In this work,we propose a new,fully automated system for multiclass skin lesion localization and classification using deep learning.The main challenge is to address the problem of imbalanced data classes,found in HAM1...In this work,we propose a new,fully automated system for multiclass skin lesion localization and classification using deep learning.The main challenge is to address the problem of imbalanced data classes,found in HAM10000,ISBI2018,and ISBI2019 datasets.Initially,we consider a pretrained deep neural network model,DarkeNet19,and fine-tune the parameters of third convolutional layer to generate the image gradients.All the visualized images are fused using a High-Frequency approach along with Multilayered Feed-Forward Neural Network(HFaFFNN).The resultant image is further enhanced by employing a log-opening based activation function to generate a localized binary image.Later,two pre-trained deep models,Darknet-53 and NasNet-mobile,are employed and fine-tuned according to the selected datasets.The concept of transfer learning is later explored to train both models,where the input feed is the generated localized lesion images.In the subsequent step,the extracted features are fused using parallel max entropy correlation(PMEC)technique.To avoid the problem of overfitting and to select the most discriminant feature information,we implement a hybrid optimization algorithm called entropy-kurtosis controlled whale optimization(EKWO)algorithm.The selected features are finally passed to the softmax classifier for the final classification.Three datasets are used for the experimental process,such as HAM10000,ISBI2018,and ISBI2019 to achieve an accuracy of 95.8%,97.1%,and 85.35%,respectively.展开更多
针对现有的基于卷积神经网络的图像去模糊算法存在图像纹理细节恢复不清晰的问题,提出了一种基于多局部残差连接注意网络的图像去模糊算法。首先,采用一个卷积层进行浅层特征提取;其次,设计了一种新的基于残差连接和并行注意机制的多局...针对现有的基于卷积神经网络的图像去模糊算法存在图像纹理细节恢复不清晰的问题,提出了一种基于多局部残差连接注意网络的图像去模糊算法。首先,采用一个卷积层进行浅层特征提取;其次,设计了一种新的基于残差连接和并行注意机制的多局部残差连接注意模块,用于消除图像模糊并提取上下文信息;再次,采用一个基于扩张卷积的成对连接模块进行细节恢复;最后,利用一个卷积层重建清晰图像。实验结果表明:在GoPro数据集上的PSNR(peak signal to noise ratio)和SSIM(structure similarity)分别为31.83 dB、0.9275,在定性和定量两方面都表明所提方法能够有效地恢复模糊图像的纹理细节,网络性能优于对比方法。展开更多
针对发动机转子表面存在磕划伤和凸起等弱对比度微小缺陷难以检测的问题,提出一种利用多方向照明结合卷积神经网络模型的发动机转子表面缺陷检测方法。首先,采用光度立体法获得增强图形凹凸性特征的曲率图和高度图,作为输入图像;其次,...针对发动机转子表面存在磕划伤和凸起等弱对比度微小缺陷难以检测的问题,提出一种利用多方向照明结合卷积神经网络模型的发动机转子表面缺陷检测方法。首先,采用光度立体法获得增强图形凹凸性特征的曲率图和高度图,作为输入图像;其次,提出一种优化的更丰富的卷积特征网络(Richer Convolutional Features Network)模型,充分利用跳层连接将首阶段与后续阶段的侧输出特征融合,提高网络深层对精细尺度下信息的保留能力;通过通道及空间注意力机制对模型侧输出进行强化,增强有效特征并抑制干扰;优化损失函数,使数据集中无缺陷信息的图像样本也能够适用于网络模型的训练;最后,以人工标注的方式制作数据集并验证优化模型的有效性。试验结果表明,与经典的缺陷检测方法相比,全卷积网络对部分缺陷的区分能力较差,本文方法对转子的表面缺陷区域具有更好的检测效果,改进模型的像素准确率达94.31%,比RCF提高了0.87个百分点。展开更多
基金supported by Korea Institute for Advancement of Technology(KIAT)grant funded by the Korea Government(MOTIE)(P0012724,The Competency Development Program for Industry Specialist)the Soonchunhyang University Research Fund.
文摘In this work,we propose a new,fully automated system for multiclass skin lesion localization and classification using deep learning.The main challenge is to address the problem of imbalanced data classes,found in HAM10000,ISBI2018,and ISBI2019 datasets.Initially,we consider a pretrained deep neural network model,DarkeNet19,and fine-tune the parameters of third convolutional layer to generate the image gradients.All the visualized images are fused using a High-Frequency approach along with Multilayered Feed-Forward Neural Network(HFaFFNN).The resultant image is further enhanced by employing a log-opening based activation function to generate a localized binary image.Later,two pre-trained deep models,Darknet-53 and NasNet-mobile,are employed and fine-tuned according to the selected datasets.The concept of transfer learning is later explored to train both models,where the input feed is the generated localized lesion images.In the subsequent step,the extracted features are fused using parallel max entropy correlation(PMEC)technique.To avoid the problem of overfitting and to select the most discriminant feature information,we implement a hybrid optimization algorithm called entropy-kurtosis controlled whale optimization(EKWO)algorithm.The selected features are finally passed to the softmax classifier for the final classification.Three datasets are used for the experimental process,such as HAM10000,ISBI2018,and ISBI2019 to achieve an accuracy of 95.8%,97.1%,and 85.35%,respectively.
文摘针对现有的基于卷积神经网络的图像去模糊算法存在图像纹理细节恢复不清晰的问题,提出了一种基于多局部残差连接注意网络的图像去模糊算法。首先,采用一个卷积层进行浅层特征提取;其次,设计了一种新的基于残差连接和并行注意机制的多局部残差连接注意模块,用于消除图像模糊并提取上下文信息;再次,采用一个基于扩张卷积的成对连接模块进行细节恢复;最后,利用一个卷积层重建清晰图像。实验结果表明:在GoPro数据集上的PSNR(peak signal to noise ratio)和SSIM(structure similarity)分别为31.83 dB、0.9275,在定性和定量两方面都表明所提方法能够有效地恢复模糊图像的纹理细节,网络性能优于对比方法。
文摘针对发动机转子表面存在磕划伤和凸起等弱对比度微小缺陷难以检测的问题,提出一种利用多方向照明结合卷积神经网络模型的发动机转子表面缺陷检测方法。首先,采用光度立体法获得增强图形凹凸性特征的曲率图和高度图,作为输入图像;其次,提出一种优化的更丰富的卷积特征网络(Richer Convolutional Features Network)模型,充分利用跳层连接将首阶段与后续阶段的侧输出特征融合,提高网络深层对精细尺度下信息的保留能力;通过通道及空间注意力机制对模型侧输出进行强化,增强有效特征并抑制干扰;优化损失函数,使数据集中无缺陷信息的图像样本也能够适用于网络模型的训练;最后,以人工标注的方式制作数据集并验证优化模型的有效性。试验结果表明,与经典的缺陷检测方法相比,全卷积网络对部分缺陷的区分能力较差,本文方法对转子的表面缺陷区域具有更好的检测效果,改进模型的像素准确率达94.31%,比RCF提高了0.87个百分点。
