基于改进的Alexnet的服装识别及FPGA加速实现

Clothing Recognition Based on Improved Alexnet and Accelerated Realization of FPGA

为了快速识别服装款式类型,针对现有的服装识别网络模型有所累赘以及速度响应仍然有待提升的问题,设计了一种改进的Alexnet。采用小卷积核替换大卷积核的方式减少计算量并提升识别准确度,相比于原神经网络提升了2个百分点;然后通过对网络进行权重提取,最后在FPGA进行网络搭建,通过FPGA的高运算功能进行加速,在稍微减少改进的Alexnet准确度的情况下大幅度地提升了识别速度,相比于ARM平台提升了3倍。

For the quick identification of the type of clothing style, the existing network models for clothing recognition are cumbersome and the speed response still needs to be improved. An improved Alexnet is designed which uses a small convolution kernel to replace a large convolution kernel to reduce the amount of calculation. The recognition accuracy is improved by 2 percentage points compared with the original neural network. Then, by extracting the weight of the network, and finally building the network in ZYNQ,it is accelerated by the high computing function of FPGA. With a cost of slight reduction of the accuracy of the improved Alexnet, the recognition speed is greatly improved, which is 3 times higher than that of the ARM platform.