【摘要】 为改善SAR图像目标识别方法低精度量化导致识别率下降严重的问题，提出采用Hessian矩阵特征分解确定量化精度分配的高效混合精度量化方法．首先采用幂迭代算法替代直接计算法来求取Hessian矩阵的主特征值，减少了运算量，提高了算法效率；然后根据卷积神经网络(CNN)不同层对权重位宽的敏感度来设置位宽，使神经网络在总体量化精度较低的情况下发挥效果；最后使用MSTAR数据集和OpenSARShip数据集对本文方法进行了验证测试．测试结果表明，在保持较高识别率下，使用高效混合精度量化方法量化后的网络与原网络相比硬件资源消耗大幅压缩．更多还原

【Abstract】 In order to address the problem of low-precision quantization of SAR image target recognition method resulting in a significant decline in recognition rate, this paper proposes an efficient mixed-precision quantization method using Hessian matrix eigenvalue decomposition to determine quantization precision allocation.Firstly, the paper employs power iteration algorithm instead of direct calculation to obtain the principal eigenvalues of Hessian matrix, which reduces the amount of computation and enhances the algorithm efficiency, and then,sets the bit width according to the sensitivity of different layers of convolutional neural network(CNN) to weight bit width, so that the neural network can play an effective role in the case of low overall quantization accuracy. Finally, the paper uses MSTAR dataset and OpenSARShip dataset to conduct a verification test on the proposed method. The test results show that, while maintaining high recognition rates, compared with the original network,the hardware resource consumption is greatly reduced of the network quantized by an efficient mixed-precision quantization method.更多还原