classification has emerged as a pivotal application across various domains including health monitoring, dietary assessment, and culinary innovation. However, efficiently categorizing food items remains a challenge, particularly in scenarios with limited labeled data. This paper introduces a novel approach for few-shot food classification using Prototypical Networks with ResNet-50 and an attention mechanism as embedding network. Leveraging the inherent capability of Prototypical Networks to learn from scarce examples, our method demonstrates exceptional adaptability and accuracy in classifying food items. Through extensive experimentation on the Food-101 dataset, employing various CNN architectures, our findings underscore the effectiveness of our approach. In particular, ResNet-50 integrated with the attention mechanism surpasses other architectures, achieving superior classification accuracies of 91.5% and 95.2% for 1-shot and 5-shot learning scenarios, respectively. This integrated approach showcases the potential of Prototypical Networks in addressing the challenges of limited labeled data in food classification tasks, marking a significant advancement in the field.

The accumulation of bird droppings on photovoltaic (PV) farms reduces power generation efficiency and necessitates manual cleaning on a regular basis, which is a challenge in large power plants. To solve this problem, this paper proposes an automatic Unmanned Aerial Vehicle (UAV) based bird droppings detection, localization, and cleaning method on large PV power plant. An automated flight route is first created, and use an UAV to fly over the solar farm to capture images of the solar panels. The captured images are then stitched together to create a high-resolution orthomosaic image of the solar farm, which enables to precisely locate the bird droppings on the solar farm. An improved YOLOv7-based model is proposed to detect the bird droppings because they are quite small in comparison to the stitched image. Then, using the ground sample distance, we calculate the distance between each of the bird droppings and the drone's takeoff point, which is used to clean the bird droppings from the solar panel. Last, the proposed model is verified by high-resolution orthomosaic images and the experimental outcomes unequivocally show that it is successful for detecting and cleaning of bird droppings on PV farms.