【摘要】 轻量级神经网络模型参数量大幅减少，且速度得到了很大的提升，然而，检测的准确率却不高。因此，对轻量级ShuffleNetV2模型进行改进，加入3×3的Depthwise卷积核，同时降低1×1的卷积核和引入注意力机制SE模块。在ImageNet数据集中进行ShuffleNetV2模型预训练。然后，将改进的ShuffleNetV2模型与其他4种网络模型进行了实验对比。结果表明：改进ShuffleNetV2模型的综合性能最佳；与SE-ShuffleNetV2模型相比，在参数量和计算量一样时，其准确率提高了7.25%。改进的ShuffleNetV2模型为移动端的声目标精确识别进一步奠定了基础。更多还原

【Abstract】 The quantity of parameters of the lightweight neural network model is greatly reduced, and the speed is greatly improved.However, the detection accuracy is not high.Therefore, the lightweight ShuffleNetV2 model is improved, add a Depthwise convolution kernel with a size of 3×3,reduce the convolution kernel with a size of 1×1 and introduce the SE module of attention mechanism.And then, the pre-training of the ShuffleNetV2 model is carried out in ImageNet dataset.The improved ShuffleNetV2 model is compared with other four network models by the experiments.The results show that the improved ShuffleNetV2 model has the best comprehensive performance.Compared with SE-ShuffleNetV2 model, the accuracy of this model is increased by 7.25 % with almost the same quantities of parameters and computation, the improved ShuffleNetV2 model lays a foundation for accurate recognition of acoustic target on mobile terminal.更多还原