A New Image Coding Algorithm Based on Self-Organizing Neural Network

The paper deals with a new VQ+DPCM+DCT algorithm based on Self-Organizing Feature Maps(SOFM) algorithm for image coding. In addition. a Frequency sensitive SOFM (FSOFM) has been also devel-oped. Simulation results show that a very good visual quality of the coded image at 0.252 bits/pixel is obtained.

Morphological self-organizing feature map neural network with applications to automatic target recognition

The rotation invariant feature of the target is obtained using the multi-direction feature extraction property of the steerable filter. Combining the morphological operation top-hat transform with the self-organizing feature map neural network, the adaptive topological region is selected. Using the erosion operation, the topological region shrinkage is achieved. The steerable filter based morphological self-organizing feature map neural network is applied to automatic target recognition of binary standard patterns and real world infrared sequence images. Compared with Hamming network and morphological shared-weight networks respectively, the higher recognition correct rate, robust adaptability, quick training, and better generalization of the proposed method are achieved.

Physics-informed neural networks for diffraction tomography

We propose a physics-informed neural network(PINN)as the forward model for tomographic reconstructions of biological samples.We demonstrate that by training this network with the Helmholtz equation as a physical loss,we can predict the scattered field accurately.It will be shown that a pretrained network can be fine-tuned for different samples and used for solving the scattering problem much faster than other numerical solutions.We evaluate our methodology with numerical and experimental results.Our PINNs can be generalized for any forward and inverse scattering problem.

Short‐term and long‐term memory self‐attention network for segmentation of tumours in 3D medical images

Tumour segmentation in medical images(especially 3D tumour segmentation)is highly challenging due to the possible similarity between tumours and adjacent tissues,occurrence of multiple tumours and variable tumour shapes and sizes.The popular deep learning‐based segmentation algorithms generally rely on the convolutional neural network(CNN)and Transformer.The former cannot extract the global image features effectively while the latter lacks the inductive bias and involves the complicated computation for 3D volume data.The existing hybrid CNN‐Transformer network can only provide the limited performance improvement or even poorer segmentation performance than the pure CNN.To address these issues,a short‐term and long‐term memory self‐attention network is proposed.Firstly,a distinctive self‐attention block uses the Transformer to explore the correlation among the region features at different levels extracted by the CNN.Then,the memory structure filters and combines the above information to exclude the similar regions and detect the multiple tumours.Finally,the multi‐layer reconstruction blocks will predict the tumour boundaries.Experimental results demonstrate that our method outperforms other methods in terms of subjective visual and quantitative evaluation.Compared with the most competitive method,the proposed method provides Dice(82.4%vs.76.6%)and Hausdorff distance 95%(HD95)(10.66 vs.11.54 mm)on the KiTS19 as well as Dice(80.2%vs.78.4%)and HD95(9.632 vs.12.17 mm)on the LiTS.

Spatial-Temporal Characterization of Atmospheric Aerosols via Airborne Spectral Imaging and Growing Hierarchical Self-Organizing Maps

Neural network analysis based on Growing Hierarchical Self-Organizing Map (GHSOM) is used to examine Spatial-Temporal characteristics in Aerosol Optical Depth (AOD), Ångström Exponent (ÅE) and Precipitation Rate (PR) over selected East African sites from 2000 to 2014. The selected sites of study are Nairobi (1°S, 36°E), Mbita (0°S, 34°E), Mau Forest (0.0° - 0.6°S;35.1°E - 35.7°E), Malindi (2°S, 40°E), Mount Kilimanjaro (3°S, 37°E) and Kampala (0°N, 32.1°E). GHSOM analysis reveals a marked spatial variability in AOD and ÅE that is associated to changing PR, urban heat islands, diffusion, direct emission, hygroscopic growth and their scavenging from the atmosphere specific to each site. Furthermore, spatial variability in AOD, ÅE and PR is distinct since each variable corresponds to a unique level of classification. On the other hand, GHSOM algorithm efficiently discriminated by means of clustering between AOD, ÅE and PR during Long and Short rain spells and dry spell over each variable emphasizing their temporal evolution. The utilization of GHSOM therefore confirms the fact that regional aerosol characteristics are highly variable be it spatially or temporally and as well modulated by PR received over each variable.

Tree species classification using deep learning and RGB optical images obtained by an unmanned aerial vehicle

The diversity of tree species and the complexity of land use in cities create challenging issues for tree species classification.The combination of deep learning methods and RGB optical images obtained by unmanned aerial vehicles(UAVs) provides a new research direction for urban tree species classification.We proposed an RGB optical image dataset with 10 urban tree species,termed TCC10,which is a benchmark for tree canopy classification(TCC).TCC10 dataset contains two types of data:tree canopy images with simple backgrounds and those with complex backgrounds.The objective was to examine the possibility of using deep learning methods(AlexNet,VGG-16,and ResNet-50) for individual tree species classification.The results of convolutional neural networks(CNNs) were compared with those of K-nearest neighbor(KNN) and BP neural network.Our results demonstrated:(1) ResNet-50 achieved an overall accuracy(OA) of 92.6% and a kappa coefficient of 0.91 for tree species classification on TCC10 and outperformed AlexNet and VGG-16.(2) The classification accuracy of KNN and BP neural network was less than70%,while the accuracy of CNNs was relatively higher.(3)The classification accuracy of tree canopy images with complex backgrounds was lower than that for images with simple backgrounds.For the deciduous tree species in TCC10,the classification accuracy of ResNet-50 was higher in summer than that in autumn.Therefore,the deep learning is effective for urban tree species classification using RGB optical images.

Review of Optical Character Recognition for Power System Image Based on Artificial Intelligence Algorithm

Optical Character Recognition(OCR)refers to a technology that uses image processing technology and character recognition algorithms to identify characters on an image.This paper is a deep study on the recognition effect of OCR based on Artificial Intelligence(AI)algorithms,in which the different AI algorithms for OCR analysis are classified and reviewed.Firstly,the mechanisms and characteristics of artificial neural network-based OCR are summarized.Secondly,this paper explores machine learning-based OCR,and draws the conclusion that the algorithms available for this form of OCR are still in their infancy,with low generalization and fixed recognition errors,albeit with better recognition effect and higher recognition accuracy.Finally,this paper explores several of the latest algorithms such as deep learning and pattern recognition algorithms.This paper concludes that OCR requires algorithms with higher recognition accuracy.

Quantitative phase imaging(QPI)through random diffusers using a diffractive optical network

Quantitative phase imaging(QPI)is a label-free computational imaging technique used in various fields,including biology and medical research.Modern QPI systems typically rely on digital processing using iterative algorithms for phase retrieval and image reconstruction.Here,we report a diffractive optical network trained to convert the phase information of input objects positioned behind random diffusers into intensity variations at the output plane,all-optically performing phase recovery and quantitative imaging of phase objects completely hidden by unknown,random phase diffusers.This QPI diffractive network is composed of successive diffractive layers,axially spanning in total~70λ,where is the illumination wavelength;unlike existing digital image reconstruction and phase retrieval methods,it forms an all-optical processor that does not require external power beyond the illumination beam to complete its QPI reconstruction at the speed of light propagation.This all-optical diffractive processor can provide a low-power,high frame rate and compact alternative for quantitative imaging of phase objects through random,unknown diffusers and can operate at different parts of the electromagnetic spectrum for various applications in biomedical imaging and sensing.The presented QPI diffractive designs can be integrated onto the active area of standard CCD/CMOS-based image sensors to convert an existing optical microscope into a diffractive QPI microscope,performing phase recovery and image reconstruction on a chip through light diffraction within passive structured layers.

GCN引导模型视点的光学遥感道路提取网络

在光学遥感图像中,道路易受遮挡物、铺装材料以及周围环境等多重因素的影响,导致其特征模糊不清。然而,现有道路提取方法即使增强其特征感知能力,仍在特征模糊区域存在大量误判。为解决上述问题,本文提出基于GCN引导模型视点的道路提取网络(RGGVNet)。RGGVNet采用编解码结构,并设计基于GCN的视点引导模块(GVPG)在编解码器的连接处反复引导模型视点,从而增强对特征模糊区域的关注。GVPG利用GCN信息传播过程具有平均特征权重的特性,将特征图中不同区域道路显著性水平作为拉普拉斯矩阵,参与到GCN信息传播从而实现引导模型视点。同时,提出密集引导视点策略(DGVS),采用密集连接的方式将编码器、GVPG和解码器相互连接,确保有效引导模型视点的同时缓解优化困难。在解码阶段设计多分辨率特征融合(MRFF)模块,最小化不同尺度道路特征在特征融合和上采样过程中的信息偏移和损失。在两个公开遥感道路数据集中,本文方法IoU分别达到65.84%和69.36%,F1-score分别达到79.40%和81.90%。从定量和定性两方面实验结果可以看出,本文所提方法性能优于其他主流方法。

乳腺钼靶肿块自监督预训练迁移检测方法研究

借助深度学习技术在乳腺钼靶领域辅助医生进行乳腺癌诊断在当下已经成为很多研究关注的热点,诊断技术主要包括良恶性分类、病灶区域检测以及病灶区域分割等。由于深度学习训练的模型性能很大程度上依赖于大量的带有标注的数据,而医学图像数据集往往存在数据量少、标注成本昂贵以及公开数据集标注质量差等现象,所以在医学图像领域应用深度学习技术具有重重困难。为使基于深度学习的乳腺钼靶计算机辅助诊断技术的开发不受限于大量有标注的数据,提出一种适用于钼靶自监督目标检测方法来完成乳腺钼靶肿块检测任务,利用大量来自肿瘤医院的数据预训练,并在公开数据集DDSM上进行微调与测试。实验结果表明,提出模型在乳腺钼靶肿块检测任务中表现优异,并且不依赖于位置标签,具有重要的研究价值与应用前景。

NHNet——新型层次化遥感图像语义分割网络

深度学习分割方法是遥感图像分割领域的热点之一,主流的深度学习方法有卷积神经网络、transformer神经网络及两者的结合。特征提取是图像分割的重要环节,除了用卷积等方式提取特征,最近的研究聚焦于一些新的特征提取范式,如图卷积、小波变换等。本文利用聚类算法的区域构建属性,将改进的聚类算法用于骨干特征提取模块,同时使用卷积和视觉transformer作为辅助模块,以获取更丰富的特征表述;在模块基础上,提出了一种新型层次化遥感图像语义分割网络(NHNet);评估了NHNet语义分割的性能,并在LoveDA遥感数据集上与其他方法进行比较。结果表明,基于多特征提取的NHNet获得了竞争性的性能表现,平均交并比为49.64%,F_(1)分数为65.7%。同时,消融实验证明辅助模块提高了聚类算法分割的精确性,给NHNet分别提升了1.03%和2.41%的平均交并比。

基于端到端深度神经网络和图搜索的OCT图像视网膜层边界分割方法

视网膜层边界的形态变化是眼部视网膜疾病出现的重要标志,光学相干断层扫描(optical coherence tomography,OCT)图像可以捕捉其细微变化,基于OCT图像的视网膜层边界分割能够辅助相关疾病的临床判断.在OCT图像中,由于视网膜层边界的形态变化多样,其中与边界相关的关键信息如上下文信息和显著性边界信息等对层边界的判断和分割至关重要.然而已有分割方法缺乏对以上信息的考虑,导致边界不完整和不连续.针对以上问题,提出一种“由粗到细”的基于端到端深度神经网络和图搜索(graph search,GS)的OCT图像视网膜层边界分割方法,避免了非端到端方法中普遍存在的“断层”现象.在粗分割阶段,提出一种端到端的深度神经网络—注意力全局残差网络(attention global residual network,AGR-Net),以更充分和有效的方式提取上述关键信息.具体地,首先设计一个全局特征模块(global feature module,GFM),通过从图像的4个方向扫描以捕获OCT图像的全局上下文信息;其次,进一步将通道注意力模块(channel attention module,CAM)与全局特征模块串行组合并嵌入到主干网络中,以实现视网膜层及其边界的上下文信息的显著性建模,有效解决OCT图像中由于视网膜层形变和信息提取不充分所导致的误分割问题.在细分割阶段,采用图搜索算法去除AGR-Net粗分割结果中的孤立区域或和孔洞等,保持边界的固定拓扑结构和连续平滑,以实现整体分割结果的进一步优化,为医学临床的诊断提供更完整的参考.最后,在两个公开数据集上从不同的角度对所提出的方法进行性能评估,并与最新方法进行比较.对比实验结果也表明所提方法在分割精度和稳定性方面均优于现有方法.

机器学习和深度学习在遥感影像分类中的对比研究

遥感影像分类是遥感技术应用的一个重要环节;机器学习和深度学习能够实现精确、自动化、迅速、可定义和规模化的遥感影像分类。本文选取机器学习算法支持向量机和深度学习算法卷积神经网络、深度置信网络、栈式自编码网络共计4种分类算法进行对比研究,并对支持向量机核函数的参数以及深度学习算法的神经元数量开展寻优以到达最高分类精度。实验结果表明,深度学习算法栈式自编码网络的总体分类精度最高,分类效果最好,在地物复杂多样地区开展遥感地物分类时具有较好的适用性和推广价值。

基于深度神经网络的数据中心光互连网络资源分配方法

在人工智能环境下为了提高数据中心光互联网络组件和软件的安全性,需要构建优化的资源分配模型,提出基于深度神经网络的数据中心光互连网络资源分配方法。采用用户关联和功率谱分配联合优化方法构建数据中心光互连网络资源调度模型,结合对网络资源粒度的服务请求QoS资源配置实现对不同种类资源的融合和聚类处理,提取数据中心光互连网络资源的空间、时间、频谱等多维网格抽象模型参数,通过深度神经网络学习方法实现对网络资源分配过程中的多种资源粒度融合和收敛性寻优控制,建立用户之间分配数据中心光互连网络资源的信道模型,通过传输链路均衡配置方案实现对网络资源的优化分配和均衡配置。仿真结果表明,本方法的资源分配传输比特率为18 bit/s,延时较小,资源分配阻塞率低,为0.05%,且资源持有度较高,可始终维持在100%,说明本方法具有对较强的资源均衡配置能力。

基于卷积神经网络的光学遥感影像道路提取方法研究进展

随着光学遥感影像空间分辨率的提升和获取渠道的丰富,利用光学遥感影像实现地物智能解译已成为高效的技术路径。由于卷积神经网络(convolutional neural networks,CNN)强大的特征提取能力以及道路信息在多个领域的应用需求,基于CNN的道路提取方法成为了当前的研究热点。鉴于此,本文根据近年来的相关研究文献,对基于CNN的道路提取方法从基于形状特征的改进、基于连通性的改进、基于多尺度特征的改进和基于提取策略的改进四个方面进行归纳总结,然后描述典型道路遮挡案例,并利用经典CNN从样本标签的局限性层面对当前的技术难点进行分析与验证,最后从多源数据协同、样本库建设、弱监督模型和域适应学习四个方面对遥感影像道路提取的发展趋势进行评估和展望。

适用于图像超分辨率的多路径融合增强网络

卷积神经网络(Convolutional Neural Network,CNN)在单幅图像的超分辨率重建方面表现出了非常强大的能力,相比传统方法有着明显的改进。然而,尽管这些方法非常成功,但是由于需要大量的计算资源,直接应用于一些边缘设备并不现实。为了解决该问题,设计了一种轻量级的图像超分辨率重建网络——多路径融合增强网络(Multi-path Fusion Enhancement Network,MFEN)。具体来说,提出了一个新颖的融合注意力增强模块(Fusion Attention Enhancement Block,FAEB)作为多路径融合增强网络的主要构建模块。融合注意力增强模块由一条主干分支和两条层级分支构成:主干分支由堆叠的增强像素注意力模块组成,负责对特征图实现深度特征学习;层级分支则负责提取并融合不同大小感受野的特征图,从而实现多尺度特征学习。层级分支的融合方式则是以相邻的增强像素注意力模块输出为分支输入,通过自适应注意力模块(Self-Adaptive Attention Module,SAAM)来动态地增强不同大小感受野特征的融合程度,进一步补全特征信息,从而实现更全面、更精准的特征学习。大量实验表明,该多路径融合增强网络在基准测试集上具有更高的准确性。

结合梯度指导和局部增强Transformer的图像去模糊网络

模糊图像不仅影响人类感知还会影响后续计算机视觉任务的性能,例如自动驾驶系统和户外监控系统中的视觉算法.针对以往基于深度学习的去模糊方法感受野较小,不能动态适应输入内容和重建图像细节信息困难等问题,提出了一种基于Transformer的图像去模糊网络.网络包含两个分支:图像内容分支和梯度分支,每条分支均以具有窗口机制的Transformer作为主干,通过梯度分支的信息指导图像去模糊重建,能够更好地恢复图像的边缘和纹理.同时,为了充分利用图像的内容信息和梯度信息,本文还设计了一个交互式融合模块来有效融合特征信息.此外,本文通过在Transformer块的自注意力机制和前馈网络中引入卷积来解决Transformer对局部信息建模不足的问题.在合成数据集和真实数据集上的大量实验结果表明,提出的算法能有效去除复杂模糊并且恢复清晰的细节,在定量指标和视觉效果上均优于目前的主流去模糊算法.

多任务衍射神经网络系统设计与实现

为探索利用衍射神经网络执行多任务图像分类识别的可行性,本文设计并搭建一种衍射神经网络系统。该系统采用空间光调制器(Spatial Light Modulator,SLM)做衍射神经网络的相位及振幅权重的调制以及网络层的光学全连接,并利用CMOS相机实现衍射神经网络中各衍射层输出的光电非线性激活与输出图像识别结果判别。设计的系统模型在MNIST和Fashion-MNIST图像分类识别中正确率达到94.1%和92.1%。最终通过搭建光路系统,光学实验正确率分别为91%和81.7%。所设计的衍射神经网络系统可满足多种图像分类识别应用需求,为衍射网络的设计与构建提供了新的思路。

基于场景上下文感知的光学遥感图像分类方法

光学遥感图像分类是对地观测领域的关键技术之一。近年来,研究人员提出利用深度神经网络对光学遥感图像进行分类,针对部分网络模型存在特征提取不充分的问题,本文提出了一种基于场景上下文感知和注意力增强的ScEfficientNet遥感图像分类方法。该方法设计了场景上下文信息感知模块(SCDM)建模目标及其周围邻域的空间关系,利用场景上下文特征增强原始特征表示,引入卷积块注意力模块(CBAM),根据通道和空间的重要性对特征图进行加权,并结合深度可分离卷积结构提取目标判别性信息,提出了ScMBConv卷积结构。在上述工作的基础上,利用基于场景上下文感知与注意力增强的ScEfficientNet网络模型进行遥感图像分类识别。实验结果表明,ScEfficientNet在AID数据集上实现了96.8%的分类准确率,较EfficientNet提升了3.3%,参数量为5.55 M,整体性能优于VGGNet19、GoogLeNet和ViT-B等图像分类算法,验证了ScEfficientNet网络模型的有效性。

整合卷积与高效自注意力机制的图像分类模型

对于传统的图像分类网络而言,卷积神经网络受限于较小且固定的感受野使其忽略了感受野之外的图像特征信息.基于Transformer模型灵活的多头自注意力机制使得其必须依赖于巨大的数据量以减少过拟合的风险,导致模型参数与计算复杂度过于庞大.针对上述问题本文提出了一种名为CSNet的多阶段图像分类模型.在模型浅层阶段利用大核卷积分解的思想扩大卷积层感受野以学习较大范围的特征信息.在深层阶段利用一种高效的自注意力机制,将卷积运算的特性加入自注意力机制中,有效减少了原始自注意力机制局部计算冗余和过分依赖数据的问题.CSNet在CIFAR-10和ImageNet-1K数据集上的分类准确率分别达到98.9%和82.6%,实验表明CSNet的模型性能优于ResNet和Vision Transformer.