This review provides an in-depth analysis of current hardware acceleration approaches for image processing and neural network inference, focusing on key operations involved in these applications and the hardware platforms used to deploy them. We examine various solutions, including traditional CPU–GPU systems, custom ASIC designs, and FPGA implementations, while also considering emerging low-power, resource-constrained devices.

Unmanned aerial vehicles (UAVs) equipped with thermal cameras are increasingly used in search and rescue (SAR) operations, where low visibility and small human footprints make detection a critical challenge. Existing datasets are mostly limited to urban or open-field scenarios, and our experiments show that models trained on such heterogeneous data achieve poor results. To address this gap, we collected and annotated thermal images in mountainous environments using a DJI M3T drone under clear daytime conditions. This mountain-specific set was integrated with ten existing sources to form an extensive benchmark of over 75,000 images. We then performed a comparative evaluation of object detection models (YOLOv8/9/10, RT-DETR) and semantic segmentation networks (U-Net variants), analyzing accuracy, inference speed, and energy consumption on an NVIDIA Jetson AGX Orin. Results demonstrate that human detection tasks can be accurately solved through both semantic segmentation and object detection, achieving 90% detection accuracy using segmentation models and 85% accuracy using the YOLOv8 X detection model in mountain scenarios. On the Jetson platform, segmentation achieves real-time performance with up to 27 FPS in FP16 mode. Our contributions are as follows: (i) the introduction of a new mountainous thermal image collection extending current benchmarks and (ii) a comprehensive evaluation of detection methods on embedded hardware for SAR applications.

Registering UAV-based thermal and visible images is a challenging task due to differences in appearance across spectra and the lack of public benchmarks. To address this issue, we introduce UAV-TIRVis, a dataset consisting of 80 accurately and manually registered UAV-based thermal (640 × 512) and visible (4K) image pairs, captured across diverse environments. We benchmark our dataset using well-known registration methods, including feature-based (ORB, SURF, SIFT, KAZE), correlation-based, and intensity-based methods, as well as a custom, heuristic intensity-based method. We evaluate the performance of these methods using four metrics: RMSE, PSNR, SSIM, and NCC, averaged per scenario and across the entire dataset. The results show that conventional methods often fail to generalize across scenes, yielding <0.6 NCC on average, whereas the heuristic method shows that it is possible to achieve 0.77 SSIM and 0.82 NCC, highlighting the difficulty of cross-spectral UAV alignment and the need for further research to improve optimization in existing registration methods.

This paper introduces a deep learning approach to photorealistic universal style transfer that extends the PhotoNet network architecture by adding extra feature-aggregation modules. Given a pair of images representing the content and the reference of style, we augment the state-of-the-art solution mentioned above with deeper aggregation, to better fuse content and style information across the decoding layers. As opposed to the more flexible implementation of PhotoNet (i.e., PhotoNAS), which targets the minimization of inference time, our method aims to achieve better image reconstruction and a more pleasant stylization. We propose several deep layer aggregation architectures to be used as wrappers over PhotoNet, to enhance the stylization and quality of the output image.

Thermal drones help to search for missing people in the wilderness, but flying high enough to cover wide areas shrinks each person to just a few pixels, causing standard detectors to miss them. We introduce MicroSlicer, a lightweight method that improves small-target detection by slicing each frame into overlapping tiles at inference time, with no extra training required. Tested on three detector architectures (YOLO26-nano, YOLO26-small, RT-DETR-Large) using the WiSARD thermal search-and-rescue dataset, MicroSlicer improves target recall (+11.7 to +18.9 percentage points overall and up to +22.7 percentage points on small targets) over standard full-frame inference. High-resolution scaling experiments further indicate that, while MicroSlicer yields higher small-target recall on YOLO26s and RT-DETR-L, single-pass full-frame inference at 1536 px achieves higher recall and throughput on the ultra-lightweight YOLO26n, demonstrating that the trade-off between spatial slicing and global input scaling depends on the model capacity and architecture. When benchmarked against an existing slicing framework, SAHI, with matched non-maximum suppression (NMS) and a slices-only configuration, MicroSlicer achieves similar overall detection accuracy while maintaining higher target recall across all architectures and reducing the post-processing latency via compiled C++ tensor operations. Furthermore, on Transformer detectors under default merging settings, SAHI exhibits substantial latency increases and bounding-box dilation at low confidence thresholds, whereas MicroSlicer maintains stable deduplication ($0.11–0.14$ ms at operational confidence and $2.1–3.6$ ms under dense candidate sets at low thresholds) and consistent accuracy. We also found that full-image training combined with MicroSlicer performed best for YOLO26s and RT-DETR-L (by up to 21.0 percentage points over patch-trained models), whereas YOLO26n benefited from sliced training. While currently evaluated on a single thermal dataset using a desktop GPU workstation, these results establish MicroSlicer as a practical, training-free option for edge-assisted search-and-rescue operations, with field validation on low-power embedded hardware identified as essential future work.