摘要
针对传统特征提取方法不能很好地表示左右不对称和弯曲叶片图像信息的问题,提出一种基于深度神经网络的多尺度特征提取方法。首先借鉴空间金字塔匹配模型思想,提取各个空间子区域的高阶Zernike矩特征,使用滑动圆形窗口提取对象域的极坐标傅里叶变换描述子;其次将Zernike矩和傅里叶特征作为深度神经网络的输入向量获取深度抽象特征。实验结果表明,与多种特征提取方法相比,该方法具有较好的特征表示性能。
For the problem that traditional feature extraction methods can not express me informauon oi me a- symmetrical and curved blade well, a new classification method of multi-scale feature extraction based on deep neural network is proposed. Firstly, the high-order Zernike features moment is extracted based on spatial pyra- mid matching model from the spatial regions in the pyramid segmentation, the polar Fourier descriptor of the object regions by using sliding window. Secondly, in order to obtain the abstract features of the deep neural network, the Zemike moment and the polar Fourier descriptor are set as the input vector of the deep neural network. Finally, the recognition of the plants is achieved through the support vector machine based on proba- bility. Compared with a variety of feature extraction methods, the experiment results show that the proposed al- gorithm can reach higher accuracy rate.
出处
《西北大学学报（自然科学版）》
CAS
CSCD
北大核心
2017年第2期215-221,共7页
Journal of Northwest University（Natural Science Edition）
基金
国家自然科学基金资助项目(61373117)
高等学校博士学科点专项科研基金资助项目(20136101110019)
研究生自主创新基金资助项目(YZZ15098)
