AFLNet: Adversarial focal loss network for RGB-D salient object detection

Because salient objects usually have fewer data in a scene, the problem of class imbalance is often encountered in salient object detection (SOD). In order to address this issue and achieve the consistent salient objects, we propose an adversarial focal loss network with improving generative adversarial networks for RGB-D SOD (called AFLNet), in which color and depth branches constitute the generator to achieve the saliency map, and adversarial branch with high-order potentials, instead of pixel-wise loss function, refines the output of the generator to obtain contextual information of objects. We infer the adversarial focal loss function to solve the problem of foreground–background class imbalance. To sufficiently fuse the high-level features of color and depth cues, an inception model is adopted in deep layers. We conduct a large number of experiments using our proposed model and its variants, and compare them with state-of-the-art methods. Quantitative and qualitative experimental results exhibit that our proposed approach can improve the accuracy of salient object detection and achieve the consistent objects.     

Self-entrainment of air on stepped spillways

The pseudo-bottom-inception point related to air entrainment is located further upstream on stepped spillways than on smooth spillways, for otherwise identical conditions. Its position is relevant concerning cavitation aspects, flow losses and flow depths. This Paper presents and discusses visual observations made with a high-speed camera and air concentration measurements in the vicinity of the pseudo-bottom air inception point on a stepped model spillway. Insight into the bottom aeration processes is provided, pointing at the effects of dynamic and turbulent air-phase surface troughs instantaneously protruding to the pseudo-bottom. In addition, the measured data were analyzed with regard to the extensions of these surface troughs. The trough bases were found to reach approximately 70–80% of the mixture flow depth upstream of the inception point, to 60–70% at the inception point and to 40–50% at the equilibrium flow region downstream of the inception point. The highly-turbulent character of developed flow is described and the general air transport process specified on the basis of air concentrations and related parameters.     

结合多层池化的卷积神经网络在表情识别中的应用

为了解决传统卷积神经网络(Convolutional Neural Network,CNN)对于人脸微表情识别的泛化能力差的问题,提出了一种改进的Inception结构与残差结构结合的卷积神经网络方法.首先在改进的Inception结构的基础上将输入特征直接映射到输出结果中构成残差结构,并针对表情局部特征复杂模糊等不足...     

基于小波分解和1D-GoogLeNet的心律失常检测

心电图(ECG)信号的准确分类对于心脏病的自动诊断非常重要。为了实现对心律失常的智能分类,该文提出一种基于小波分解和1D-GoogLeNet的精确分类方法。在该方法中,利用Db6小波对ECG信号进行8级分解,得到既含时域信息又有频域信息的多维数据。随后,分解的样本用作1D-GoogLeNet的输入训练该模型。在提出的1D-GoogLeNet模型中,借鉴Inception在图像特征提取中的优异性能,将2维卷积变换为1维卷积学习ECG的特征,并且简化各个Inception的结构,降低模型参数。该文提出的神经网络分类器能够有效缓解计算效率低、收敛困难和模型退化的问题。在实验中,选用MIT-BIH心律失常数据集测试所提模型的性能,对比了信号的不同分解分量组合作为输入的检测结果,当输入数据由{d2-d7}组合时,所提1D-GoogLeNet模型可以达到96.58%的平均准确率。此外,还对比了该模型与未经结构优化的简单1维GoogLeNet在数据集上的表现,前者在准确率上比后者提高了4.7%,训练效率提高了118%。     

VI-NET: A hybrid deep convolutional neural network using VGG and inception V3 model for copy-move forgery classification

Nowadays, various image editing tools are available that can be utilized for manipulating the original images; here copy-move forgery is most common forgery. In copy-move forgery, some part of the original image is copied and pasted into the same image at some other location. However, Artificial Intelligence (AI) based approaches can extract manipulated features easily. In this study, a deep learning-based method is proposed to classify the copy-move forged images. For classifying the forged images, a deep learning (DL) based hybrid model is presented named as VI-NET using fusion of two DL architectures, i.e., VGG16 and Inception V3. Further, output of two models is concatenated and connected with two additional convolutional layers. Cross-validation protocols, K10 (90 % training, 10 % testing), K5 (80 % training, 20 % testing), and K2 (50 % training, 50 % testing) are applied on the COMOFOD dataset. Moreover, the performance of VI-NET is compared with transfer learning and machine learning models using evaluation metrics such as accuracy, precision, recall, F1 score, etc. Proposed hybrid model performed better than other approaches with classification accuracy of 99 ± 0.2 % in comparison to accuracy of 95 ± 4 % (Inception V3), 93 ± 5 % (MobileNet), 59 ± 8 % (VGG16), 60 ± 1 % (Decision tree), 87 ± 1 % (KNN), 54 ± 1 % (Naïve Bayes) and 65 ± 1 % (random forest) under K10 protocol. Similarly, results are evaluated based on K2 and K5 validation protocols. It is experimentally observed that the proposed model performance is better than existing standard and customized deep learning architectures.     

Deep Compressive Sensing on ECG Signals with Modified Inception Block and LSTM

In practical electrocardiogram (ECG) monitoring, there are some challenges in reducing the data burden and energy costs. Therefore, compressed sensing (CS) which can conduct under-sampling and reconstruction at the same time is adopted in the ECG monitoring application. Recently, deep learning used in CS methods improves the reconstruction performance significantly and can removes of some of the constraints in traditional CS. In this paper, we propose a deep compressive-sensing scheme for ECG signals, based on modified-Inception block and long short-term memory (LSTM). The framework is comprised of four modules: preprocessing; compression; initial; and final reconstruction. We adaptively compressed the normalized ECG signals, sequentially using three convolutional layers, and reconstructed the signals with a modified Inception block and LSTM. We conducted our experiments on the MIT-BIH Arrhythmia Database and Non-Invasive Fetal ECG Arrhythmia Database to validate the robustness of our model, adopting Signal-to-Noise Ratio (SNR) and percentage Root-mean-square Difference (PRD) as the evaluation metrics. The PRD of our scheme was the lowest and the SNR was the highest at all of the sensing rates in our experiments on both of the databases, and when the sensing rate was higher than 0.5, the PRD was lower than 2%, showing significant improvement in reconstruction performance compared to the comparative methods. Our method also showed good recovering quality in the noisy data.     

基于双通道注意力机制的ResNet果实外观品质分类北大核心CSCD

为了实现对采摘后的果实进行快速、精确的外观品质分类,并配合分拣生产线完成果实大规模集中分拣,该研究提出了一种基于改进ResNet的果实分类方法。首先,将深度残差神经网络(deep residual neural network,ResNet)网络中的残差模块与双通道SE模块(dual channel squeeze-and-excitation block,DC-SE Block)结合,增强有效的通道特征并抑制低效或无效的通道特征,提高特征图的表达能力,从而提升识别精度;其次,在原始ResNet模型中加入Inception模块,将果实不同尺度的特征进行融合,增强对较小缺陷的识别能力;最后,对收集到的4类不同外观品质的果实图像进行数据增强并利用迁移学习的方法对模型进行初始化。以苹果为例进行的试验结果表明:经过数据集训练之后的改进模型,在测试集下的准确率达到99.7%,高于原模型的98.5%;精确率达到99.7%,高于原模型的98.3%;召回率达到99.7%,高于原模型的98.7%;在图形处理器(graphic processing unit,GPU)下的平均检测速度达到32.3帧/s,略低于原模型的35.7帧/s。与GoogleNet、MobileNet等几种目前先进的分类方法进行比较并对不同改进模型进行对比试验的结果表明,该方法具有良好的分类性能,对解决果实外观品质的精准分级问题具有重要参考价值。     

Simulation of the electric field in wire-plate type electrostatic precipitators

The paper presents a general Control Volume model for electric field simulation in wire-plate type electrostatic precipitators, along with a new injection law for charge density. The model is validated against empirical equations and experimental data in the literature when applied to the wire-plate and point-plate configurations. The voltage current characteristics and detailed distribution of field and charge density are characterized, particularly for the case of barbed wire electrode. The effects of geometric variations, such as the sharpness of the tip and the direction of the needles with respect to the plate, are investigated.     

基于改进Inception V4的面部表情识别算法的研究

面部表情识别是地铁、火车站、机场等复杂环境中安检监控的重要任务,通过识别监控图像中行人的面部表情可以筛选出可疑分子。针对因监控图像模糊和面部表情拍摄不全而引起的识别准确率低等问题,提出一种改进的InceptionV4面部表情识别算法,改进InceptionV4的网络结构,使其更好地适应面部表情识别任务。基于深度学习中的Tensorflow平台对面部表情类数据进行训练,在面部表情验证集上进行测试,在输入图像为299×299时,识别准确率高达97.9%,改进后的算法在保证识别精度的同时,降低表情在类内差距较大、图像模糊和面部表情拍摄不全情况下的误识率,提高系统鲁棒性。