A Novel CNN Architecture on FPGA-based SoC for Remote Sensing Image Classification

The convolutional neural network (CNN) has achieved extraordinary results for image classification and recognition in-orbit satellite remote sensing applications. However, the implementation of CNN is difficult due to the increasing computational accuracy and complexity of CNN and the special working environment for satellites in orbit. A promising solution for this problem is field-programmable gate array (FPGA) due to its sufficient supporting for parallel computing with low power consumption. And system-on-chip (SoC) is a system that supports parameter reconfiguration. Hence, through applying the Lenet-5 based remote sensing image classification method to SoC, we propose several network mapping methods to reduce the resource consumption and enhance the computational efficiency of FPGA, then, we utilize the proposed methods to design several key modules for the implementation of CNNs, i.e., the convolution module, sigmoid module, pooling module, full connected module, system control module, on-chip cache model, and off-chip DRAM module. The test results show that the proposed hardware system architecture can implement Lenet-5 network and realize the realtime requirements of the system.