FPGA-Based Implementation of the Quantized Neural Network for Image Super-Resolution

The convolutional neural network(CNN)-based method for image super-resolution(SR) reconstruction has been becoming an important part in the fields of image processing such as security monitoring, object tracking. However, the huge computational complexity of neural networks make it shackled to satisfy the requirements for power consumption and complexity of reconstruction systems in specific application circumstances. In this paper, we propose a structure-unsophisticated and parameter-slight CNN model, and design a quantization scheme that allows inference to be performed by using integer-only arithmetic to alleviate parameter storage and communication pressure in hardware. Furthermore, convolution loop unroll and parallel acceleration analysis are performed for the model under the premise of board resources to optimize hardware inference. The results show that when the quantization precision is 8 bits, memory requirement of parameters is reduced by 75%, and the inference peak signal-to-noise ratio(PSNR) value loss is less than 1%. Quantization model was performed on the Xilinx Zynq7035 platform with inference losses not exceeding 2%, while hardware resources consumption is greatly reduced.