融合DenseNet201网络与Xception网络的外周血白细胞五分类方法研究

目的:为解决白细胞图像五分类中单一分类网络精度不高、泛化能力差的问题,提出一种融合DenseNet201网络与Xception网络的外周血白细胞分类方法。方法:对输入的白细胞图像分别通过DenseNet201网络与Xception网络的特征提取层进行特征提取,将提取到的特征进行串联式组合后再通过一个由全连接层、Dropout层、Softmax层构成的白细胞分类器实现白细胞五分类。为验证该方法的适用性和分类性能,分别在公开的单一来源白细胞数据集1和混合来源数据集2上,与基于经典卷积神经网络VGG16、ResNet50、InceptionV3、DenseNet201和Xception的分类方法进行对比实验。结果:在图像质量较好、颜色分布一致的数据集1和图像质量较差、颜色分布各异的数据集2上,融合DenseNet201网络与Xception网络的分类方法的平均分类准确率分别达到99.4%和98.2%,均优于基于经典卷积神经网络的分类方法。结论:提出的融合DenseNet201网络与Xception网络的外周血白细胞分类方法对数据集适用性较好、分类精度较高,可作为一种有效的外周血白细胞五分类方法。

基于改进DenseNet201网络的织物疵点检测算法

针对当前织物疵点检测存在大多数采用人工检测、速度慢、检测准确率低等问题,提出一种改进DenseNet201网络的织物检测算法。先对数据集图像进行预处理,可视化各种织物疵点类型的数量,把数据集划分为正常织物、8种常见织物疵点,对疵点图像进行数据增强,从而扩增训练集数量;然后提取在数据集ImageNet下预训练好的DenseNet201权重参数进行迁移学习,改进卷积层第1层、添加SPP层和本研究9分类的分类层;最后经过反复调整参数训练得到织物疵点检测模型。试验表明:改进后的DenseNet201模型对正常织物、8种常见织物疵点识别精度为96.8%。认为:改进DenseNet201网络模型具有良好的泛化性和鲁棒性。

基于深度学习的玉米病虫害智能诊断系统开发

使用自定义CNN和DenseNet201两种基于深度学习的网络,对大斑病、南方锈病、玉米黏虫、玉米蚜虫、玉米叶螨等10种常见玉米病虫害图像样本开展模型训练,并对部分训练结果进行了对比分析。发现所得val_accuracy大于0.8的模型中,基于CNN网络的模型相对稳定,val_loss值相对较小,说明在特定情况下基于CNN网络的模型收敛性相对较好,但DenseNet201网络更容易取得较高准确率的模型。面向Android系统开发基于深度学习的玉米病虫害智能诊断系统,并对系统开展诊断结果验证。验证结果:系统对于小斑病、纹枯病、茎腐病3种病害的诊断错误率较高,泛化能力不足。结论:开发基于深度学习的玉米病虫害智能诊断系统是可行的,但还需进一步调整完善。

Detection Algorithm of Knee Osteoarthritis Based on Magnetic Resonance Images

Knee osteoarthritis(OA)is a common disease that impairs knee function and causes pain.Currently,studies on the detection of knee OA mainly focus on X-ray images,but X-ray images are insensitive to the changes in knee OA in the early stage.Since magnetic resonance(MR)imaging can observe the early features of knee OA,the knee OA detection algorithm based on MR image is innovatively proposed to judge whether knee OA is suffered.Firstly,the knee MR images are preprocessed before training,including a region of interest clipping,slice selection,and data augmentation.Then the data set was divided by patient-level and the knee OA was classified by the deep transfer learning method based on the DenseNet201 model.The method divides the training process into two stages.The first stage freezes all the base layers and only trains the weights of the embedding neural networks.The second stage unfreezes part of the base layers and trains the unfrozen base layers and the weights of the embedding neural network.In this step,we design a block-by-block fine-tuning strategy for training based on the dense blocks,which improves detection accuracy.We have conducted training experiments with different depth modules,and the experimental results show that gradually adding more dense blocks in the fine-tuning can make the model obtain better detection performance than only training the embedded neural network layer.We achieve an accuracy of 0.921,a sensitivity of 0.960,a precision of 0.885,a specificity of 0.891,an F1-Score of 0.912,and an MCC of 0.836.The comparative experimental results on the OAI-ZIB dataset show that the proposed method outperforms the other detection methods with the accuracy of 92.1%.

Detection of COVID-19 and Pneumonia Using Deep Convolutional Neural Network

COVID-19 has created a panic all around the globe.It is a contagious dis-ease caused by Severe Acute Respiratory Syndrome Coronavirus 2(SARS-CoV-2),originated from Wuhan in December 2019 and spread quickly all over the world.The healthcare sector of the world is facing great challenges tackling COVID cases.One of the problems many have witnessed is the misdiagnosis of COVID-19 cases with that of healthy and pneumonia cases.In this article,we propose a deep Convo-lutional Neural Network(CNN)based approach to detect COVID+(i.e.,patients with COVID-19),pneumonia and normal cases,from the chest X-ray images.COVID-19 detection from chest X-ray is suitable considering all aspects in compar-ison to Reverse Transcription Polymerase Chain Reaction(RT-PCR)and Computed Tomography(CT)scan.Several deep CNN models including VGG16,InceptionV3,DenseNet121,DenseNet201 and InceptionResNetV2 have been adopted in this pro-posed work.They have been trained individually to make particular predictions.Empirical results demonstrate that DenseNet201 provides overall better performance with accuracy,recall,F1-score and precision of 94.75%,96%,95%and 95%respec-tively.After careful comparison with results available in the literature,we have found to develop models with a higher reliability.All the studies were carried out using a publicly available chest X-ray(CXR)image data-set.

基于卷积神经网络与ECOC-SVM的输电线路异物检测

输电线路悬挂异物会引发输电线路单相接地、相间短路等停电事故,因此本文提出一种基于卷积神经网络与ECOC-SVM的输电线路异物检测方法。首先,本文构建气球、风筝、塑料和鸟巢4种输电线路异物图像数据集;然后采用Otsu自适应阈值分割、形态学处理等方法提取感兴趣区域;再利用DenseNet201提取感兴趣区域的特征;最后对ECOC-SVM模型进行训练、测试与结果分析。所用方法在4类异物上的平均识别准确率可达93.3%,有助于提高输电线路运维的效率。