基于焦点分割与改进SuperPoint的多聚焦图像配准算法研究

栾庆磊; 屈紫浩; 郭继智

文章摘要

栾庆磊,屈紫浩,郭继智.基于焦点分割与改进SuperPoint的多聚焦图像配准算法研究[J].安徽建筑大学学报,2025,33(1):51-61

基于焦点分割与改进SuperPoint的多聚焦图像配准算法研究

Research on Multi-focus Image Registration Algorithm Based on Focus Segmentation with Improved SuperPoint

DOI：

中文关键词: 多聚焦显微图像焦点分割 SuperPoint GhostNetV2

英文关键词: multi-focused microscopic image focus segmentation SuperPoint GhostNetV2

基金项目:安徽省科技重大专项项目（202203a05020022）

作者	单位
栾庆磊	安徽建筑大学机械与电气工程学院，安徽合肥 230601；安徽省工程机械智能制造重点实验室，安徽合肥 230601
屈紫浩	安徽建筑大学机械与电气工程学院，安徽合肥 230601；安徽省工程机械智能制造重点实验室，安徽合肥 230601
郭继智	安徽建筑大学机械与电气工程学院，安徽合肥 230601；安徽省工程机械智能制造重点实验室，安徽合肥 230601

摘要点击次数: 4923

全文下载次数: 0

中文摘要:

现有多聚焦显微图像配准存在特征提取难度大、易受噪声干扰、时间复杂度高等问题。针对以上问题，提出一种基于自适应Canny聚焦区域分割算法与轻量化SuperPoint网络相融合的多聚焦显微图像配准算法。首先，使用自适应中值滤波与大津算法改进的Canny算法分割出相邻多聚焦显微图像的聚焦区域，在去除相邻帧共同背景区域的同时减小图像尺寸，以提高后续特征检测的精度和速度；接着，针对特征提取难度大、时间复杂度高的问题，使用轻量化的SuperPoint网络提取特征点，针对原始SuperPoint网络的VGG架构编码层参数量多、计算量大的缺点，使用GhostNetV2代替原本的VGG编码层，在保证精度的同时降低了计算量和参数量；然后，使用K最近邻算法对特征点进行匹配。最后，使用退化采样一致性算法（DEGENSAC）代替普通的随机采样一致性算法（RANSAC）去除误匹配并计算单应矩阵对多聚焦显微图像进行配准。经过实验验证，所提算法相比于其他算法拥有更高匹配精度和速度，相比于原始SuperPoint，参数量、计算量和模型大小分别下降了51.69%、88.04%、50.07%，FPS增加了3倍左右。

英文摘要:

Considering the feature extraction, noise susceptibility, and high time complexity in the multi-focused microscopic imageregistration, a multi-focus microscope image registration algorithm based on the fusion of adaptive Canny focus area segmentationalgorithm and lightweight SuperPoint network is proposed. Firstly, the Canny algorithm is improved by adaptive median filtering andOtsu algorithm to segment the focusing region of adjacent multi-focused microscopic image, which reduces the image size while removing thecommon background region of adjacent frames to improve the accuracy and speed of subsequent feature detection. Next, a lightweight SuperPointnetwork is used to address the difficulty and high time complexity of feature extraction, and GhostNetV2 instead of the originalVGG coding layer is used to deal with a large number of parameters in the coding layer of the VGG architecture and a large amount ofcomputation of the original SuperPoint network, which ensures the accuracy and reduces the amount of computation and the number of pa⁃rameters at the same time. Then, the feature points are matched using the K-nearest neighbour algorithm. Finally, the DegenerateSampling Consistency Algorithm（DEGENSAC）is used instead of the ordinary Random Sampling Consistency Algorithm（RANSAC）toremove the false matches and calculate the Homography matrix for the multi-focused microscopy image registration. Throughthe experimental validation, the proposed algorithm has higher matching accuracy and speed compared with other algorithms,and compared with the original SuperPoint, the number of parameters, computation amount and model size decreased by 51.69%,88.04% and 50.07%, respectively, and the FPS increased by about 3 times.

查看全文查看/发表评论下载PDF阅读器

关闭