Reparameterization Method Research Articles

YOLOv7-Rep: a re-parameterization method for surface defect detection in workpieces

YOLOv7-Rep: a re-parameterization method for surface defect detection in workpieces

Dual-Spatial Domain Generalization for Fundus Lesion Segmentation in Unseen Manufacturer's OCT Images.

Optical Coherence Tomography (OCT) images can provide non-invasive visualization of fundus lesions; however, scanners from different OCT manufacturers largely vary from each other, which often leads to model deterioration to unseen OCT scanners due to domain shift. To produce the T-styles of the potential target domain, an Orthogonal Style Space Reparameterization (OSSR) method is proposed to apply orthogonal constraints in the latent orthogonal style space to the sampled marginal styles. To leverage the high-level features of multi-source domains and potential T-styles in the graph semantic space, a Graph Adversarial Network (GAN) is constructed to align the generated samples with the source domain samples. To align features with the same label based on the semantic feature in the graph semantic space, Graph Semantic Alignment (GSA) is performed to focus on the shape and the morphological differences between the lesions and their surrounding regions. Comprehensive experiments have been performed on two OCT image datasets. Compared to state-of-the-art methods, the proposed method can achieve better segmentation. The proposed fundus lesion segmentation method can be trained with labeled OCT images from multiple manufacturers' scanners and be tested on an unseen manufacturer's scanner with better domain generalization. The proposed method can be used in routine clinical occasions when an unseen manufacturer's OCT image is available for a patient.

Neural quantile optimization for edge–cloud networking

We seek the best traffic allocation scheme for the edge–cloud networking subject to SD-WAN architecture and burstable billing. First, we formulate a family of quantile-based integer programming problems for a fixed network topology with random parameters describing the traffic demands. Then, to overcome the difficulty caused by the discrete feature, we generalize the Gumbel-softmax reparameterization method to induce an unconstrained continuous optimization problem as a regularized continuation of the discrete problem. Finally, we introduce the Gumbel-softmax sampling neural network to solve optimization problems via unsupervised learning. The neural network structure reflects the edge–cloud networking topology and is trained to minimize the expectation of the cost function for unconstrained continuous optimization problems. The trained network works as an efficient traffic allocation scheme sampler, outperforming the random strategy in feasibility and cost value. Besides testing the quality of the output allocation scheme, we examine the generalization property of the network by increasing the time steps and the number of users. We also feed the solution to existing integer optimization solvers as initial conditions and verify the warm-starts can accelerate the short-time iteration process. The framework is general, and the decoupled feature of the random neural networks is adequate for practical implementations.

Slim-YOLOv8: A fast and accurate algorithm for surface defect detection of steel strips

Steel strip is an extremely important industrial material and is widely used in various industrial fields. During the production process, surface defects need to be detected quickly and accurately. This study proposes a model. This study proposes a new slim-YOLOv8 (lightweight YOLOv8) detection model. The model is based on YOLOv8 and adopts a lightweight design paradigm, which reduces the number of parameters of the model and enhances the detection real-time performance. At the same time, an online reparameterization method is introduced to enhance the feature extraction capability of the network without raising the inference cost, and to improve the model's detection accuracy for complex defects. Finally, an auxiliary training head that can provide richer gradient information is added to the model to help train the model while preventing model overfitting. The performance of slim-YOLOv8 in mean average precision and parameters was evaluated on the well-known steel strip surface defect detection dataset NEU-DET, reaching a mAP of 85.8% at IoU 0.50 and 50.3% in the IoU 0.50–0.95 range. This is an improvement of 8.3% and 3%, respectively, compared to the baseline model. Meanwhile, the number of parameters of the model was reduced from 3.0 M to 2.7 M, which is 7% lower than the baseline model. The experimental results show that slim-YOLOv8 uses a smaller number of parameters, but has higher accuracy and is able to detect various defects in the dataset well.

SIAlex: Species identification and monitoring based on bird sound features

The combination of deep learning and bird sound recognition is widely employed in bird species conservation monitoring. A complex network structure is not conducive for deploying bird sound recognition devices, resulting in problems such as long inference time and low efficiency. Using AlexNet as the backbone model, we explore the potential of shallow and straightforward models without complex connection techniques or attention mechanisms, named SIAlex, to recognise and classify 20 bird sound datasets, which are simultaneously validated on a 10 class UrbanSound8k dataset. Using the structural re-parameterization method, the number of model layers is reduced, computational efficiency is improved, and the inference time is significantly reduced, achieving a decoupling of training and inference time in the structure. To increase the nonlinearity of the model, a cascaded approach is utilised to increase the number of activation functions, thereby significantly improving the generalisation performance of the model. Simultaneously, in the classifier section, convolutional layer replaces the original fully connected layer, thereby reducing the inference time and increasing the feature extraction ability of the model, improving accuracy, and effectively recognising bird speech. The experimental data show that the SIAlex network on the Birdsdata dataset improves the accuracy to 93.66%, and the inference time for a piece of data is only 2.466 ms. The accuracy of the UrbanSound8k dataset reaches 96.04%, and the inference time for a piece of data is 3.031 ms. A large number of experimental comparisons have shown that the method proposed in this paper achieves good results in reducing the inference time of the model, bringing breakthroughs in the application of shallow, simple models.

基于轻量级深度学习网络RepYOLO模型的麦穗实时检测

Real-time detection has become an essential component in intelligent agriculture and industry. In this paper, a real-time wheat spike detection method based on the lightweight deep learning network RepYOLO is proposed. Addressing the small and densely packed phenotype characteristics of wheat spikes, the channel attention mechanism module CBAM from the traditional YOLOv4 algorithm is introduced and multiple convolutional kernels are merged using a structural reparameterization method. Additionally, the ATSS algorithm is incorporated to enhance the accuracy of object detection. These approaches significantly reduce the model size, improve the inference speed, and lower the memory access cost. To validate the effectiveness of the model, it is trained and tested on a large dataset of diverse wheat spike images representing various phenotypes. The experimental results demonstrate that the RepYOLO algorithm achieves an average accuracy of 98.42% with a detection speed of 8.2 FPS. On the Jetson Nano platform, the inference speed reaches 34.20 ms. Consequently, the proposed model effectively reduces the memory access cost of deep learning networks without compromising accuracy and successfully improves the utilization of CPU/MCU limited performance.

Fish detection based on Gather-and-Distribute mechanism Multi-scale feature fusion network and Structural Re-parameterization method

ABSTRACT To solve the problems of localization and identification of fish in the complex fishway environment, improving the accuracy of fish detection, this paper proposes an object detection algorithm YOLORG, and a fishway fish detection dataset (FFDD). The FFDD contains 4,591 images from the web and lab shots and labeled with the LabelIMG tool, covering fish in a wide range of complex scenarios. The YOLORG algorithm, based on YOLOv8, improves the traditional FPN–PAN network into a C2f Multi-scale feature fusion network with a Gather-and-Distribute mechanism, which solves the problem of information loss accompanied by the network in the fusion of feature maps of different sizes. Also, we propose a C2D Structural Re-parameterization module with a rich gradient flow and good performance to further improve the detection accuracy of the algorithm. The experimental results show that the YOLORG algorithm improves the mAP50 and mAP50-95 by 1.2 and 1.8% compared to the original network under the joint VOC dataset, and also performs very well in terms of accuracy compared to other state-of-the-art object detection algorithms, and is able to detect fish in very turbid environments after training on the FFDD.

Ultra-Short-Term Power Prediction of Large Offshore Wind Farms Based on Spatiotemporal Adaptation of Wind Turbines

Accurately predicting the active power output of offshore wind power is of great significance for reducing the uncertainty in new power systems. By utilizing the spatiotemporal correlation characteristics among wind turbine unit outputs, this paper embeds the Diffusion Convolutional Neural Network (DCNN) into the Gated Recurrent Unit (GRU) for the feature extraction of spatiotemporal correlations in wind turbine unit outputs. It also combines graph structure learning to propose a sequence-to-sequence model for ultra-short-term power prediction in large offshore wind farms. Firstly, the electrical connection graph within the wind farm is used to preliminarily determine the reference adjacency matrix for the wind turbine units within the farm, injecting prior knowledge of the adjacency matrix into the model. Secondly, a convolutional neural network is utilized to convolve the historical curves of units within the farm along the time dimension, outputting a unit connection probability vector. The Gumbel–softmax reparameterization method is then used to make the probability vector differentiable, thereby generating an optimal adjacency matrix for the prediction task based on the probability vector. At the same time, the difference between the two adjacency matrices is added as a regularization term to the loss function to reduce model overfitting. The simulation of actual cases shows that the proposed model has good predictive performance in ultra-short-term power prediction for large offshore wind farms.

Damage identification for railway tracks using ultrasound guided wave and hybrid probabilistic deep learning

Rails are often damaged by train extrusion and other external forces, which seriously affects track safety. Although deep learning is widely used in rail damage recognition, its usual inability to understand own uncertainty that is critical in engineering application. Hence, this paper proposes a hybrid probabilistic deep learning for rail damage identification. It combines the hierarchical representation capabilities of deep learning with probability distributions. The posterior distribution provides support for the uncertainty quantification in identification results. Firstly, ultrasonic guided waves were used to sense the rail damage and time-frequency feature map dataset was obtained by analysing its time-domain signal using wavelet transform. Then, two different weight perturbation methods are used for comparative study. The reparameterization method performs more consistently and efficiently in this task, with a recognition accuracy of 0.9400. Subsequently, the effect of the number of probabilistic and non-probabilistic layers in the hybrid network on recognition results was analysed. Experimental results showed that hybrid probabilistic deep learning achieved the highest testing accuracy of 0.9900. The uncertainty quantification metrics of recognition results from hybrid probabilistic deep learning are mostly less than 0.2, demonstrating favourable reliability.

Height-diameter model of broad-leaved mixed forest based on species classification in Maoershan, Northeast China.

The complex stand structure and high species diversity of natural forests pose great challenges for analyzing stand growth and formulating reasonable plans for forest management. The height-diameter relationship is of great significance for predicting stand growth and formulating forest management measures. Based on survey data of 48 broad-leaved mixed forest plots in Maoershan, we classified 23 tree species into four groups based on species structure, growth characteristics and bionomics. We established a generalized model including stand, tree competition, species mixing and species diversity variables by reparameterization method, and a two-level mixed effect model of plot and tree species group. We tested the prediction ability of the model by leave-one-out cross-validation method. The results showed that the Ratkowsky (1990) model was the optimal basic model. The introduction of dominant height, basal area of trees larger than the object tree, basal area proportion of each species, and Shannon index could better explain the height-diameter relationship of broad-leaved mixed forest in Maoershan. The introduction of the mixed effect model of plot and tree species group could significantly improve the prediction accuracy of the model, with a Ra2 of 0.83. Under the same gradient of environmental factors, intolerant tree species exhibited higher tree heights than shade-tolerant tree species. In this study, we used the constructed tree height-diameter model to analyze the effects of species mixing and tree functional traits on tree height, which provided a theoretical basis for accurately predicting height of different tree species and analyzing the growth relationships in broadleaved mixed forests.

Rep-MCA-former: An efficient multi-scale convolution attention encoder for text-independent speaker verification

In many speaker verification tasks, the quality of speaker embedding is an important factor in affecting speaker verification systems. Advanced speaker embedding extraction networks aim to capture richer speaker features through the multi-branch network architecture. Recently, speaker verification systems based on transformer encoders have received much attention, and many satisfactory results have been achieved because transformer encoders can efficiently extract the global features of the speaker (e.g., MFA-Conformer). However, the large number of model parameters and computational latency are common problems faced by the above approaches, which make them difficult to apply to resource-constrained edge terminals. To address this issue, this paper proposes an effective, lightweight transformer model (MCA-former) with multi-scale convolutional self-attention (MCA), which can perform multi-scale modeling and channel modeling in the temporal direction of the input with low computational cost. In addition, in the inference phase of the model, we further develop a systematic re-parameterization method to convert the multi-branch network structure into the single-path topology, effectively improving the inference speed. We investigate the performance of the MCA-former for speaker verification under the VoxCeleb1 test set. The results show that the MCA-based transformer model is more advantageous in terms of the number of parameters and inference efficiency. By applying the re-parameterization, the inference speed of the model is increased by about 30%, and the memory consumption is significantly improved.

A YOLO-NL object detector for real-time detection

In recent years, YOLO object detection models have undergone significant advancement due to the success of novel deep convolutional networks. The success of these YOLO models is often attributed to their use of guidance techniques, such as expertly tailored deeper backbone and meticulously crafted detector head, which provides effective mechanisms to tradeoff between accuracy and efficiency. However, these sluggish-reasoning models are not capable of handling false detection and negative phenomena, facing challenges include improving the robustness of scaled objects detection against occlude and densely sophisticated scenarios.To address these limitations, we propose a novel object detector, You Only Look Once and None Left (YOLO-NL). Our model includes a novel global dynamic label assignment strategy, which allocates labels for specific anchors to maintain a balance between higher precision detection and finer localization. To enhance the detection capability of multi-scale objects in complex scenes, we separately upgrade CSPNet and PANet using the shortest-longest gradient strategy and self-attention mechanism. To meet the need for fast inference, we propose the Rep-CSPNet network using the reparameterization method to convert residual convolutions to ghost linear operations. Additionally, we accelerate the feature extraction process by deploying the serial SSPP structure. The proposed model is robust to scale objects against negative effectives such as dust, dense, ambiguous, and obstructed scenes. YOLO-NL achieved a mAP of 52.9% on the COCO 2017 test dataset, exhibiting a significant improvement of 2.64% compared to the baseline YOLOX. It is worth noting that YOLO-NL can perform high-accuracy and high-speed face mask detection in real-life scenarios. The YOLO-NL model was employed on self-built FMD and large open-source datasets, and the results show that it outperforms the other state-of-the-art methods, achieving 98.8% accuracy while maintaining 130 FPS.

Rational reparameterization of unstructured quadrilateral meshes for isogeometric analysis with optimal convergence

Unstructured quadrilateral meshes are widely used in representing surfaces of arbitrary topology. However, in connection with isogeometric analysis, unstructured meshes suffer from sub-optimal convergence rates in extraordinary regions, especially for solving higher-order PDEs (partial differential equations). In this work, a rational reparameterization method is proposed to achieve optimal convergence rates for isogeometric analysis using Catmull-Clark surfaces. The scheme lifts the convergence of analysis solutions by tuning the weights of control vertices of the input geometry in extraordinary regions, which are often neglected in constructing analysis-suitable parameterizations. Preferred tuning parameters with optimal convergence rates in L2-norm are determined statistically based on a test suite composed of typical testing solutions with a range of parameters. Numerical experiments confirm the validity of the proposed method for PDEs up to the fourth-order. Comparisons with the original Catmull-Clark scheme and a state-of-the-art tuned subdivision scheme also show the superiority of the proposed method in terms of both optimal convergence and reduced absolute errors of resulting analysis solutions.

High-performance one-stage detector for SiC crystal defects based on convolutional neural network

SiC (silicon carbide), as the most important third-generation semiconductor material, has huge market prospects in numerous fields, such as 5G base stations and new energy vehicle charging piles. The identification of SiC crystal defects is essential for improving crystal quality. Currently, this study relies mainly on artificial methods to identify defects, which have significant limitations in terms of accuracy and efficiency. Thus, to quickly detect and classify different SiC crystal defects in complex scenarios, a convolutional neural network-based SiC crystal defect detection (SCDD-Net) model is presented for the first time in this study. SCDD-Net uses an improved online convolutional re-parameterization method that can effectively extract the features of SiC crystal defects and decrease the large training overhead. We devised a new spatial pyramid pooling module that, when combined with the global context block, enables the fast fusion of high-level crystal defects and underlying features. We also designed an anchor-based decoupling detection head network to identify smaller crystal defects. By collecting and processing more than 5300 high-quality microscopic images, we built a fine-grained labeled SiC crystal defect image dataset, SiC-Crystal-5K, for the first time. The experimental results show that the SCDD-Net has excellent detection accuracy compared to other state-of-the-art models. The mean average precision for high-resolution SiC crystal defect identification reached 99.53%, corresponding to a single-image detection speed of 102 fps. In addition to crystal-defect detection, the SCDD-Net model can be used as a general-purpose detector in a wide range of scenarios.

TbsNet: the importance of thin-branch structures in CNNs.

The performance of a convolutional neural network (CNN) model is influenced by several factors, such as depth, width, network structure, size of the receptive field, and feature map scaling. The optimization of the best combination of these factors poses as the main difficulty in designing a viable architecture. This article presents an analysis of key factors influencing network performance, offers several strategies for constructing an efficient convolutional network, and introduces a novel architecture named TbsNet (thin-branch structure network). In order to minimize computation costs and feature redundancy, lightweight operators such as asymmetric convolution, pointwise convolution, depthwise convolution, and group convolution are implemented to further reduce the network's weight. Unlike previous studies, the TbsNet architecture design rejects the reparameterization method and adopts a plain, simplified structure which eliminates extraneous branches. We conduct extensive experiments, including network depth, width, etc. TbsNet performs well on benchmark platforms, Top 1 Accuracy on CIFAR-10 is 97.02%, on CIFAR-100 is 83.56%, and on ImageNet-1K is 86.17%. Tbs-UNet's DSC on the Synapse dataset is 78.39%, higher than TransUNet's 0.91%. TbsNet can be competent for some downstream tasks in computer vision, such as medical image segmentation, and thus is competitive with prior state-of-the-art deep networks such as ResNet, ResNeXt, RepVgg, ParNet, ConvNeXt, and MobileNet.

Resist: Robust Network Training for Memristive Crossbar-Based Neuromorphic Computing Systems

In recent years, memristive crossbar-based neuromorphic computing systems (NCS) have provided a promising solution for neural network acceleration. However, stuck-at faults (SAFs) in the memristor devices significantly degrade the computing accuracy of NCS. Besides, memristors suffer from process variations, causing the deviation of actual programmed conductance from its target value. In this paper, we propose a unified robust network training framework for a memristive crossbar-based NCS, simultaneously taking the impacts of SAFs and variations into account. In order to incorporate SAFs and variations into the training process, an effective sampling strategy for SAF and an efficient variation injection technique based on the local reparameterization method are developed. Experimental results clearly demonstrate that the proposed training framework can boost the computation accuracy of NCS and improve the NCS robustness.

A High-Precision Plant Disease Detection Method Based on a Dynamic Pruning Gate Friendly to Low-Computing Platforms.

Timely and accurate detection of plant diseases is a crucial research topic. A dynamic-pruning-based method for automatic detection of plant diseases in low-computing situations is proposed. The main contributions of this research work include the following: (1) the collection of datasets for four crops with a total of 12 diseases over a three-year history; (2) the proposition of a re-parameterization method to improve the boosting accuracy of convolutional neural networks; (3) the introduction of a dynamic pruning gate to dynamically control the network structure, enabling operation on hardware platforms with widely varying computational power; (4) the implementation of the theoretical model based on this paper and the development of the associated application. Experimental results demonstrate that the model can run on various computing platforms, including high-performance GPU platforms and low-power mobile terminal platforms, with an inference speed of 58 FPS, outperforming other mainstream models. In terms of model accuracy, subclasses with a low detection accuracy are enhanced through data augmentation and validated by ablation experiments. The model ultimately achieves an accuracy of 0.94.

Bayesian kinetic modeling for tracer-based metabolomic data

BackgroundStable Isotope Resolved Metabolomics (SIRM) is a new biological approach that uses stable isotope tracers such as uniformly 13C\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^{13}C$$\\end{document}-enriched glucose (13C6\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^{13}C_6$$\\end{document}-Glc) to trace metabolic pathways or networks at the atomic level in complex biological systems. Non-steady-state kinetic modeling based on SIRM data uses sets of simultaneous ordinary differential equations (ODEs) to quantitatively characterize the dynamic behavior of metabolic networks. It has been increasingly used to understand the regulation of normal metabolism and dysregulation in the development of diseases. However, fitting a kinetic model is challenging because there are usually multiple sets of parameter values that fit the data equally well, especially for large-scale kinetic models. In addition, there is a lack of statistically rigorous methods to compare kinetic model parameters between different experimental groups.ResultsWe propose a new Bayesian statistical framework to enhance parameter estimation and hypothesis testing for non-steady-state kinetic modeling of SIRM data. For estimating kinetic model parameters, we leverage the prior distribution not only to allow incorporation of experts’ knowledge but also to provide robust parameter estimation. We also introduce a shrinkage approach for borrowing information across the ensemble of metabolites to stably estimate the variance of an individual isotopomer. In addition, we use a component-wise adaptive Metropolis algorithm with delayed rejection to perform efficient Monte Carlo sampling of the posterior distribution over high-dimensional parameter space. For comparing kinetic model parameters between experimental groups, we propose a new reparameterization method that converts the complex hypothesis testing problem into a more tractable parameter estimation problem. We also propose an inference procedure based on credible interval and credible value. Our method is freely available for academic use at https://github.com/xuzhang0131/MCMCFlux.ConclusionsOur new Bayesian framework provides robust estimation of kinetic model parameters and enables rigorous comparison of model parameters between experimental groups. Simulation studies and application to a lung cancer study demonstrate that our framework performs well for non-steady-state kinetic modeling of SIRM data.

An Explainable Brain Tumor Detection Framework for MRI Analysis

Explainability in medical images analysis plays an important role in the accurate diagnosis and treatment of tumors, which can help medical professionals better understand the images analysis results based on deep models. This paper proposes an explainable brain tumor detection framework that can complete the tasks of segmentation, classification, and explainability. The re-parameterization method is applied to our classification network, and the effect of explainable heatmaps is improved by modifying the network architecture. Our classification model also has the advantage of post-hoc explainability. We used the BraTS-2018 dataset for training and verification. Experimental results show that our simplified framework has excellent performance and high calculation speed. The comparison of results by segmentation and explainable neural networks helps researchers better understand the process of the black box method, increase the trust of the deep model output, and make more accurate judgments in disease identification and diagnosis.

Skip-Concatenated Image Super-Resolution Network for Mobile Devices

Single-image super-resolution technology has been widely studied in various applications to improve the quality and resolution of degraded images acquired from noise-sensitive low-resolution sensors. As most studies on single-image super-resolution focused on the development of deep learning networks operating on high-performance GPUs, this study proposed an efficient and lightweight super-resolution network that enables real-time performance on mobile devices. To replace the relatively slow element-wise addition layer on mobile devices, we introduced a skip connection layer by directly concatenating a low-resolution input image with an intermediate feature map. In addition, we introduced weighted clipping to reduce the quantization errors commonly encountered during float-to-int8 model conversion. Moreover, a reparameterization method was selectively applied without increasing the cost in terms of inference time and number of parameters. Based on the contributions, the proposed network has been recognized as the best solution in Mobile AI & AIM 2022 Real-Time Single-Image Super-Resolution Challenge with PSNR of 30.03 dB and NPU runtime of 19.20 ms.