Sub-Regional Infrared-Visible Image Fusion Using Multi-Scale Transformation

Infrared-visible image fusion plays an important role in multi-source data fusion,which has the advantage of integrating useful information from multi-source sensors.However,there are still challenges in target enhancement and visual improvement.To deal with these problems,a sub-regional infrared-visible image fusion method(SRF)is proposed.First,morphology and threshold segmentation is applied to extract targets interested in infrared images.Second,the infrared back-ground is reconstructed based on extracted targets and the visible image.Finally,target and back-ground regions are fused using a multi-scale transform.Experimental results are obtained using public data for comparison and evaluation,which demonstrate that the proposed SRF has poten-tial benefits over other methods.

YOLO-MFD:Remote Sensing Image Object Detection with Multi-Scale Fusion Dynamic Head

Remote sensing imagery,due to its high altitude,presents inherent challenges characterized by multiple scales,limited target areas,and intricate backgrounds.These inherent traits often lead to increased miss and false detection rates when applying object recognition algorithms tailored for remote sensing imagery.Additionally,these complexities contribute to inaccuracies in target localization and hinder precise target categorization.This paper addresses these challenges by proposing a solution:The YOLO-MFD model(YOLO-MFD:Remote Sensing Image Object Detection withMulti-scale Fusion Dynamic Head).Before presenting our method,we delve into the prevalent issues faced in remote sensing imagery analysis.Specifically,we emphasize the struggles of existing object recognition algorithms in comprehensively capturing critical image features amidst varying scales and complex backgrounds.To resolve these issues,we introduce a novel approach.First,we propose the implementation of a lightweight multi-scale module called CEF.This module significantly improves the model’s ability to comprehensively capture important image features by merging multi-scale feature information.It effectively addresses the issues of missed detection and mistaken alarms that are common in remote sensing imagery.Second,an additional layer of small target detection heads is added,and a residual link is established with the higher-level feature extraction module in the backbone section.This allows the model to incorporate shallower information,significantly improving the accuracy of target localization in remotely sensed images.Finally,a dynamic head attentionmechanism is introduced.This allows themodel to exhibit greater flexibility and accuracy in recognizing shapes and targets of different sizes.Consequently,the precision of object detection is significantly improved.The trial results show that the YOLO-MFD model shows improvements of 6.3%,3.5%,and 2.5%over the original YOLOv8 model in Precision,map@0.5 and map@0.5:0.95,separately.These results illustrate the clear advantages of the method.

Transfer learning framework for multi-scale crack type classification with sparse microseismic networks

Rock fracture mechanisms can be inferred from moment tensors(MT)inverted from microseismic events.However,MT can only be inverted for events whose waveforms are acquired across a network of sensors.This is limiting for underground mines where the microseismic stations often lack azimuthal coverage.Thus,there is a need for a method to invert fracture mechanisms using waveforms acquired by a sparse microseismic network.Here,we present a novel,multi-scale framework to classify whether a rock crack contracts or dilates based on a single waveform.The framework consists of a deep learning model that is initially trained on 2400000+manually labelled field-scale seismic and microseismic waveforms acquired across 692 stations.Transfer learning is then applied to fine-tune the model on 300000+MT-labelled labscale acoustic emission waveforms from 39 individual experiments instrumented with different sensor layouts,loading,and rock types in training.The optimal model achieves over 86%F-score on unseen waveforms at both the lab-and field-scale.This model outperforms existing empirical methods in classification of rock fracture mechanisms monitored by a sparse microseismic network.This facilitates rapid assessment of,and early warning against,various rock engineering hazard such as induced earthquakes and rock bursts.

MSC-YOLO:Improved YOLOv7 Based onMulti-Scale Spatial Context for Small Object Detection in UAV-View

Accurately identifying small objects in high-resolution aerial images presents a complex and crucial task in thefield of small object detection on unmanned aerial vehicles(UAVs).This task is challenging due to variations inUAV flight altitude,differences in object scales,as well as factors like flight speed and motion blur.To enhancethe detection efficacy of small targets in drone aerial imagery,we propose an enhanced You Only Look Onceversion 7(YOLOv7)algorithm based on multi-scale spatial context.We build the MSC-YOLO model,whichincorporates an additional prediction head,denoted as P2,to improve adaptability for small objects.We replaceconventional downsampling with a Spatial-to-Depth Convolutional Combination(CSPDC)module to mitigatethe loss of intricate feature details related to small objects.Furthermore,we propose a Spatial Context Pyramidwith Multi-Scale Attention(SCPMA)module,which captures spatial and channel-dependent features of smalltargets acrossmultiple scales.This module enhances the perception of spatial contextual features and the utilizationof multiscale feature information.On the Visdrone2023 and UAVDT datasets,MSC-YOLO achieves remarkableresults,outperforming the baseline method YOLOv7 by 3.0%in terms ofmean average precision(mAP).The MSCYOLOalgorithm proposed in this paper has demonstrated satisfactory performance in detecting small targets inUAV aerial photography,providing strong support for practical applications.

Multi-Scale Mixed Attention Tea Shoot Instance Segmentation Model

Tea leaf picking is a crucial stage in tea production that directly influences the quality and value of the tea.Traditional tea-picking machines may compromise the quality of the tea leaves.High-quality teas are often handpicked and need more delicate operations in intelligent picking machines.Compared with traditional image processing techniques,deep learning models have stronger feature extraction capabilities,and better generalization and are more suitable for practical tea shoot harvesting.However,current research mostly focuses on shoot detection and cannot directly accomplish end-to-end shoot segmentation tasks.We propose a tea shoot instance segmentation model based on multi-scale mixed attention(Mask2FusionNet)using a dataset from the tea garden in Hangzhou.We further analyzed the characteristics of the tea shoot dataset,where the proportion of small to medium-sized targets is 89.9%.Our algorithm is compared with several mainstream object segmentation algorithms,and the results demonstrate that our model achieves an accuracy of 82%in recognizing the tea shoots,showing a better performance compared to other models.Through ablation experiments,we found that ResNet50,PointRend strategy,and the Feature Pyramid Network(FPN)architecture can improve performance by 1.6%,1.4%,and 2.4%,respectively.These experiments demonstrated that our proposed multi-scale and point selection strategy optimizes the feature extraction capability for overlapping small targets.The results indicate that the proposed Mask2FusionNet model can perform the shoot segmentation in unstructured environments,realizing the individual distinction of tea shoots,and complete extraction of the shoot edge contours with a segmentation accuracy of 82.0%.The research results can provide algorithmic support for the segmentation and intelligent harvesting of premium tea shoots at different scales.

Improved multi-scale inverse bottleneck residual network based on triplet parallel attention for apple leaf disease identification

Accurate diagnosis of apple leaf diseases is crucial for improving the quality of apple production and promoting the development of the apple industry. However, apple leaf diseases do not differ significantly from image texture and structural information. The difficulties in disease feature extraction in complex backgrounds slow the related research progress. To address the problems, this paper proposes an improved multi-scale inverse bottleneck residual network model based on a triplet parallel attention mechanism, which is built upon ResNet-50, while improving and combining the inception module and ResNext inverse bottleneck blocks, to recognize seven types of apple leaf(including six diseases of alternaria leaf spot, brown spot, grey spot, mosaic, rust, scab, and one healthy). First, the 3×3 convolutions in some of the residual modules are replaced by multi-scale residual convolutions, the convolution kernels of different sizes contained in each branch of the multi-scale convolution are applied to extract feature maps of different sizes, and the outputs of these branches are multi-scale fused by summing to enrich the output features of the images. Second, the global layer-wise dynamic coordinated inverse bottleneck structure is used to reduce the network feature loss. The inverse bottleneck structure makes the image information less lossy when transforming from different dimensional feature spaces. The fusion of multi-scale and layer-wise dynamic coordinated inverse bottlenecks makes the model effectively balances computational efficiency and feature representation capability, and more robust with a combination of horizontal and vertical features in the fine identification of apple leaf diseases. Finally, after each improved module, a triplet parallel attention module is integrated with cross-dimensional interactions among channels through rotations and residual transformations, which improves the parallel search efficiency of important features and the recognition rate of the network with relatively small computational costs while the dimensional dependencies are improved. To verify the validity of the model in this paper, we uniformly enhance apple leaf disease images screened from the public data sets of Plant Village, Baidu Flying Paddle, and the Internet. The final processed image count is 14,000. The ablation study, pre-processing comparison, and method comparison are conducted on the processed datasets. The experimental results demonstrate that the proposed method reaches 98.73% accuracy on the adopted datasets, which is 1.82% higher than the classical ResNet-50 model, and 0.29% better than the apple leaf disease datasets before preprocessing. It also achieves competitive results in apple leaf disease identification compared to some state-ofthe-art methods.

A multi-scale second-order autoregressive recursive filter approach for the sea ice concentration analysis

To effectively extract multi-scale information from observation data and improve computational efficiency,a multi-scale second-order autoregressive recursive filter(MSRF)method is designed.The second-order autoregressive filter used in this study has been attempted to replace the traditional first-order recursive filter used in spatial multi-scale recursive filter(SMRF)method.The experimental results indicate that the MSRF scheme successfully extracts various scale information resolved by observations.Moreover,compared with the SMRF scheme,the MSRF scheme improves computational accuracy and efficiency to some extent.The MSRF scheme can not only propagate to a longer distance without the attenuation of innovation,but also reduce the mean absolute deviation between the reconstructed sea ice concentration results and observations reduced by about 3.2%compared to the SMRF scheme.On the other hand,compared with traditional first-order recursive filters using in the SMRF scheme that multiple filters are executed,the MSRF scheme only needs to perform two filter processes in one iteration,greatly improving filtering efficiency.In the two-dimensional experiment of sea ice concentration,the calculation time of the MSRF scheme is only 1/7 of that of SMRF scheme.This means that the MSRF scheme can achieve better performance with less computational cost,which is of great significance for further application in real-time ocean or sea ice data assimilation systems in the future.

Multi-Scale Design and Optimization of Composite Material Structure for Heavy-Duty Truck Protection Device

In this paper,to present a lightweight-developed front underrun protection device(FUPD)for heavy-duty trucks,plain weave carbon fiber reinforced plastic(CFRP)is used instead of the original high-strength steel.First,the mechanical and structural properties of plain carbon fiber composite anti-collision beams are comparatively analyzed from a multi-scale perspective.For studying the design capability of carbon fiber composite materials,we investigate the effects of TC-33 carbon fiber diameter(D),fiber yarn width(W)and height(H),and fiber yarn density(N)on the front underrun protective beam of carbon fiber compositematerials.Based on the investigation,a material-structure matching strategy suitable for the front underrun protective beam of heavy-duty trucks is proposed.Next,the composite material structure is optimized by applying size optimization and stack sequence optimization methods to obtain the higher performance carbon fiber composite front underrun protection beam of commercial vehicles.The results show that the fiber yarn height(H)has the greatest influence on the protective beam,and theH1matching scheme for the front underrun protective beamwith a carbon fiber composite structure exhibits superior performance.The proposed method achieves a weight reduction of 55.21% while still meeting regulatory requirements,which demonstrates its remarkable weight reduction effect.

Disparity estimation for multi-scale multi-sensor fusion

The perception module of advanced driver assistance systems plays a vital role.Perception schemes often use a single sensor for data processing and environmental perception or adopt the information processing results of various sensors for the fusion of the detection layer.This paper proposes a multi-scale and multi-sensor data fusion strategy in the front end of perception and accomplishes a multi-sensor function disparity map generation scheme.A binocular stereo vision sensor composed of two cameras and a light deterction and ranging(LiDAR)sensor is used to jointly perceive the environment,and a multi-scale fusion scheme is employed to improve the accuracy of the disparity map.This solution not only has the advantages of dense perception of binocular stereo vision sensors but also considers the perception accuracy of LiDAR sensors.Experiments demonstrate that the multi-scale multi-sensor scheme proposed in this paper significantly improves disparity map estimation.

Nonlinear Flap-Wise Vibration Characteristics ofWind Turbine Blades Based onMulti-Scale AnalysisMethod

This work presents a novel approach to achieve nonlinear vibration response based on the Hamilton principle.We chose the 5-MW reference wind turbine which was established by the National Renewable Energy Laboratory(NREL),to research the effects of the nonlinear flap-wise vibration characteristics.The turbine wheel is simplified by treating the blade of a wind turbine as an Euler-Bernoulli beam,and the nonlinear flap-wise vibration characteristics of the wind turbine blades are discussed based on the simplification first.Then,the blade’s large-deflection flap-wise vibration governing equation is established by considering the nonlinear term involving the centrifugal force.Lastly,it is truncated by the Galerkin method and analyzed semi-analytically using the multi-scale analysis method,and numerical simulations are carried out to compare the simulation results of finite elements with the numerical simulation results using Campbell diagram analysis of blade vibration.The results indicated that the rotational speed of the impeller has a significant impact on blade vibration.When the wheel speed of 12.1 rpm and excitation amplitude of 1.23 the maximum displacement amplitude of the blade has increased from 0.72 to 3.16.From the amplitude-frequency curve,it can be seen that the multi-peak characteristic of blade amplitude frequency is under centrifugal nonlinearity.Closed phase trajectories in blade nonlinear vibration,exhibiting periodic motion characteristics,are found through phase diagrams and Poincare section diagrams.

Integrated multi-scale approach combining global homogenization and local refinement for multi-field analysis of high-temperature superconducting composite magnets

Second-generation high-temperature superconducting(HTS)conductors,specifically rare earth-barium-copper-oxide(REBCO)coated conductor(CC)tapes,are promising candidates for high-energy and high-field superconducting applications.With respect to epoxy-impregnated REBCO composite magnets that comprise multilayer components,the thermomechanical characteristics of each component differ considerably under extremely low temperatures and strong electromagnetic fields.Traditional numerical models include homogenized orthotropic models,which simplify overall field calculation but miss detailed multi-physics aspects,and full refinement(FR)ones that are thorough but computationally demanding.Herein,we propose an extended multi-scale approach for analyzing the multi-field characteristics of an epoxy-impregnated composite magnet assembled by HTS pancake coils.This approach combines a global homogenization(GH)scheme based on the homogenized electromagnetic T-A model,a method for solving Maxwell's equations for superconducting materials based on the current vector potential T and the magnetic field vector potential A,and a homogenized orthotropic thermoelastic model to assess the electromagnetic and thermoelastic properties at the macroscopic scale.We then identify“dangerous regions”at the macroscopic scale and obtain finer details using a local refinement(LR)scheme to capture the responses of each component material in the HTS composite tapes at the mesoscopic scale.The results of the present GH-LR multi-scale approach agree well with those of the FR scheme and the experimental data in the literature,indicating that the present approach is accurate and efficient.The proposed GH-LR multi-scale approach can serve as a valuable tool for evaluating the risk of failure in large-scale HTS composite magnets.

Spatio-temporal Evaluation of Multi-scale Cultivated Land System Resilience in Black Soil Region from 2000 to 2019:A Case Study of Liaoning Province,Northeast China

It is of great significance to systematically analyze the cultivated land system resilience(CLSR) for the black soil protection and national food security.The CLSR is impacted by planting structure adjustment and cultivated land quality decline,posing major hidden dangers to food security.It is urgent to evaluate the CLSR at multiple spatio-temporal scales.This study took Liaoning Province in the black soil region of Northeast China as an example.Based on the resilience theory,this study constructed the CLSR evaluation system from the input-feedback perspective at the provincial-scale and the city-scale,and used the rank-sum ratio comprehensive evaluation method(RSR) to analyze the key influencing factors of CLSR in Liaoning Province and its 14 cities from 2000 to 2019.The results showed that:1) the time series changes of CLSR at the provincial-scale and the city-scale in Liaoning Province were similar,both showing an increasing trend.2) The CLSR in Liaoning Province presented a spatial pattern of ‘high in the west and low in the east’ at the city-scale.3) There were seven and six main influencing factors of CLSR at the provincial-scale and the city-scale,respectively.In addition to the net income per capita of rural households,other influencing factors of CLSR were different at the provincial-scale and the city-scale.The feedback factors were dominant at the provincial-scale,and the input factors and feedback factors were dominant at the city-scale.The results could provide a reference for the utilization of black soil and draw on the experience of regional agricultural planning and adjustment.

Unconditionally stable Crank-Nicolson algorithm with enhanced absorption for rotationally symmetric multi-scale problems in anisotropic magnetized plasma

Large calculation error can be formed by directly employing the conventional Yee’s grid to curve surfaces.In order to alleviate such condition,unconditionally stable CrankNicolson Douglas-Gunn(CNDG)algorithm with is proposed for rotationally symmetric multi-scale problems in anisotropic magnetized plasma.Within the CNDG algorithm,an alternative scheme for the simulation of anisotropic plasma is proposed in body-of-revolution domains.Convolutional perfectly matched layer(CPML)formulation is proposed to efficiently solve the open region problems.Numerical example is carried out for the illustration of effectiveness including the efficiency,resources,and absorption.Through the results,it can be concluded that the proposed scheme shows considerable performance during the simulation.

Enhancing the Quality of Low-Light Printed Circuit Board Images through Hue, Saturation, and Value Channel Processing and Improved Multi-Scale Retinex

To address the issue of deteriorated PCB image quality in the quality inspection process due to insufficient or uneven lighting, we proposed an image enhancement fusion algorithm based on different color spaces. Firstly, an improved MSRCR method was employed for brightness enhancement of the original image. Next, the color space of the original image was transformed from RGB to HSV, followed by processing the S-channel image using bilateral filtering and contrast stretching algorithms. The V-channel image was subjected to brightness enhancement using adaptive Gamma and CLAHE algorithms. Subsequently, the processed image was transformed back to the RGB color space from HSV. Finally, the images processed by the two algorithms were fused to create a new RGB image, and color restoration was performed on the fused image. Comparative experiments with other methods indicated that the contrast of the image was optimized, texture features were more abundantly preserved, brightness levels were significantly improved, and color distortion was prevented effectively, thus enhancing the quality of low-lit PCB images.

A Lightweight Convolutional Neural Network with Hierarchical Multi-Scale Feature Fusion for Image Classification

Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.

GMTS: GNN-based multi-scale transformer siamese network for remote sensing building change detection

With the remarkable success of change detection(CD)in remote sensing images in the context of deep learning,many convolutional neural network(CNN)based methods have been proposed.In the current research,to obtain a better context modeling method for remote sensing images and to capture more spatiotemporal characteristics,several attention-based methods and transformer(TR)-based methods have been proposed.Recent research has also continued to innovate on TR-based methods,and many new methods have been proposed.Most of them require a huge number of calculation to achieve good results.Therefore,using the TR-based mehtod while maintaining the overhead low is a problem to be solved.Here,we propose a GNN-based multi-scale transformer siamese network for remote sensing image change detection(GMTS)that maintains a low network overhead while effectively modeling context in the spatiotemporal domain.We also design a novel hybrid backbone to extract features.Compared with the current CNN backbone,our backbone network has a lower overhead and achieves better results.Further,we use high/low frequency(HiLo)attention to extract more detailed local features and the multi-scale pooling pyramid transformer(MPPT)module to focus on more global features respectively.Finally,we leverage the context modeling capabilities of TR in the spatiotemporal domain to optimize the extracted features.We have a relatively low number of parameters compared to that required by current TR-based methods and achieve a good effect improvement,which provides a good balance between efficiency and performance.

An infrared and visible image fusion method based upon multi-scale and top-hat transforms

The high-frequency components in the traditional multi-scale transform method are approximately sparse, which can represent different information of the details. But in the low-frequency component, the coefficients around the zero value are very few, so we cannot sparsely represent low-frequency image information. The low-frequency component contains the main energy of the image and depicts the profile of the image. Direct fusion of the low-frequency component will not be conducive to obtain highly accurate fusion result. Therefore, this paper presents an infrared and visible image fusion method combining the multi-scale and top-hat transforms. On one hand, the new top-hat-transform can effectively extract the salient features of the low-frequency component. On the other hand, the multi-scale transform can extract highfrequency detailed information in multiple scales and from diverse directions. The combination of the two methods is conducive to the acquisition of more characteristics and more accurate fusion results. Among them, for the low-frequency component, a new type of top-hat transform is used to extract low-frequency features, and then different fusion rules are applied to fuse the low-frequency features and low-frequency background; for high-frequency components, the product of characteristics method is used to integrate the detailed information in high-frequency. Experimental results show that the proposed algorithm can obtain more detailed information and clearer infrared target fusion results than the traditional multiscale transform methods. Compared with the state-of-the-art fusion methods based on sparse representation, the proposed algorithm is simple and efficacious, and the time consumption is significantly reduced.

基于Transformer和自适应特征融合的矿井低照度图像亮度提升和细节增强方法

高质量矿井影像为矿山安全生产提供保障,也有利于提高后续图像分析技术的性能。矿井影像受低照度环境的影响,易出现亮度低,照度不均,颜色失真,细节信息丢失严重等问题。针对上述问题,提出一种基于Transformer和自适应特征融合的矿井低照度图像亮度提升和细节增强方法。基于生成对抗思想搭建生成对抗式主体模型框架,使用目标图像域而非单一参考图像驱动判别器监督生成器的训练,实现对低照度图像的充分增强;基于特征表示学习理论搭建特征编码器,将图像解耦为亮度分量和反射分量,避免图像增强过程中亮度与颜色特征相互影响从而导致颜色失真问题;设计CEM-Transformer Encoder通过捕获全局上下文关系和提取局部区域特征,能够充分提升整体图像亮度并消除局部区域照度不均;在反射分量增强过程中,使用结合CEM-Cross-Transformer Encoder的跳跃连接将低级特征与深层网络处特征进行自适应融合,能够有效避免细节特征丢失,并在编码网络中添加ECA-Net,提高浅层网络的特征提取效率。制作矿井低照度图像数据集为矿井低照度图像增强任务提供数据资源。试验显示,在矿井低照度图像数据集和公共数据集中,与5种先进的低照度图像增强算法相比,该算法增强图像的质量指标PSNR、SSIM、VIF平均提高了16.564%,10.998%,16.226%和14.438%,10.888%,14.948%,证明该算法能够有效提升整体图像亮度,消除照度不均,避免颜色失真和细节丢失,实现矿井低照度图像增强。

基于多层次特征融合的Transformer人脸识别方法

卷积神经网络中的卷积操作只能捕获局部信息,而Transformer能保留更多的空间信息且能建立图像的长距离连接.在视觉领域的应用中,Transformer缺乏灵活的图像尺寸及特征尺度适应能力,通过利用层级式网络增强不同尺度建模的灵活性,且引入多尺度特征融合模块丰富特征信息.本文提出了一种基于改进的Swin Transformer人脸模型——Swin Face模型.Swin Face以Swin Transformer为骨干网络,引入多层次特征融合模块,增强了模型对人脸的特征表达能力,并使用联合损失函数优化策略设计人脸识别分类器,实现人脸识别.实验结果表明,与多种人脸识别方法相比,Swin Face模型通过使用分级特征融合网络,在LFW、CALFW、AgeDB-30、CFP数据集上均取得最优的效果,验证了此模型具有良好的泛化性和鲁棒性.

基于Depth-wise卷积和视觉Transformer的图像分类模型

图像分类作为一种常见的视觉识别任务,有着广阔的应用场景。在处理图像分类问题时,传统的方法通常使用卷积神经网络,然而,卷积网络的感受野有限,难以建模图像的全局关系表示,导致分类精度低,难以处理复杂多样的图像数据。为了对全局关系进行建模,一些研究者将Transformer应用于图像分类任务,但为了满足Transformer的序列化和并行化要求,需要将图像分割成大小相等、互不重叠的图像块,破坏了相邻图像数据块之间的局部信息。此外,由于Transformer具有较少的先验知识,模型往往需要在大规模数据集上进行预训练,因此计算复杂度较高。为了同时建模图像相邻块之间的局部信息并充分利用图像的全局信息,提出了一种基于Depth-wise卷积的视觉Transformer(Efficient Pyramid Vision Transformer,EPVT)模型。EPVT模型可以实现以较低的计算成本提取相邻图像块之间的局部和全局信息。EPVT模型主要包含3个关键组件:局部感知模块(Local Perceptron Module,LPM)、空间信息融合模块(Spatial Information Fusion,SIF)和“+卷积前馈神经网络(Convolution Feed-forward Network,CFFN)。LPM模块用于捕获图像的局部相关性;SIF模块用于融合相邻图像块之间的局部信息,并利用不同图像块之间的远距离依赖关系,提升模型的特征表达能力,使模型学习到输出特征在不同维度下的语义信息;CFFN模块用于编码位置信息和重塑张量。在图像分类数据集ImageNet-1K上,所提模型优于现有的同等规模的视觉Transformer分类模型,取得了82.6%的分类准确度,证明了该模型在大规模数据集上具有竞争力。