基于多尺度分析和神经网络的目标识别方法

Target recognition method based on multi-scale analysis and neural network

阐述了一种基于多尺度分析理论和神经网络的目标识别方法,对图像数据进行几何重构,包括图像预处理和飞行姿态的重构。采用Contourlet变换提取图像的低频及高频特征向量作为基础输入训练集和修正输入训练集。在此基础上进行BP神经网络的设计,利用高频细节数据修正低频轮廓数据,并确定输入输出层、中间层个数和算法。训练好的网络显示对不同光照条件、姿态的机动目标具有较高的识别能力,说明该方法具有工程可适用性。

A target recognition method is discussed based on the theory of multi-scale analysis and neural network.Firstly,the geometric reconstruction of the image data is performed including image preprocessing and flight attitude reconstruction.Then,the feature vectors of low and high frequency of the images extracted through the Contourlet transform are used as the basic input training set and modified input training set.On this basis,the BP neural network is designed.The low frequency contour data are modified by using the high-frequency details,and the number and algorithm of the input middle and output layer are determined.The trained network has strong recognition capability under different lights and flight attitudes,indicating that the method is of high engineering applicability.