基于模糊数学的对抗网络图像配准建模

为了解决传统方法特征提取结果受外界环境影响大,且没有考虑对抗网络图像中高频信息的特殊作用,影响配准精度的问题,通过模糊数学方法分析对抗网络图像配准建模问题。分析了对抗网络,在生成模型与判别模型中添加条件变量,通过对抗网络,利用表征向量对图像进行重构,生成图像数据。通过变换对图像对比度进行扩展,通过反变换获取原空间域中的边缘增强图像,通过抑制干扰能力强的Susan算子提取对抗网络图像边缘特征。在边缘特征提取的基础上,引入模糊数学中的模糊隶属度,对图像中不同点属于梯度的模糊隶属度进行定义,构造图像的模糊梯度场,通过模糊数学中的贴进度构造模糊梯度相似性测度,将模糊梯度相似性高的图像作为配准图像,实现对抗网络图像配准。结果表明:研究方法配准效果好;针对存在平移、灰度变化、细节变化、区域变化和尺度差异情况下的图像,可保持很高的性能。研究结果应用性强,配准准确性好。

Modeling and analysis of confrontation network image registration based on Fuzzy Mathematics

In order to solve the problem that the traditional feature extraction results are greatly influenced by the external environment, and the special role of high frequency information in countermeasure network images is not considered, which affects the registration accuracy. The modeling of confrontation network image registration is analyzed by fuzzy mathematics. The antagonistic network is analyzed. Conditional variables are added to the generation model and discrimination model. Through the antagonistic network, the image is reconstructed using the representation vector to generate image data. The contrast of the image is expanded by transformation, and the edge enhancement image in the original spatial domain is obtained by inverse transformation. The edge features of the image against the network are extracted by SUSAN operator with strong interference suppression ability. On the basis of edge feature extraction, the fuzzy membership degree in fuzzy mathematics is introduced to define the fuzzy membership degree of gradient at different points in the image, and the fuzzy gradient field of the image is constructed. The fuzzy gradient similarity measure is constructed by the sticking schedule in fuzzy mathematics, and the image with high similarity of the fuzzy gradient is used as the registration map. For example, the network image registration is realized. The results show that the proposed method has good registration effect, and can maintain high performance for images with translation, gray level change, detail change, regional change and scale difference. It is obvious that the proposed method has strong applicability and good registration accuracy.