| 遥感图像分割的低维纹理特征算子与双变异蝴蝶优化算法 |
| |
| 引用本文： | 高鹏,李佳田,杨汝春,张泽龙,杨超,张兴忆.遥感图像分割的低维纹理特征算子与双变异蝴蝶优化算法[J].武汉大学学报(信息科学版),2023,48(1):165-174. |
| |
| 作者姓名： | 高鹏 李佳田 杨汝春 张泽龙 杨超 张兴忆 |
| |
| 作者单位： | 1.昆明理工大学国土资源工程学院, 云南 昆明, 650093 |
| |
| 基金项目： | 国家自然科学基金资助项目41561082 |
| |
| 摘 要： | 针对光学遥感影像受天气影响导致部分地物间形成弱边缘问题，提出一种低维纹理特征算子与双变异蝴蝶优化算法。首先提出一种适用于遥感影像的低维完备局部三值模式的纹理特征提取算子，并将其引入简单线性迭代聚类算法，对遥感影像进行初始分割，减小了噪音影响，同时增强算法对弱边缘的敏感度和分割准确性；然后采用双变异蝴蝶优化的支持向量机合并同质超像素块，以简单线性迭代聚类算法和低维纹理特征算子得到的综合特征作为输入，得到最终分割图像。利用2组高分辨率遥感影像进行分割实验，并与当下流行的卷积神经网络进行对比，实验结果表明，所提算法相较于传统算法对弱边缘有更好的分割效果，数据一的边界回归率(boundary recall，BR)值较对比算法平均提高了1.9%，Kappa系数平均提高了0.036；数据二的BR值较对比算法平均提高了2.33%，Kappa系数平均提高了0.027。对比实验证明了所提算法相较于卷积神经网络有更好的泛化性。
|
| 关 键 词： | 高分辨率遥感图像 超像素 纹理特征 支持向量机 图像分割 |
| 收稿时间： | 2022-03-20 |
Remote Sensing Images Segmentation Based on Low-Dimensional Texture Feature Operator and Double-Mutant Butterfly Optimization Algorithm |
| |
| Affiliation: | 1.Faculty of Land Resources Engineering, Kunming University of Science and Technology, Kunming 650093, China |
| |
| Abstract: | Objectives Remote sensing image segmentation is one of the key steps in the remote sensing image analysis progress. Due to the influence of illumination and other cases, the high-resolution remote sensing images show weak edges between different features, which is of great disadvantages to the subsequent image segmentation. The low?dimensional texture feature operator combined with simple linear iterative clustering (SLIC) and the support vector machine (SVM) optimized by double-mutant butterfly optimization algorithm (DBOA) are proposed to solve the problem. Methods Firstly, the texture extraction operator called low-dimensional completed local ternary pattern (LCLTP) suitable for remote sensing image is proposed, which is based on completed local ternary pattern (CLTP) and has less dimension compared with CLTP. Meanwhile, because of its robustness to illumination, LCLTP can identify weak edges very well. Secondly, we introduce LCLTP into SLIC to perform initial super-pixel segmentation and combine the feature of LCLTP and SLIC as the comprehensive feature. And DBOA is used to optimize the parameter selection of SVM. Finally, the trained SVM model is used to classify the super-pixels and the final segmentation image is obtained. Results Compared with traditional algorithms, the proposed method is more sensitive to the weak edges and more robust to independent noises. It performs better segmentation to pivotal objectives such as buildings, plants and roads. Compared with convolutional neural network, the boundary recall (BR) values of the proposed method are increased by 1.9% and 2.33% on dataset-1 and dataset-2, respectively. Kappa coefficient are increased by 0.036 and 0.027, respectively. Conclusions The proposed method has better generalization ability with higher segmentation accuracy for the weak edge by combining LCLTP and SLIC, meaning that, the comprehensive features can distinguish different objects more effectively. By taking the comprehensive features as the input of SVM, the accuracy of classification is improved and the time of sample production is saved. DBOA is used to optimize the parameter selection process of SVM, which improves the parameter selection precision and the running efficiency. |
| |
| Keywords: | |
|
| 点击此处可从《武汉大学学报(信息科学版)》浏览原始摘要信息 |
|
点击此处可从《武汉大学学报(信息科学版)》下载全文 |
|
