A Winograd-based CNN Accelerator for Cloud Detection on FPGA

Cloud detection with a convolutional neural network is an effective method for improving the accuracy of remote sensing image processing, but it takes a large number of convolutions with repeated floating-point multiplications which is difficult for high-speed onboard processing. Winograd is always used for accelerating convolution. However, the efficiency of Winograd implemented on the general-purpose processors is low, which requires FPGA to provide parallel computing capabilities for the Winograd optimized cloud detection network. This paper proposes a Winograd-based CNN accelerator for cloud detection on FPGA. Firstly, a cloud detection network is proposed based on the improved U-Net, which provides the capability of high accuracy of cloud detection with low computing complexity. Then, the convolution in the cloud detection network is improved by the Winograd algorithm to reduce the number of floating-point multiplications. And a novel cloud detection accelerator based on FPGA is proposed. The accelerator performs a sufficient parallel operation on the convolution improved by Winograd, which further improves the processing efficiency. Experimental results indicate that based on ensuring that the average pixel accuracy is greater than 95%, the processing time for a 256×256 image can reach 996ms, which is 35 times faster than the ARM deployment only.