一种基于在轨深度学习的压缩率确定方法

Method for Determining the Compression Rate of On-orbit Deep Learning

如何对遥感图像中的重要目标进行精确的识别分类,是卫星遥感领域的一个难点和重要的研究方向。深度学习能较好解决识别分类的问题,但其学习模型有大量参数需要确定,会消耗大量计算和存储资源,不利于在轨实现。模型压缩是降低资源需求的有效方法,但会导致分类准确率降低。剪枝是模型压缩的主要方法之一,目前剪枝技术大多研究的是在降低计算量的情况下如何减少准确率损失,如何确定压缩率是有待研究的问题。本文提出了一种通过函数拟合准确率与压缩率关系的方法,可以据此确定相应的压缩率,并对不同的压缩方法进行比较。仿真结果表明:该方法的函数模型可以在不同场景下用较少的点拟合出准确率与压缩率关系曲线,且均方根误差最大为1.09,平均值为0.51,拟合效果较好,可据此针对不同的应用条件与需求确定相应的模型压缩率。

How to accurately identify and classify the important targets in remote sensing images is a difficult and important research direction in the field of satellite remote sensing. Deep learning can well solve the problem of recognition and classification. However, the learning model has a large number of parameters to be determined, which will consume a lot of calculation and storage resources, and thus is not suitable to on-orbit implementation. Model compression is an effective way to reduce the resource requirements. However, it will result in a decrease in the classification accuracy. Pruning is one of the main methods for model compression. However, most of the studies on pruning methods at present are focused on how to reduce the loss of accuracy while reducing the amount of calculation.How to determine the compression rate is a problem to be studied. In this paper, a method of fitting the relationship between the accuracy and the compression rate by a function model is proposed, with which the corresponding compression rate can be determined. The simulation results show that the function model can fit the curves of accuracy and compression rate with fewer points in different scenarios. The maximum root mean square error is 1.09, and the average value is 0.51. The fitting effect is good. With the proposed method, the model compression rates under different application conditions and requirements can be determined.