深度迁移学习的相干斑噪声图像标注算法研究

由于相干斑噪声会导致图像特征提取困难,普通的图像处理算法无法对相干斑噪声图像进行有效分类标注。针对其图像特征设计了具有正则与拟合项的求解模型,并提出了深度迁移学习标注算法。在正则项中引入滤波算法和惩罚策略,用于过滤相干斑噪声;拟合项控制估计结果向真实结果的逼近。为满足深度学习网络处理的凸特性要求,对模型采取非凸优化。在深度学习过程中,将图像标注整体分为两个子任务,通过参数迁移进行并行处理。在各个子任务的最末层,分别设计相应的损失函数,对各个特征标签采取计分评价,改善网络学习的搜索能力和收敛性。通过和数据库的仿真,验证了深度迁移学习标注算法能够有效过滤图像中的相干斑噪声,获得更好的图像标注准确性和稳定性。

Research on Image Annotation Algorithm of Speckle Noise Based on Deep Migration Learning

Because the speckle noise may lead to the difficulty of image feature extraction, the common image processing algorithm can not effectively classify and label the speckle noise image. Therefore, a solution model with regular and fitting terms was designed for its image features, and a depth migration learning annotation algorithm was proposed. Filtering algorithm and penalty strategy were introduced into the regular term to filter the speckle noise. The fitting term controled the approximation of the estimated result to the real one. In addition, in order to meet the convexity requirements of deep learning network processing, nonconvex optimization was adopted for the model. In the process of deep learning, the whole image annotation was divided into two sub tasks, which were processed in parallel by parameter migration. At the end of each sub task, the corresponding loss function was designed. To improve the search ability and convergence of network learning, the feature tags were evaluated by scoring. Through the simulation experiment of corel5 k and iaprtc-12 database, it was verified that the algorithm can effectively filter the speckle noise in the image and obtain better accuracy and stability of image annotation.