基于卷积神经网络的材质分类识别研究

目前,空间目标中约6%为正在工作的航天器,而约94%的空间目标为太空垃圾,严重干扰和限制了航天器发射、运行等正常的太空活动轨道,在有效清除空间碎片之前,必须对其进行有效识别。本文基于散射光谱,使用卷积神经网络对空间碎片四种材质进行分类识别,并与BP神经网络的识别结果分析比较。鉴于试验所得的材质的原始光谱信噪比低、特征信息弱等特点,需要对光谱信号进行预处理包括去噪、BRDF计算和归一化处理。然后各取四种材质的200帧样本数据进行训练,另各取50帧数据预测,结果表明：卷积神经网络的总体精度比BP神经网络低2%,耗时少101 s；而增加训练样本数据量达到每个材质各500帧时,卷积神经网络的总体精度仅比BP神经网络低0.05%,耗时则少了891 s,卷积神经网络极大的体现了其时间的优越性。该方法对大数据量的空间碎片材质的分类,具有较大的实用性和借鉴意义。

Study on classification and recognition of materials based on convolutional neural network

Before the space debris are effectively cleaned,space debris must be identified.Based on scattering spectrum,four kinds of space debris materials were classified and recognized by convolutional neural network,and the obtained results were compared with that of BP neural network.As material spectral data have the characteristics of low signal-to-noise ratio and poor feature information,spectral signals were pretreated by de-noising,BRDF calculation and normalization.It takes 200 frames as training data and another 50 frames as predicted data for four kinds of space debris materials.The classification and identification tests were done by using two methods.The results show that the overall accuracy of convolutional neural network is 2% lower than one of BP neural network,but its time-consuming is 101 s less than that of BP neural network.When the sampling amount of training data reaches 500 frames in each material,the overall accuracy of convolution nerve network is only 0.05% lower than that of BP neural network,but its time-consuming is 891 s less than that of BP neural network.Convolutional neural network has shown its extraordinary time superiority.The convolutional neural network has a large practical value for classification and identification of space debris materials.