【摘要】 针对具有复杂纹理特征的细胞显微图像分割问题,提出了融合BN(Batch Normalization)与全卷积深度学习网络的细胞图像分割算法。构建全卷积增强型U-Net网络来获取细胞图像的特征信息;在构建的网络中融合改进的BN算法缓解训练时网络中间层数据分布改变而降低网络泛化能力的问题,既固定了每层数据的分布,又避免破坏已学习的数据特征,标准化处理输入层及每一隐层的输入数据;为缓解网络过拟合,在网络中增加了Dropout操作,构建了错误率最小的网络循环次数;最后,采用自适应矩估计优化函数来优化模型参数,提高了网络训练的速度及收敛性能。实验结果表明,相对于经典的最大类间方差(Otsu)算法、脉冲耦合神经网络(PCNN)算法、空洞卷积算法、及U-Net网络分割算法,所提算法分割细胞显微图像的PRI提高0.04以上,VoI降低0.24以上,BDE减少0.2以上,GCE降低0.04以上,分割的图像边缘与细节的清晰度、精细度相对于其它算法均有较大程度改善。更多还原

【Abstract】 Aiming at the problem of cell micro image segmentation with complex texture features, a cell image segmentation algorithm is proposed,which combines Batch Normalization(BN) and fully convolutional deep learning network. The full convolution enhanced u-net network was constructed to obtain the characteristic information of cell image. In the constructed network, the improved BN algorithm was integrated to alleviate the problem of reducing the generalization ability of the network due to the change of the data distribution in the middle layer of the network during trainingm, which not only fixes the data distribution of each layer, but also avoids damaging the learned data characteristics, and standardizes the input data of the input layer and each hidden layer. In order to alleviate over fitting, the dropout operation was added in the network, and the number of network loops with the smallest error rate was constructed. Finally, the adaptive moment estimation optimization function was used to optimize the model parameters, which improved the speed and convergence performance of the network training. The experimental results show that compared with the classical Otsu algorithm, PCNN algorithm, hole convolution algorithm and u-net network segmentation algorithm, the PRI of the proposed algorithm is increased by more than 0. 04, voi is reduced by more than0. 24, BDE is reduced by more than 0. 2 and GCE is reduced by more than 0. 04, The definition and fineness of the segmented image edge and detail are greatly improved compared with other algorithms.更多还原