适合星上图像处理的深度卷积网络轻量化方法

As the most advanced technology in machine learning, Deep Convolutional Neural Network brings great opportunities to enhance the on-board intelligent processing capability of satellite remote sensing images.However, the convolutional neural network not only has a large model and is computationally intensive, so it can't be deployed directly on onboard embedded devices with limited hardware resources and tight energy budgets. This has seriously affected the intelligent development of space-based systems. This project aims to research the lightweight method of deep convolutional networks for on-board image processing, as well as efficient hardware design and optimization techniques for thinning models, and is intended to conduct research in the following three areas. First, a pruning training method based on dynamic relative restraint and confrontation of connection is proposed, combined with sparse constrained relations among channels and influence factors, pruning is carried out to achieve a thinning and lightweight representation of the network structure. Second, a dynamic adaptive network parameter quantification mechanism is proposed to overcome the shortcomings of the existing methods such as monotonous quantification mode, poor matching between software and hardware, and large loss of quantization accuracy. Through adaptive non-uniform quantization and sparse coding, a network is compressed on computational and storage space greatly. Finally, we study the efficient hardware processing structure of lightweight and asymmetric networks, and try to solve the problem of high-speed real-time processing of convolutional neural networks on typical space-borne devices.

深度卷积神经网络作为机器学习最具代表性的先进技术，为提升星上遥感图像智能处理能力带来重大机遇。但由于卷积神经网络模型庞大、计算密集，无法在硬件资源有限、能量预算紧张的星载嵌入式设备上直接部署应用，已严重影响了天基系统的智能化发展。本项目旨在研究适合星上图像处理的深度卷积网络的轻量化方法，以及稀疏化模型的高效硬件设计与优化技术，拟在开展以下研究工作：1）提出以动态阈值为基础的强弱连接动态相对抑制和对抗剪枝训练方法，实现网络结构的稀疏化和轻量化表示；2）提出一种动态自适应的网络参数量化机制，克服现有方法量化模式单一、软硬件匹配度差、量化精度损失大的缺点，通过自适应非均匀量化和稀疏编码，以实现对网络模型在计算量和存储空间上的压缩。

深度卷积神经网络作为机器学习最具代表性的先进技术，为提升星上遥感图像智能处理能力带来重大机遇。但由于卷积神经网络模型庞大、计算密集，无法在硬件资源有限、能量预算紧张的星载嵌入式设备上直接部署应用，已严重影响了天基系统的智能化发展。本项目研究适合星上图像处理的深度卷积网络的轻量化方法，以及稀疏化模型的高效硬件设计与优化技术，在资助的一年周期内，开展的主要工作包括：建立了适合硬件平台性能的模型量化策略，制定适应平台资源的量化压缩机制，设计了基于硬件资源和不同位率的网络模型动态自适应量化方法；完成轻量化卷积神经网络的高效VLSI设计，并在FPGA处理器上完成对核心算法的验证。本项目的研究成果能够有效降低CNN模型对计算资源和存储资源的需求，灵活的实现针对不同星上资源的CNN模型的自适应量化，在保证CNN模型精度的前提下加快推理速度。此外，所提出的量化方法以及优化方案可应用于对实时性、稳定性要求较高的星上图像处理，为星上的深度学习技术开展和实施提供技术方案支持。本项目的部分研究成果已在KX联盟的某卫星上进行了应用推广。