基于改进Deeplab V3+网络的语义分割

深度学习的语义分割在计算机视觉领域中有非常广阔的发展前景,但许多分割效果较好网络模型占用内存大和处理单张图片耗时长.针对这个问题,把Deeplab V3+模型的骨干网(ResNet101)的瓶颈单元设计为1D非瓶颈单元,且对空洞空间金字塔池化模块(Atrous Spatial Pyramid Pooling, ASPP)的卷积层进行分解.该算法能大幅度降低Deeplab V3+网络的参数量,提高网络推理速度.基于PASCAL VOC 2012数据集进行对比实验,实验结果显示改进网络模型拥有更快的处理速度和更优的分割效果,且消耗更少的内存.     

Is attention all you need for intraday Forex trading?

The main objective of this paper is to analyse whether the Transformer neural network, which has become one of the most influential algorithms in Artificial Intelligence over the last few years, exhibits predictive capabilities for high-frequency Forex data. The prediction task is to classify short-term Forex movements for six currency pairs and five different time intervals from 60 to 720 min. We find that the Transformer exhibits high predictive power in the context of intraday Forex trading. This performance is slightly better than for the carefully selected benchmark – ResNet-LSTM, which currently is a state-of-the-art algorithm. Since intraday Forex trading based on deep learning models is largely unexplored, we offer insight on which currency pair and time interval are amenable to devising a profitable trading strategy. We also show that high predictive accuracy can be misleading in real world trading for short time intervals, as models trained on OHLC data tend to report the highest accuracy when the spread cost is the highest. This renders assessment based on typical machine learning metrics overly optimistic. Therefore, it is critical to backtest frequent intraday Forex trading strategies with realistic cost assumptions, which is rarely the case in empirical literature. Lastly, sensitivity analysis shows that the length of the time interval used for training does not play a critical role in the Transformer's predictive capabilities, whereas features derived from technical analysis are essential.     

交叉连接的少层残差卷积神经网络

最近的研究表明,卷积神经网络的性能可以通过采用跨层连接来提高,典型的残差网络(ResNet)便通过恒等映射方法取得了非常好的图像识别效果.但是通过理论分析,在残差模块中,跨层连接线的布局并没有达到最优设置,造成信息的冗余和层数的浪费,为了进一步提高卷积神经网络的性能,文章设计了两种新型的网络结构,分别命名为C-Fnet...     

基于侧链连接卷积神经网络的手掌静脉图像识别

针对手掌静脉图像数量少且质量参差不齐,进而导致掌脉识别系统的性能降低的现象,提出一种基于侧链连接卷积神经网络的手掌静脉图像识别方法。首先,在ResNet模型的基础上,用卷积层和池化层提取掌脉特征。然后,采用指数线性单元（ELU）激活函数、批归一化（BN）和Dropout技术来改进和优化模型,以缓解梯度消失、防止过拟合、加快收敛及增强模型泛化能力。最后,引入稠密连接网络（DenseNet）,使提取到的手掌静脉特征更具丰富性和有效性。在两个公开库和一个自建库上分别进行实验,结果表明所提方法在三个数据库上的识别率分别为99.98%、97.95%、97.96%。可见该方法能有效提高掌脉识别系统的性能,且更适用于掌脉识别的实际应用。     

深度残差神经网络高分辨率遥感图像建筑物分割

针对高分辨率遥感图像建筑物分割问题,提出一种Encoder-Decoder的深度学习框架,建立输入图像到分割结果之间的端对端的分割模型。其中Encoder以残差网络为基础,自动提取建筑物的特征;Decoder采用反卷积实现对特征图的上采样,从而完成对建筑物的分割;同时引入批量规范化处理,降低了神经网络权重训练过程中的梯度竞争,从而减小了神经网络的训练难度。实验表明:提出的建筑物分割算法能有效提取建筑物的块状特征和边缘信息,降低复杂道路等干扰的影响,提升建筑物的分割精准度,算法对邻近复杂道路的建筑物、规律性建筑物、单体复杂建筑物等3种典型建筑物的分割精度分别为：0.837、0.892和0.630;F值分别为：0.851、0.879和0.730。同时,多分辨率条件下的分割实验结果表明,该算法对于一定范围内的多分辨率遥感图像具有较好的泛化能力。     

基于深度学习的电网监控视频中工作人员检测与识别

针对电网监控视频场景多样,电网工作人员姿态变化严重影响工作人员识别精度的问题,提出了一种基于深度学习的电网监控视频中工作人员检测与识别算法.该算法使用Res Net50网络提取行人特征,Faster-Rcnn检测方法快速、精确地检测出电网中的工作人员,识别网络对检测出的工作人员进行身份确认,并使用各种组合损失来训练检测与识别网络.在电网监控视频数据集上的测试结果表明,所提出的方法具有更高的检测和识别精度,且对遮挡及低光照图片具有较好的鲁棒性.     

Brain Tumor Auto-Segmentation on Multimodal Imaging Modalities Using Deep Neural Network

Due to the difficulties of brain tumor segmentation, this paper proposes a strategy for extracting brain tumors from three-dimensional Magnetic Resonance Image (MRI) and Computed Tomography (CT) scans utilizing 3D U-Net Design and ResNet50, taken after by conventional classification strategies. In this inquire, the ResNet50 picked up accuracy with 98.96%, and the 3D U-Net scored 97.99% among the different methods of deep learning. It is to be mentioned that traditional Convolutional Neural Network (CNN) gives 97.90% accuracy on top of the 3D MRI. In expansion, the image fusion approach combines the multimodal images and makes a fused image to extricate more highlights from the medical images. Other than that, we have identified the loss function by utilizing several dice measurements approach and received Dice Result on top of a specific test case. The average mean score of dice coefficient and soft dice loss for three test cases was 0.0980. At the same time, for two test cases, the sensitivity and specification were recorded to be 0.0211 and 0.5867 using patch level predictions. On the other hand, a software integration pipeline was integrated to deploy the concentrated model into the webserver for accessing it from the software system using the Representational state transfer (REST) API. Eventually, the suggested models were validated through the Area Under the Curve–Receiver Characteristic Operator (AUC–ROC) curve and Confusion Matrix and compared with the existing research articles to understand the underlying problem. Through Comparative Analysis, we have extracted meaningful insights regarding brain tumour segmentation and figured out potential gaps. Nevertheless, the proposed model can be adjustable in daily life and the healthcare domain to identify the infected regions and cancer of the brain through various imaging modalities.     

Brain Tumor Identification Using Data Augmentation and Transfer Learning Approach

A brain tumor is a lethal neurological disease that affects the average performance of the brain and can be fatal. In India, around 15 million cases are diagnosed yearly. To mitigate the seriousness of the tumor it is essential to diagnose at the beginning. Notwithstanding, the manual evaluation process utilizing Magnetic Resonance Imaging (MRI) causes a few worries, remarkably inefficient and inaccurate brain tumor diagnoses. Similarly, the examination process of brain tumors is intricate as they display high unbalance in nature like shape, size, appearance, and location. Therefore, a precise and expeditious prognosis of brain tumors is essential for implementing the of an implicit treatment. Several computer models adapted to diagnose the tumor, but the accuracy of the model needs to be tested. Considering all the above mentioned things, this work aims to identify the best classification system by considering the prediction accuracy out of AlexNet, ResNet 50, and Inception V3. Data augmentation is performed on the database and fed into the three convolutions neural network (CNN) models. A comparison line is drawn between the three models based on accuracy and performance. An accuracy of 96.2% is obtained for AlexNet with augmentation and performed better than ResNet 50 and Inception V3 for the 120th epoch. With the suggested model with higher accuracy, it is highly reliable if brain tumors are diagnosed with available datasets.     

基于CycleGAN的灰度图像彩色化方法

当前主流的图片彩色化方法包括传统算法和深度学习方法.随着深度学习模型的发展,基于深度学习的灰度图像彩色化方法能带来更好的着色效果,但仍然存在细节损失和着色枯燥问题.针对上述问题,本文将CycleGAN模型应用在非单一类别的灰度图像彩色化上,使其在动物、植物、风景等图片上有逼真的着色效果.模型结构上对CycleGAN模型的激活函数加以改进,在生成器使用PReLU激活函数,使模型更易于训练.在判别器使用PatchGAN提高图片高分辨率上的颜色细节.通过ImageNet数据集5个热门类别图像的训练后,模型对动植物与风景图彩色化的效果十分逼真.在图像评估指标中,该模型在PSNR中比GAN高了0.603 dB约有2.1%的提升,在SSIM中明显高于其他模型,在效果上有5.1%的提升.从视觉感受来看,通过CycleGAN彩色化的图片饱和度更高,在视觉真实性上高于VGG和GAN等模型,解决了着色枯燥问题,而且更容易还原图片中的颜色细节,避免细节损失.     

基于残差网络和GRU的XSS攻击检测方法

传统的XSS攻击及其漏洞检测方法在面对多样化的攻击payload时其效果难以令人满意,需要大量人工参与,具有较大的主观性;而如CNN、RNN等深度学习方法只能单一地学习数据样本的空间特征或时序特征.提出一种基于残差网络和GRU的XSS攻击检测方法,在CNN基础上引入残差框架并与GRU相结合来学习数据的时空特征,且通过利...