A High Resolution Convolutional Neural Network with Squeeze and Excitation Module for Automatic Modulation Classification

Automatic modulation classification(AMC) technology is one of the cutting-edge technologies in cognitive radio communications. AMC based on deep learning has recently attracted much attention due to its superior performances in classification accuracy and robustness. In this paper, we propose a novel, high resolution and multi-scale feature fusion convolutional neural network model with a squeeze-excitation block, referred to as HRSENet,to classify different kinds of modulation signals.The proposed model establishes a parallel computing mechanism of multi-resolution feature maps through the multi-layer convolution operation, which effectively reduces the information loss caused by downsampling convolution. Moreover, through dense skipconnecting at the same resolution and up-sampling or down-sampling connection at different resolutions, the low resolution representation of the deep feature maps and the high resolution representation of the shallow feature maps are simultaneously extracted and fully integrated, which is benificial to mine signal multilevel features. Finally, the feature squeeze and excitation module embedded in the decoder is used to adjust the response weights between channels, further improving classification accuracy of proposed model.The proposed HRSENet significantly outperforms existing methods in terms of classification accuracy on the public dataset “Over the Air” in signal-to-noise(SNR) ranging from-2dB to 20dB. The classification accuracy in the proposed model achieves 85.36% and97.30% at 4dB and 10dB, respectively, with the improvement by 9.71% and 5.82% compared to LWNet.Furthermore, the model also has a moderate computation complexity compared with several state-of-the-art methods.

基于改进ResNet-50与迁移学习的苹果叶片病害的图像识别

为解决下述问题,在自然场景中获取的苹果叶病害图像大多包含复杂的背景,同时,由于拍摄要求不那么严格,疾病特征更有可能出现在图像中的任意位置,所有这些因素都会影响卷积神经网络的识别精度,且随着网络加深其性能提升困难。提出了一种基于改进ResNet-50的苹果叶片病害识别方法。首先引入SE(squeeze-and-excitation)注意力机制,改进残差模块,在一定程度上去除复杂背景等干扰信息,另一方面降低模型对特征定位的过度敏感度,使模型能够学习更重要的疾病特征,减少复杂背景等信息的干扰,然后加入Dropout层抑制过拟合改进模型结构,最后采用迁移学习的训练方式加快模型的收敛速度。在苹果叶病数据集上的实验结果表明,改进模型的准确率达到98.35%,较ResNet-50提高5%。与其他一些传统的卷积神经网络相比,该模型收敛速度更快,具有更高的识别精度,能够较好地识别苹果叶片病害,适用于自然场景下获取的苹果叶病图像等优点,具有较强的实用性。

A Comprehensive Evaluation of State-of-the-Art Deep Learning Models for Road Surface Type Classification

In recent years,as intelligent transportation systems(ITS)such as autonomous driving and advanced driver-assistance systems have become more popular,there has been a rise in the need for different sources of traffic situation data.The classification of the road surface type,also known as the RST,is among the most essential of these situational data and can be utilized across the entirety of the ITS domain.Recently,the benefits of deep learning(DL)approaches for sensor-based RST classification have been demonstrated by automatic feature extraction without manual methods.The ability to extract important features is vital in making RST classification more accurate.This work investigates the most recent advances in DL algorithms for sensor-based RST classification and explores appropriate feature extraction models.We used different convolutional neural networks to understand the functional architecture better;we constructed an enhanced DL model called SE-ResNet,which uses residual connections and squeeze-and-excitation mod-ules to improve the classification performance.Comparative experiments with a publicly available benchmark dataset,the passive vehicular sensors dataset,have shown that SE-ResNet outperforms other state-of-the-art models.The proposed model achieved the highest accuracy of 98.41%and the highest F1-score of 98.19%when classifying surfaces into segments of dirt,cobblestone,or asphalt roads.Moreover,the proposed model significantly outperforms DL networks(CNN,LSTM,and CNN-LSTM).The proposed RE-ResNet achieved the classification accuracies of asphalt roads at 98.98,cobblestone roads at 97.02,and dirt roads at 99.56%,respectively.

A Novel SE-CNN Attention Architecture for sEMG-Based Hand Gesture Recognition

In this article,to reduce the complexity and improve the generalization ability of current gesture recognition systems,we propose a novel SE-CNN attention architecture for sEMG-based hand gesture recognition.The proposed algorithm introduces a temporal squeeze-and-excite block into a simple CNN architecture and then utilizes it to recalibrate the weights of the feature outputs from the convolutional layer.By enhancing important features while suppressing useless ones,the model realizes gesture recognition efficiently.The last procedure of the proposed algorithm is utilizing a simple attention mechanism to enhance the learned representations of sEMG signals to performmulti-channel sEMG-based gesture recognition tasks.To evaluate the effectiveness and accuracy of the proposed algorithm,we conduct experiments involving multi-gesture datasets Ninapro DB4 and Ninapro DB5 for both inter-session validation and subject-wise cross-validation.After a series of comparisons with the previous models,the proposed algorithm effectively increases the robustness with improved gesture recognition performance and generalization ability.

基于轻量级残差网络的植物叶片病害识别

针对基于卷积神经网络的植物叶片病害识别方法存在网络参数众多、计算量大且复杂的问题,结合植物叶片病害特征,提出了一种基于轻量级残差网络(Scale-Down ResNet)的植物叶片病害识别方法。网络基于Residual Network(ResNet),通过缩减网络卷积核数目和轻量级残差模块(SD-BLOCK),在大幅减少网络参数、降低计算复杂度的同时保持了低识别错误率,然后加入Squeeze-and-Excitation模块进一步降低识别错误率。在PlantVillage数据集上的实验表明,在网络参数量8×10^(4),计算量MFLOPs为55的情况下,模型识别错误率为0.55%。当参数量达到2.8×10^(5),计算量MFLOPs为176时,模型识别错误率为0.32%,低于ResNet-18,并且参数量约为其1/39,计算量约为其1/10。相比MobileNet V3和ShuffleNet V2,所提网络模型更为轻量,识别错误率更低。同时网络在自建苹果叶片病害数据集上获得了1.52%的低识别错误率。

基于可能性聚类和卷积神经网络的道路交通标识识别算法

为解决图像采集中噪声和复杂背景对图片的影响以及深度神经网络的高耗时问题,基于可能性聚类算法与卷积神经网络,提出一种道路交通标识识别算法.该方法运用了图像分割技术,并结合卷积神经网络模型对道路交通标识进行更准确的识别.首先,通过色彩增强、图像分割、特征提取、数据增强和归一化等批量预处理操作,形成一个完整的数据集;然后,结合Squeeze-and-Excitation思想和残差网络结构,充分训练出MRESE(My Residual-Squeeze and Excitation)卷积神经网络模型;最后,将优化的网络模型用于道路交通标志的识别.实验结果表明,该方法使训练时间缩短了5%左右,识别精度可达99.02%.

基于注意力特征融合的SqueezeNet细粒度图像分类模型

针对现有细粒度图像分类算法普遍存在的模型结构复杂、参数多、分类准确率较低等问题,提出一种注意力特征融合的SqueezeNet细粒度图像分类模型.通过对现有细粒度图像分类算法和轻量级卷积神经网络的分析,首先使用3个典型的预训练轻量级卷积神经网络,对其微调后在公开的细粒度图像数据集上进行验证,经比较后选择了模型性能最佳的SqueezeNet作为图像的特征提取器;然后将两个具有注意力机制的卷积模块嵌入至SqueezeNet网络的每个Fire模块;接着提取出改进后的SqueezeNet的中间层特征进行双线性融合形成新的注意力特征图,与网络的全局特征再融合后分类;最后通过实验对比和可视化分析,网络嵌入Convolution Block Attention Module(CBAM)模块的分类准确率在鸟类、汽车、飞机数据集上依次提高了8.96%、4.89%和5.85%,嵌入Squeeze-and-Excitation(SE)模块的分类准确率依次提高了9.81%、4.52%和2.30%,且新模型在参数量、运行效率等方面比现有算法更具优势.

Research on behavior recognition algorithm based on SE-I3D-GRU network

In order to effectively solve the problems of low accuracy and large amount of calculation of current human behavior recognition,a behavior recognition algorithm based on squeeze-and-excitation network(SENet) combined with 3 D Inception network(I3 D) and gated recurrent unit(GRU) network is proposed.The algorithm first expands the Inception module to three-dimensional,and builds a network based on the three-dimensional module,and expands SENet to three-dimensional,making it an attention mechanism that can pay attention to the three-dimensional channel.Then SENet is introduced into the 13 D network,named SE-I3 D,and SENet is introduced into the CRU network,named SE-GRU.And,SE-13 D and SE-GRU are merged,named SE-13 D-GRU.Finally,the network uses Softmax to classify the results in the UCF-101 dataset.The experimental results show that the SE-I3 D-GRU network achieves a recognition rate of 93.2% on the UCF-101 dataset.

用于语音控制的低资源关键词检索系统

基于深度神经网络的低资源条件下关键词检索已经取得了很大的进展,但这些方法仍旧需要较多的参数才能保证模型的精度。为了进一步减少模型的参数量,本文将Squeeze-and-Excitation网络和深度可分离卷积应用在关键词检索任务中。首先利用Squeeze-and-Excitation网络对不同特征通道之间的相互依赖关系建模的能力进一步提升模型的精度,然后通过将标准卷积替换为深度可分离卷积来有效的减少模型所需要的参数。在谷歌语音命令数据集上的实验证明我们的模型可以在保证高精度的同时把参数量限制在一定的范围内。

基于改进型MobileNet网络的车型识别方法

车型识别是车辆结构化系统的重要方面。针对嵌入式环境下车型识别精度的要求,提出改进型MobileNet网络模型。利用Squeeze-and-Excitation模块对MobileNet网络结构进行改进,提高车型识别精度。实验表明,所提出的改进策略可以有效地提高车型识别的精度。

基于SE-CaraNet的全方位最大密度投影图像颅内动脉瘤自动检测方法

传统的单一方位最大密度投影(MIP)图像在检测颅内动脉瘤时容易忽略部分形态特征,造成漏检和误检。针对该问题,本文提出一种新的基于全方位MIP图像的颅内动脉瘤检测方法。首先,对三维磁共振血管造影(MRA)图像进行全方位最大密度投影,获得MIP图像;然后,利用匹配滤波对颅内动脉瘤区域进行预定位;最后,使用Squeeze and Excitation(SE)模块对CaraNet模型进行了改进,并用改进后的模型对全方位MIP图像中的预定位区域进行检测,确定是否患有颅内动脉瘤。本文收集了245例图像对所提方法进行了测试实验。实验结果表明本文所提方法的精确率和特异性分别可以达到93.75%和93.86%,显著提高了对MIP图像中颅内动脉瘤的检测性能。

Fine-grained detection of caged-hen head states using adaptive Brightness Adjustment in combination with Convolutional Neural Networks

Timely identification and tracking of abnormal hens in stacked cages are of great significance for precision treatment and the elimination of sick individuals.The head features of the caged-hens are used to overcome observation difficulties caused by the cage and feathers blocking,but it is still hard to identify similar head states.To solve this problem,the fine-grained detection of caged-hens head states was developed using adaptive Brightness Adjustment in combination with Convolutional Neural Networks(FBA-CNN).Grid Region-based CNN(R-CNN),a convolution neural network(CNN),was optimized with the Squeeze-and-Excitation(SE)and Depthwise Over-parameterized Convolutional(DO-Conv)to detect layer heads from cages and to accurately cut them as single-head images.The brightness of each single-head image was adjusted adaptively and classified through the deep convolution neural network based on SE-Resnet50.Finally,we returned to the original image to realize multi-target detection with coordinate mapping.The results showed that the AP@0.5 of layer head detection using the optimized Grid R-CNN was 0.947,the accuracy of classification with SE-Resnet50 was 0.749,the F1 score was 0.637,and the mAP@0.5 of FBA-CNN was 0.846.In summary,this automated method can accurately identify different layer head states in layer cages to provide a basis for follow-up studies of abnormal behavior including dyspnea and cachexia.

基于卷积神经网络和XGBoost的摔倒检测

提出了一种基于卷积神经网络和XGBoost的摔倒检测算法。采用基于squeeze-and-excitation模块的YOLO-v3算法对图片进行人体区域检测,在此基础上使用人体姿态估计网络获取人体关节点并提取出特征向量,再将其输入XGBoost进行训练,进而判断人体是否摔倒。实验结果表明,所提出的摔倒检测算法准确率较高,达到98.3%。

嵌入SENet的改进YOLOv4眼底图像微动脉瘤自动检测算法

微动脉瘤是糖尿病视网膜病变的初期症状,消除该病灶可在早期非常有效地预防糖尿病视网膜病变。但由于视网膜结构复杂,同时眼底图像的成像由于患者、环境、采集设备等因素的不同会存在不同的亮度和对比度,现有的微动脉瘤检测算法难以实现该病灶的精确检测和定位,为此本文提出嵌入SENet(squeeze-andexcitation networks)的改进YOLO(you only look once)v4自动检测算法。该算法在YOLOv4网络基础上,首先通过使用一种改进的快速模糊C均值聚类算法对目标样本进行先验框参数优化,以提高先验框与特征图的匹配度;然后,在主干网络嵌入SENet模块,通过强化关键信息,抑制背景信息,提高微动脉瘤的置信度;此外,还在网络颈部增加空间金字塔池化结构以增强主干网络输出特征的接受域,从而有助于分离出重要的上下文信息;最后,在Kaggle数据集上进行模型验证,并与其他方法进行对比。实验结果表明,与其他各种结构的YOLOv4网络模型相比,所提出的嵌入SENet的改进YOLOv4网络模型能显著提高检测结果(与原始YOLOv4相比Fscore提升了12.68%);与其他网络模型以及方法相比,所提出的嵌入SENet的改进YOLOv4网络模型的自动检测精度明显更优,且可实现精准定位。故本文所提出的嵌入SENet的改进YOLOv4算法性能较优,能准确、有效地检测并定位出眼底图像中的微动脉瘤。