基于卷积神经网络的材质分类识别研究

发布时间：2018-03-12 08:19

本文选题：空间碎片　切入点：散射光谱　出处：《激光与红外》2017年08期 　论文类型：期刊论文

【摘要】：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。
[Abstract]:At present, about 6% of the space targets are active spacecraft, while about 94% of the space targets are space junk, which seriously interferes and restricts the normal orbit of space activities, such as the launch and operation of the spacecraft, before effectively removing space debris. It must be recognized effectively. In this paper, four materials of space debris are classified and identified by convolution neural network based on scattering spectrum. Compared with the recognition results of BP neural network, in view of the low signal-to-noise ratio of the original spectrum and the weak characteristic information of the material obtained from the experiment, It is necessary to preprocess the spectral signal including denoising BRDF calculation and normalized processing. Then we take 200 frames of sample data of four different materials for training and 50 frames each for prediction. The results show that the overall accuracy of the convolutional neural network is lower than that of the BP neural network, and it takes less than 101 s, and when the training sample data is increased to 500 frames per material, The overall accuracy of convolution neural network is only 0.05 lower than that of BP neural network, and the time consuming is 891 s less. The convolution neural network greatly reflects the superiority of time. It has great practicability and reference significance.
【作者单位】：长春理工大学理学院;
【分类号】：TP183;V528

【相似文献】