受限资源下制导武器末制导机器视觉技术研究

Research on Terminal Guidance Machine Vision Technology for Guided Weapons with Limited Hardware Resources

关键词:  自主检测识别  神经网络性能评估  硬件资源定量评估  网络高精度压缩  嵌入式应用

Key words:  Autonomous detection and recognition  Neural network performance evaluation  Quantitative evaluation of hardware resources  High-precision network compression  Embed-ded applications

摘要:

针对精确制导武器末制导机器视觉技术应用需求，研究了基于卷积神经网络的、针对复杂背景及小目标的自主目标检测识别算法，并分别进行了网络性能评估和硬件资源需求定量评估。针对最优算法，提出了基于嵌入式受限资源下的高精度神经网络压缩算法，并对算法进行了普适性评估。基于GPU嵌入式平台，实现TensorRT路线网络优化，并在速度和精度两方面均衡考虑下，对裁剪与量化算法进行了详细实验验证。实验结果表明，高精度神经网络压缩算法在硬件资源受限条件下，可以有效提升推理速度，最终经算法优化后的网络结构，可以获得3倍以上的速度提升，网络精度损失小于5%。

Abstract:

Aiming at the application requirement of machine vision technology in the terminal guidance of precision guided weapon, the autonomous target detection and recognition algorithms based on the convolutional neural network for complex background and small target are studied, and the network performance evaluation and hardware resource requirement quantitative evaluation are carried out respectively. Aiming at the optimal algorithm, high-precision neural network compression algorithm based on embedded limited resources is proposed, and the general applicability of the algorithm is evaluated. Based on the GPU embedded platform, the TensorRT network optimization is realized, Furthermore, the balance of speed and precision is verified by detailed experiments. The experimental results show that the high-precision neural network compression algorithm can effectively improve the inference speed under the condition of limited hardware resources. Finally, the network structure optimized by the algorithm can achieve more than three times the speed improvement, and the network accuracy loss is less than 5%.