FPGA Implementation of an Improved ShuffleNetV2 Rice Diseases Identification Algorithm

In recent years, the development of agricultural technology has opened up new possibilities for the realization of precision agriculture. In rice production management, its yield and quality are very much affected by the impact of pests and diseases. In order to identify and manage rice diseases in a timely and accurate manner, researchers have proposed an image identification method based on CNN (Convolution Neural Networks) with the help of deep learning technology. However, the traditional convolutional neural network has many parameters, large computation volume, slow inference speed deployed on CPU and GPU, and large power consumption, etc. To address the above problems, we use the new AugShuffle Block1 module on top of the original ShuffleNetV2, reduce the number of stacks of Shuffle Block1, replace pooling by convolutional stepping, and add the ECA module after the last convolutional layer. The new network AugShuffle _ Eca _lite was constructed by replacing pooling with convolutional stepping and adding the ECA module after the convolutional layer, which significantly reduced the network size. Then for the special structure of AugShuffle_ECA_lite network, the accelerator system was optimized using streaming parallel architecture design, mixing layer design, convolutional computation optimization and templating. A lightweight convolutional neural network AugShuffle _ Eca _lite rice diseases identification system with excellent performance and low power consumption was finally implemented on the zynq7020 development board. The processing speed is about 17.5 times faster than CPU and 1.47ms faster than GPU, while the power consumption is about 52.6% lower than GPU, which effectively improves the accelerator inference speed.