Graphics Processing Units Memory Research Articles

Iterative Reconstruction of Micro Computed Tomography Scans Using Multiple Heterogeneous GPUs.

Graphics processing units (GPUs) facilitate massive parallelism and high-capacity storage, and thus are suitable for the iterative reconstruction of ultrahigh-resolution micro computed tomography (CT) scans by on-the-fly system matrix (OTFSM) calculation using ordered subsets expectation maximization (OSEM). We propose a finite state automaton (FSA) method that facilitates iterative reconstruction using a heterogeneous multi-GPU platform through parallelizing the matrix calculations derived from a ray tracing system of ordered subsets. The FSAs perform flow control for parallel threading of the heterogeneous GPUs, which minimizes the latency of launching ordered-subsets tasks, reduces the data transfer between the main system memory and local GPU memory, and solves the memory-bound of a single GPU. In the experiments, we compared the operation efficiency of OS-MLTR for three reconstruction environments. The heterogeneous multiple GPUs with job queues for high throughput calculation speed is up to five times faster than the single GPU environment, and that speed up is nine times faster than the heterogeneous multiple GPUs with the FIFO queues of the device scheduling control. Eventually, we proposed an event-triggered FSA method for iterative reconstruction using multiple heterogeneous GPUs that solves the memory-bound issue of a single GPU at ultrahigh resolutions, and the routines of the proposed method were successfully executed on each GPU simultaneously.

Multi-residual 2D network integrating spatial correlation for whole heart segmentation

Whole heart segmentation (WHS) has significant clinical value for cardiac anatomy, modeling, and analysis of cardiac function. This study aims to address the WHS accuracy on cardiac CT images, as well as the fast inference speed and low graphics processing unit (GPU) memory consumption required by practical clinical applications. Thus, we propose a multi-residual two-dimensional (2D) network integrating spatial correlation for WHS. The network performs slice-by-slice segmentation on three-dimensional cardiac CT images in a 2D encoder–decoder manner. In the network, a convolutional long short-term memory skip connection module is designed to perform spatial correlation feature extraction on the feature maps at different resolutions extracted by the sub-modules of the pre-trained ResNet-based encoder. Moreover, a decoder based on the multi-residual module is designed to analyze the extracted features from the perspectives of multi-scale and channel attention, thereby accurately delineating the various substructures of the heart. The proposed method is verified on a dataset of the multi-modality WHS challenge, an in-house WHS dataset, and a dataset of the abdominal organ segmentation challenge. The dice, Jaccard, average symmetric surface distance, Hausdorff distance, inference time, and maximum GPU memory of the WHS are 0.914, 0.843, 1.066 mm, 15.778 mm, 9.535 s, and 1905 MB, respectively. The proposed network has high accuracy, fast inference speed, minimal GPU memory consumption, strong robustness, and good generalization. It can be deployed to clinical practical applications for WHS and can be effectively extended and applied to other multi-organ segmentation fields. The source code is publicly available at https://github.com/nancy1984yan/MultiResNet-SC.

H3DM: A High-bandwidth High-capacity Hybrid 3D Memory Design for GPUs

Graphics Processing Units (GPUs) are widely used for modern applications with huge data sizes. However, the performance benefit of GPUs is limited by their memory capacity and bandwidth. Although GPU vendors improve memory capacity and bandwidth using 3D memory technology (HBM), many important workloads with terabytes of data still cannot fit in the provided capacity and are bound by the provided bandwidth. With a limited GPU memory capacity, programmers should handle the data movement between GPU and host memories by themselves, causing a significant programming burden. To improve programming ease, GPUs use a unified address space with the host that allows over-subscribing GPU memory, but this approach is not effective in terms of performance once GPUs encounter memory page faults. Many recent works have tried to remedy capacity and bandwidth bottlenecks using dense non-volatile memories (NVMs) and true-3D stacking. However, these works mainly focus on one bottleneck or do not provide a scalable solution that fits future requirements. In this paper, we investigate true-3D stacking of dense, low-power, and refresh-free non-volatile phase change memory (PCM) on top of state-of-the-art GPU configurations to provide higher capacity and bandwidth within the available area and power budget. The higher density and lower power consumption of PCM provide higher capacity through integrating more cells in each 3D layer and enabling stacking more layers. However, we observe that stacking more than six layers of pure-PCM memory violates the thermal constraint and severely harms the performance and power efficiency due to its higher write latency and energy. Further, it degrades the lifetime of GPU to less than one year. Utilizing a hybrid architecture that leverages the benefits of both DRAM and PCM memories has been widely studied by prior proposals; however, true-3D integration of such a hybrid memory architecture especially on top of state-of-the-art powerful GPU architecture has not been investigated yet. We experimentally demonstrate that by covering 80% of write requests in DRAM and eliminating refresh overhead, true-3D stacking of eight 32GB layers of PCM along with two 8GB layers of DRAM is possible resulting in a total of 272GB memory capacity. Based on the explored design requirements, We propose a 3D high-bandwidth high-capacity hybrid memory (H3DM) system utilizing a hybrid-3D (H3D)-aware remapping scheme to reduce expensive PCM writes to under 20% while avoiding DRAM refresh overhead. H3DM improves the performance up to 291% compared to the baseline GPU architecture while remaining within only 3% of an ideal case with DRAM-like access latency, on average. Moreover, by increasing the dataset size above the baseline GPU memory space, H3DM improves performance and power up to 648% and 87% compared to the baseline GPU architecture since it avoids expensive data transfers through off-chip communication links.

Soft- and Hard-Kill Hybrid GPU-based Bi-Directional Evolutionary Structural Optimization

Abstract Bi-directional evolutionary structural optimization (BESO) is a well-recognized method for generating optimal topologies of structures. Its soft-kill variant has a high computational cost, especially for large-scale structures, whereas the hard-kill variant often faces convergence issues. Addressing these issues, this paper proposes a hybrid BESO model tailored for graphics processing units (GPUs) by combining the soft-kill and hard-kill approaches for large-scale structures. A GPU-based algorithm is presented for dynamically isolating the solid/hard elements from the void/soft elements in the finite element analysis (FEA) stage. The hard-kill approach is used in the FEA stage with an assembly-free solver to facilitate the use of high-resolution meshes without exceeding the GPU memory limits, whereas for the rest of the optimization procedure, the soft-kill approach with a material interpolation scheme is implemented. Furthermore, the entire BESO method pipeline is accelerated for both the proposed hybrid and the standard soft-kill BESO. The comparison of the hybrid BESO with the GPU-accelerated soft-kill BESO using four benchmark problems with more than two million degrees of freedom reveals three key benefits of the proposed hybrid model: reduced execution time, decreased memory consumption, and improved FEA convergence, all of which mitigate the major computational issues associated with BESO.

Memory-efficient semantic segmentation of large microscopy images using graph-based neural networks.

We present a graph neural network (GNN)-based framework applied to large-scale microscopy image segmentation tasks. While deep learning models, like convolutional neural networks (CNNs), have become common for automating image segmentation tasks, they are limited by the image size that can fit in the memory of computational hardware. In a GNN framework, large-scale images are converted into graphs using superpixels (regions of pixels with similar color/intensity values), allowing us to input information from the entire image into the model. By converting images with hundreds of millions of pixels to graphs with thousands of nodes, we can segment large images using memory-limited computational resources. We compare the performance of GNN- and CNN-based segmentation in terms of accuracy, training time and required graphics processing unit memory. Based on our experiments with microscopy images of biological cells and cell colonies, GNN-based segmentation used one to three orders-of-magnitude fewer computational resources with only a change in accuracy of $-2\;%$ to $+0.3\;%$. Furthermore, errors due to superpixel generation can be reduced by either using better superpixel generation algorithms or increasing the number of superpixels, thereby allowing for improvement in the GNN framework's accuracy. This trade-off between accuracy and computational cost over CNN models makes the GNN framework attractive for many large-scale microscopy image segmentation tasks in biology.

GPUTucker: Large-Scale GPU-Based Tucker Decomposition Using Tensor Partitioning

Tucker decomposition is used extensively for modeling multi-dimensional data represented as tensors. Owing to the increasing magnitude of nonzero values in real-world tensors, a growing demand has emerged for expeditious and scalable Tucker decomposition techniques. Several graphics processing unit(GPU)-accelerated techniques have been proposed for Tucker decomposition to decrease the decomposition speed. However, these approaches often encounter difficulties in handling extensive tensors owing to their huge memory demands, which exceed the available capacity of GPU memory. This study presents an expandable GPU-based technique for Tucker decomposition called GPUTucker. The proposed method meticulously partitions sizable tensors into smaller sub-tensors, which are referred to as tensor blocks, and effectively implements the GPU-based data pipeline by handling these tensor blocks asynchronously. Extensive experiments demonstrate that GPUTucker outperforms state-of-the-art Tucker decomposition methods in terms of the decomposition speed and scalability.

Effect of image resolution on automated classification of chest X-rays.

Deep learning (DL) models have received much attention lately for their ability to achieve expert-level performance on the accurate automated analysis of chest X-rays (CXRs). Recently available public CXR datasets include high resolution images, but state-of-the-art models are trained on reduced size images due to limitations on graphics processing unit memory and training time. As computing hardware continues to advance, it has become feasible to train deep convolutional neural networks on high-resolution images without sacrificing detail by downscaling. This study examines the effect of increased resolution on CXR classification performance. We used the publicly available MIMIC-CXR-JPG dataset, comprising 377,110 high resolution CXR images for this study. We applied image downscaling from native resolution to , , , and and then we used the DenseNet121 and EfficientNet-B4 DL models to evaluate clinical task performance using these four downscaled image resolutions. We find that while some clinical findings are more reliably labeled using high resolutions, many other findings are actually labeled better using downscaled inputs. We qualitatively verify that tasks requiring a large receptive field are better suited to downscaled low resolution input images, by inspecting effective receptive fields and class activation maps of trained models. Finally, we show that stacking an ensemble across resolutions outperforms each individual learner at all input resolutions while providing interpretable scale weights, indicating that diverse information is extracted across resolutions. This study suggests that instead of focusing solely on the finest image resolutions, multi-scale features should be emphasized for information extraction from high-resolution CXRs.

GPU-HADVPPM V1.0: a high-efficiency parallel GPU design of the piecewise parabolic method (PPM) for horizontal advection in an air quality model (CAMx V6.10)

Abstract. With semiconductor technology gradually approaching its physical and thermal limits, graphics processing units (GPUs) are becoming an attractive solution for many scientific applications due to their high performance. This paper presents an application of GPU accelerators in an air quality model. We demonstrate an approach that runs a piecewise parabolic method (PPM) solver of horizontal advection (HADVPPM) for the air quality model CAMx on GPU clusters. Specifically, we first convert the HADVPPM to a new Compute Unified Device Architecture C (CUDA C) code to make it computable on the GPU (GPU-HADVPPM). Then, a series of optimization measures are taken, including reducing the CPU–GPU communication frequency, increasing the data size computation on the GPU, optimizing the GPU memory access, and using thread and block indices to improve the overall computing performance of the CAMx model coupled with GPU-HADVPPM (named the CAMx-CUDA model). Finally, a heterogeneous, hybrid programming paradigm is presented and utilized with GPU-HADVPPM on the GPU clusters with a message passing interface (MPI) and CUDA. The offline experimental results show that running GPU-HADVPPM on one NVIDIA Tesla K40m and an NVIDIA Tesla V100 GPU can achieve up to a 845.4× and 1113.6× acceleration. By implementing a series of optimization schemes, the CAMx-CUDA model results in a 29.0× and 128.4× improvement in computational efficiency by using a GPU accelerator card on a K40m and V100 cluster, respectively. In terms of the single-module computational efficiency of GPU-HADVPPM, it can achieve 1.3× and 18.8× speedup on an NVIDIA Tesla K40m GPU and NVIDIA Tesla V100 GPU, respectively. The multi-GPU acceleration algorithm enables a 4.5× speedup with eight CPU cores and eight GPU accelerators on a V100 cluster.

Fully 3D implementation of the end-to-end deep image prior-based PET image reconstruction using block iterative algorithm

Objective. Deep image prior (DIP) has recently attracted attention owing to its unsupervised positron emission tomography (PET) image reconstruction method, which does not require any prior training dataset. In this paper, we present the first attempt to implement an end-to-end DIP-based fully 3D PET image reconstruction method that incorporates a forward-projection model into a loss function. Approach. A practical implementation of a fully 3D PET image reconstruction could not be performed at present because of a graphics processing unit memory limitation. Consequently, we modify the DIP optimization to a block iteration and sequential learning of an ordered sequence of block sinograms. Furthermore, the relative difference penalty (RDP) term is added to the loss function to enhance the quantitative accuracy of the PET image. Main results. We evaluated our proposed method using Monte Carlo simulation with [18F]FDG PET data of a human brain and a preclinical study on monkey-brain [18F]FDG PET data. The proposed method was compared with the maximum-likelihood expectation maximization (EM), maximum a posteriori EM with RDP, and hybrid DIP-based PET reconstruction methods. The simulation results showed that, compared with other algorithms, the proposed method improved the PET image quality by reducing statistical noise and better preserved the contrast of brain structures and inserted tumors. In the preclinical experiment, finer structures and better contrast recovery were obtained with the proposed method. Significance. The results indicated that the proposed method could produce high-quality images without a prior training dataset. Thus, the proposed method could be a key enabling technology for the straightforward and practical implementation of end-to-end DIP-based fully 3D PET image reconstruction.

CUDA-bigPSF: An optimized version of bigPSF accelerated with graphics processing Unit

Accurate and fast short-term load forecasting is crucial in efficiently managing energy production and distribution. As such, many different algorithms have been proposed to address this topic, including hybrid models that combine clustering with other forecasting techniques. One of these algorithms is bigPSF, an algorithm that combines K-means clustering and a similarity search optimized for its use in distributed environments. The work presented in this paper aims to improve the time required to execute the algorithm with two main contributions. First, some of the issues of the original proposal that limited the number of cores simultaneously used are studied and highlighted. Second, a version of the algorithm optimized for Graphics Processing Unit (GPU) is proposed, solving the previously mentioned issues while taking into account the GPU architecture and memory structure. Experimentation was done with seven years of real-world electric demand data from Uruguay. Results show that the proposed algorithm executed consistently faster than the original version, achieving speedups up to 500 times faster during the training phase.

Computationally effective estimation of supersonic flow field around airfoils using sparse convolutional neural network

This work proposes an innovative approach for supersonic flow field modeling around airfoils based on sparse convolutional neural networks (SCNNs) and Bézier generative adversarial network (GAN), where (1) the SCNN model is built to end-to-end predict supersonic compressible physical flow fields around airfoils from spatially-sparse geometries and (2) the trained Bézier-GAN is utilized to generate plenty of smooth airfoils as well as the latent codes representing airfoils. The spatially-sparse positions of airfoil geometry are represented using signed distance function (SDF). Particularly, the latent codes are merged with the SDF matrix and the Mach number to form the input of the SCNN model, effectively making the SCNN model possess more robust geometric adaptability to different flow conditions. The most significant contribution compared to the regular convolutional neural network is that SCNN introduces sparse convolutional operations to process spatially-sparse input matrix, specifically, which only focuses on the local area with flow information when performing convolution, eventually saving memory usage and improving the network’s attention on the flow area. Further, the testing results show that the SCNN model can more accurately predict supersonic flow fields with a mean absolute error lower than 5% and save 40% of graphics processing unit memory. These results indicate that the proposed SCNN model can capture the shock wave features of supersonic flow fields and improve learning efficiency and computing efficiency.

Single‐image super‐resolution using lightweight transformer‐convolutional neural network hybrid model

AbstractWith constant advances in deep learning methods as applied to image processing, deep convolutional neural networks (CNNs) have been widely explored in single‐image super‐resolution (SISR) problems and have attained significant success. These CNN‐based methods cannot fully use the internal and external information of the image. The authors add a lightweight Transformer structure to capture this information. Specifically, the authors apply a dense block structure and residual connection to build a residual dense convolution block (RDCB) that reduces the parameters somewhat and extracts shallow features. The lightweight transformer block (LTB) further extracts features and learns the texture details between the patches through the self‐attention mechanism. The LTB comprises an efficient multi‐head transformer (EMT) with small graphics processing unit (GPU) memory footprint, and benefits from feature preprocessing by multi‐head attention (MA), reduction, and expansion. The EMT significantly reduces the use of GPU resources. In addition, a detail‐purifying attention block (DAB) is proposed to explore the context information in the high‐resolution (HR) space to recover more details. Extensive evaluations of four benchmark datasets demonstrate the effectiveness of the authors’ proposed model in terms of quantitative metrics and visual effects. The proposed EMT only uses about 40% as much GPU memory as other methods, with better performance.

A novel GPU based Geo-Location Inference Attack on WebGL framework

In the past few years, graphics processing units (GPUs) have become an indispensable part of modern computer systems, not only for graphics rendering but also for intensive parallel computing. Given that many tasks running on GPUs contain sensitive information, security concerns have been raised, especially about potential GPU information leakage. Previous works have shown such concerns by showing that attackers can use GPU memory allocations or performance counters to measure victim side effects. However, such an attack has a critical drawback that it requires a victim to install desktop applications or mobile apps yielding it uneasy to be deployed in the real world. In this paper, we solve this drawback by proposing a novel GPU-based side-channel Geo-Privacy inference attack on the WebGL framework, namely, GLINT (stands for Geo-Location Inference Attack). GLINT merely utilizes a lightweight browser extension to measure the time elapsed to render a sequence of frames on well-known map websites, e.g., Google Maps, or Baidu Maps. The measured stream of time series is then employed to infer geologically privacy-sensitive information, such as a search on a specific location. Upon retrieving the stream, we propose a novel online segmentation algorithm for streaming data to determine the start and end points of privacy-sensitive time series. We then combine the DTW algorithm and KNN algorithm on these series to conclude the final inference on a user’s geo-location privacy.We conducted real-world experiments to testify our attack. The experiments show that GeoInfer can correctly infer more than 83% of user searches regardless of the locations and map websites, meaning that our Geo-Privacy inference attack is accurate, practical, and robust. To counter this attack, we implemented a defense strategy based on Differential Privacy to hinder obtaining accurate rendering data. We found that this defense mechanism managed to reduce the average accuracy of the attack model by more than 70%, indicating that the attack was no longer effective. We have fully implemented GLINT and open-sourced it for future follow-up research.

Optimized IANSegNet: Deep Segmentation for the Detection of Inferior Alveolar Nerve Canal

Imaging studies in dentistry and maxillofacial pathology have recently concentrated on detecting the inferior alveolar nerve (IAN) canal. In spite of the minor dimensions of 3D maxillofacial datasets, deep learning-based algorithms have shown encouraging consequences in this study area. This study describes a mandibular cone-beam CT (CBCT) dataset with 2D and 3D hand comments. It is huge and freely available. It was possible to utilise this dataset by applying the residual neural network (IANSegNet), which consumed less GPU memory and computational complexity. As an encoder, IANSegNet uses the computationally efficient 3D ShuffleNetV2 network to reduce graphics processing unit (GPU) memory usage and improve efficiency. After that, a decoder with leftover blocks is added to keep the quality high. To address network convergence and data inequity, Dice’s loss and cross-entropy loss were created. Optimized postprocessing techniques are also recommended for fine-tuning the coarse segmentation findings that are generated by IANSegNet. The results of the validation show that IANSegNet outperformed other deep learning models in a variety of criteria.

Large‐scale 3D fast Fourier transform computation on a GPU

AbstractWe propose a novel graphics processing unit (GPU) algorithm that can handle a large‐scale 3D fast Fourier transform (i.e., 3D‐FFT) problem whose data size is larger than the GPU's memory. A 1D FFT‐based 3D‐FFT computational approach is used to solve the limited device memory issue. Moreover, to reduce the communication overhead between the CPU and GPU, we propose a 3D data‐transposition method that converts the target 1D vector into a contiguous memory layout and improves data transfer efficiency. The transposed data are communicated between the host and device memories efficiently through the pinned buffer and multiple streams. We apply our method to various large‐scale benchmarks and compare its performance with the state‐of‐the‐art multicore CPU FFT library (i.e., fastest Fourier transform in the West [FFTW]) and a prior GPU‐based 3D‐FFT algorithm. Our method achieves a higher performance (up to 2.89 times) than FFTW; it yields more performance gaps as the data size increases. The performance of the prior GPU algorithm decreases considerably in massive‐scale problems, whereas our method's performance is stable.

Space-to-speed architecture supporting acceleration on VHR image processing

One of the major focuses in the remote sensing community is the rapid processing of deep neural networks (DNNs) on very high resolution (VHR) aerial images. Few studies have investigated the acceleration of training and prediction by optimizing the architecture of the DNN system rather than designing a lightweight DNN. Parallel processing using multiple graphics processing units (GPUs) increases VHR image processing performance. It drives extremely large and frequent data transfers (input/output(I/O)) from random access memory (RAM) to GPU memory. As a result, the system bus congestion causes the system to hang, resulting in long latency in training/predicting. In this paper, we evaluate the causes of long latency and propose a space-to-speed (S2S) DNN system to overcome the aforementioned challenges. A three-level memory system aiming to reduce data transfer during system operation is presented. Distribution optimization with parallel processing was used to accelerate the training. Training optimizations on VHR images (such as hot-zone searching and image/ground truth queues for data saving) were used to train the VHR images efficiently. Inference optimization was performed to speed up prediction in the release mode. To verify the efficiency of the proposed system, we used aerial image labeling from the Institut National de Recherche en Informatique et en Automatique (INRIA) and benchmarks from the Massachusetts Institute of Technology Aerial Imagery for Roof Segmentation (MITAIRS) to test the system performance and accuracy. Without the loss of accuracy, the S2S system improved prediction speed on the testing dataset by eight GPUs in a normal setting in both the INRIA dataset (from 534 to 72 s) and the MITAIRS dataset (818 to 120 s). With the prediction in half-float (using float-16 data), an 8-GPU parallel processing increased the speed to 38 s in the INRIA dataset and 83 s in the MITAIRS dataset. In a pressure test, our proposed system operated on 18,000 images with a size of 5000 × 5000 from 18.2 to 1.8 h with the prediction in full-float (using float-32 data) and 43 min with the prediction in half-float, increasing the speed by a factor of 9.78 and 25.3, respectively, when compared to system runs without optimization.

Network-Assisted Noncontiguous Transfers for GPU-Aware MPI Libraries

The importance of graphics processing units (GPUs) in accelerating HPC applications is evident by the fact that a large number of supercomputing clusters are GPU enabled. Many of these HPC applications use message passing interface (MPI) as their programming model. These MPI applications frequently exchange data that is noncontiguous in GPU memory. MPI provides derived datatypes (DDTs) to represent such data. Past research on DDTs mainly focused on optimizing the pack–unpack kernels. Modern HCAs are capable of gathering/scattering data from/to noncontiguous GPU memory regions. We propose a low-overhead HCA-assisted scheme to improve the performance of GPU-based noncontiguous exchanges without the GPU-based pack–unpack kernels. We show that the proposed scheme provides up to 2× benefits compared to the existing pack-based scheme at the benchmark level. Furthermore, we show up to 17% improvement with the SW4Lite application compared to other MPI libraries, such as MVAPICH2-GDR and OpenMPI+UCX.

Aerial-Based Weed Detection Using Low-Cost and Lightweight Deep Learning Models on an Edge Platform

Highlights Lightweight deep learning models were trained on an edge device to identify weeds in aerial images. A customized configuration file was setup to train the models. These models were deployed to detect weeds in aerial images and videos (near real-time). CSPMobileNet-v2 and YOLOv4-lite are recommended models for weed detection using edge platform. Abstract. Deep learning (DL) techniques have proven to be a successful approach in detecting weeds for site-specific weed management (SSWM). In the past, most of the research work has trained and deployed pre-trained DL models on high-end systems coupled with expensive graphical processing units (GPUs). However, only a limited number of research studies have used DL models on an edge system for aerial-based weed detection. Therefore, while focusing on hardware cost minimization, eight DL models were trained and deployed on an edge device to detect weeds in aerial-image context and videos in this study. Four large models, namely CSPDarkNet-53, DarkNet-53, DenseNet-201, and ResNet-50, along with four lightweight models, CSPMobileNet-v2, YOLOv4-lite, EfficientNet-B0, and DarkNet-Ref, were considered for training a customized DL architecture. Along with trained model performance scores (average precision score, mean average precision (mAP), intersection over union, precision, and recall), other model metrics to assess edge system performance such as billion floating-point operations/s (BFLOPS), frame rates/s (FPS), and GPU memory usage were also estimated. The lightweight CSPMobileNet-v2 and YOLOv4-lite models outperformed others in detecting weeds in aerial image context. These models were able to achieve a mAP score of 83.2% and 82.2%, delivering an FPS of 60.9 and 61.1 during near real-time weed detection in aerial videos, respectively. The popular ResNet-50 model achieved a mAP of 79.6%, which was the highest amongst all the large models deployed for weed detection tasks. Based on the results, the two lightweight models, namely, CSPMobileNet-v2 and YOLOv4-lite, are recommended, and they can be used on a low-cost edge system to detect weeds in aerial image context with significant accuracy. Keywords: Aerial image, Deep learning, Edge device, Precision agriculture, Weed detection.

Lightweight Deep Learning Models for High-Precision Rice Seedling Segmentation from UAV-Based Multispectral Images.

Accurate segmentation and detection of rice seedlings is essential for precision agriculture and high-yield cultivation. However, current methods suffer from high computational complexity and poor robustness to different rice varieties and densities. This article proposes 2 lightweight neural network architectures, LW-Segnet and LW-Unet, for high-precision rice seedling segmentation. The networks adopt an encoder-decoder structure with hybrid lightweight convolutions and spatial pyramid dilated convolutions, achieving accurate segmentation while reducing model parameters. Multispectral imagery acquired by unmanned aerial vehicle (UAV) was used to train and test the models covering 3 rice varieties and different planting densities. Experimental results demonstrate that the proposed LW-Segnet and LW-Unet models achieve higher F1-scores and intersection over union values for seedling detection and row segmentation across varieties, indicating improved segmentation accuracy. Furthermore, the models exhibit stable performance when handling different varieties and densities, showing strong robustness. In terms of efficiency, the networks have lower graphics processing unit memory usage, complexity, and parameters but faster inference speeds, reflecting higher computational efficiency. In particular, the fast speed of LW-Unet indicates potential for real-time applications. The study presents lightweight yet effective neural network architectures for agricultural tasks. By handling multiple rice varieties and densities with high accuracy, efficiency, and robustness, the models show promise for use in edge devices and UAVs to assist precision farming and crop management. The findings provide valuable insights into designing lightweight deep learning models to tackle complex agricultural problems.

Autosegmentation of brain metastases using 3D FCNN models and methods to manage GPU memory limitations

Aims. To explore the efficacy of two different approaches to train a Fully Convolutional Neural Network (FCNN) with Graphical Processing Unit (GPU) memory limitations and investigate if pre-trained two-dimensional weights can be transferred into a three-dimensional model for the purpose of brain tumour segmentation. Materials and methods. Models were developed in Python using TensorFlow and Keras. T1 contrast-enhanced MRI scans and associated contouring data from 104 patients were used to train and validate the model. The data was resized to one-quarter of its original resolution, and the original data was also split into four quarters for comparison to fit within GPU limitations. Transferred weights from a two-dimensional VGG16 model trained on ImageNet were transformed into three dimensions for comparison with randomly generated initial weights. Results. Resizing the data produced superior dice similarity coefficients with fewer false positives than quartering the data. Quartering the data yielded a superior sensitivity. Transforming and transferring two-dimensional weights was not able to consistently produce improvement in training or final metrics. Conclusion. For segmentation of brain tumours, resizing the data results in better performance than quartering the data. For the model and approaches used in this report, transferring weights were not able to demonstrate any benefit.