| With the advent of the era of big data,with the help of massive data information,deep neural networks with complex structures with powerful feature learning and expression capabilities have surpassed the traditional method of manually extracting features in the field of computer vision.With the improvement of network performance,the number of parameters in the network has also increased exponentially,which has brought great challenges to deploy models on hardware devices with limited resources such as computing and storage.Therefore,loading an excellent model algorithm on the edge device and running it to improve the performance of the deep neural network on the resource-constrained platform is conducive to the landing of the technology and has great application value.In view of the above problems,this paper proposes a FPGA-based pruning algorithm that can perceive the bit width of the memory bus to compress the image superresolution model.The proposed pruning algorithm uses the geometric median characteristics of the filters in the convolutional layer,combined with the L1 norm of the filter itself,to remove filters that contribute little to the network or have redundant information in order to achieve the purpose of reducing network parameters.Then use the INT8 quantization algorithm to low-bit quantize the parameters in the network model after pruning,and further compress the size of the model.Finally,we design an FPGA-based convolutional network accelerator to accelerate the compressed model.The proposed pruning algorithm comprehensively considers the bit width of the memory bus of the FPGA.By rearranging the data stream,the memory bus and the data bit width are aligned to reduce the memory access overhead and improve the execution efficiency of the FPGA.At the same time,FPGA’s high flexibility and parallelism are used to customize the accelerator hardware circuit according to the optimized algorithm characteristics.Experiments show that compared with the original model,the pruning algorithm designed in this paper improves the peak signal-to-noise ratio to 35.58 d B and achieves a 1.3% accuracy improvement.The number of parameters is reduced by 46.93%,and the inference speed is reduced from the average single image of the original model in 5.1s to 1.7s,achieving a speed increase of about 3 times. |
