【摘要】 VGGNet能提供高精度的火星图像分类，但需消耗大量内存资源。鉴于器载计算机内存资源有限，为解决这一矛盾，本文提出了基于迭代剪枝VGGNet的火星图像分类方法。首先，采用迁移学习训练网络的连通性，以便评估神经元的重要性；其次，通过迭代剪枝方法修剪不重要的神经元，以便将全连接层的参数量和内存占用量减少；最后，采用K-means++聚类实现权重参数的量化，利用霍夫曼编码压缩迭代剪枝与量化后的VGGNet权重参数，达到减少存储量和浮点数运算量的作用。此外，通过5种数据增强方法进行数据扩充，目的是解决类别不平衡的问题。实验结果表明，压缩后的VGGNet模型的所占内存、Flops和准确率分别为62.63 Mb、150.6 MFlops和96.15%。与ShuffleNet、MobileNet和EfficientNet等轻量级图像分类算法相比，所提模型具有更好的性能。更多还原

【Abstract】 VGGNet can provide high-precision Martian image classification, but consumes vast memory resources. Considering the limitation of memory resources of the onboard computer, a Martian image classification method based on iterative pruning VGGNet is proposed to solve this contradiction. Firstly, the transfer learning is used to train the connectivity of the network in order to evaluate the importance of neurons. Secondly, to reduce the number of fully connected layer parameters and memory consumption, the iterative pruning method is used to prune unimportant neurons. Finally, K-means++ clustering is used to quantify the weight parameters, and Huffman coding compresses the weight parameters of VGGNet after iterative pruning and quantization to reduce the storage capacity and floating point arithmetic.Furthermore, the data augmentation is carried out through five data augmentation methods to address the class imbalance. Experimental results show that the memory, Flops and accuracy of the compressed VGGNet model are 62. 63 Mb, 150. 6 MFlops and 96. 15%, respectively. Compared with lightweight image classification algorithms such as ShuffleNet, MobileNet and EfficientNet, the performance of the proposed model is better.更多还原