基于迁移学习的嵌入式数字识别系统

Embedded digital recognition system based on transfer learning

针对现有神经网络数字识别模型因结构复杂在终端设备部署面临的计算性能瓶颈,设计了一种轻型数字识别网络EmbedNet,来降低模型对高性能硬件的依赖.同时针对模型识别真实图片效果差的问题,使用迁移学习和灰度分布调整结合的方法提高模型的识别准确度.实验结果显示:轻量级网络EmbedNet在数据集MNIST中的识别精度达到97.4%,经改进迁移学习方法训练后,模型对真实样本的识别准确率从29%提高到82%,该方法有效提升了模型的识别能力.整体方案能够在嵌入式设备中运行.

To tackle the computational performance bottleneck encountered by the existing neural network digital recognition model when deployed at the terminal equipment due to its complex structure,a lightweight digital recognition network EmbedNet was designed to reduce the model's dependence on high-performance hardware.At the same time,in view of the problem of the model's poor recognition of real images,transfer learning and gray distribution adjustment methods were used to improve the recognition accuracy of the model.The experimental results show that the recognition accuracy of the lightweight network EmbedNet in the data set MNIST reaches 97.4%.After training by the improved transfer learning method,the model's recognition accuracy of real samples has increased from 29%to 82%,which effectively improves the model's recognition ability.The overall solution can be run in embedded devices.