基于改进卷积神经网络的高光谱图像分类方法(特邀)
周德滨 Zhou Debin, 马金英 Ma Jinying, 曹佳琦 Cao Jiaqi, 江俊峰 Jiang Junfeng, 刘铁根 Liu Tiegen
Tianjin University of Technology and Education Tianjin University
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Aiming at the bottleneck problems of noise interference, spectral aliasing, and low computational efficiency in hyperspectral image classification, a lightweight network based on multimodal collaborative optimization (MC-DetNet) is proposed. The content features and high-frequency noise features of the original image are extracted through the dual branch architecture. The original branch uses deformable convolution to enhance the adaptability of geometric deformation. The noise branch captures local abnormal features based on spatial rich model (SRM) filtering and lightweight residual module. In order to solve the adaptability problem of multi-modal feature fusion, a gradient guided attention mechanism is proposed, and the feature weight is dynamically adjusted through the multi-directional gradient map to strengthen the identification ability of boundary sensitive regions. Further, the frequency domain spatial domain collaborative module is constructed, which combines the global energy analysis of fast Fourier transform and cross modal contrast loss to suppress the feature confusion caused by spectral aliasing. In order to improve the efficiency of the model, the optimization strategy of deep separable convolution and multi-scale hole convolution is adopted, which significantly reduces the computational complexity. Experiments on public hyperspectral datasets show that the overall classification accuracy of the proposed method is 99.1%, 93.7% and 98.9%, respectively, which is 0.9%?4.1% higher than the mainstream models. The number of parameters is only 6.2×106, and the reasoning delay is 47 ms. The proposed method effectively improves the classification accuracy in noisy and complex scenes, and has better model interpretability and edge device adaptability, providing a lightweight solution for hyperspectral edge intelligent interpretation.
逐年被引趋势
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AIRemote-Sensing Image Classification
Advanced Neural Network Applications · Spectroscopy and Chemometric Analyses
参考文献 18
此处列出前 3 条
引用本文 2
按被引量排序,此处列出前 3 条