FPGA-Based Inter-layer Pipelined Accelerators for Filter-Wise Weight-Balanced Sparse Fully Convolutional Networks with Overlapped Tiling

Convolutional neural networks (CNNs) exhibit state-of-the-art performance while performing computer-vision tasks. CNNs require high-speed, low-power, and high-accuracy hardware for various scenarios, such as edge environments. However, the number of weights is so large that embedded systems cannot store them owing to their limited on-chip memory. A different method is used to minimize the input image size, for real-time processing, but it causes a considerable drop in accuracy. Although pruned sparse CNNs and special accelerators are proposed, the requirement of random access incurs a large number of wide multiplexers for a high degree of parallelism, which becomes more complicated and unsuitable for FPGA implementation. To address this problem, we propose filter-wise pruning with distillation and block RAM (BRAM)-based zero-weight skipping accelerator. It eliminates weights such that each filter has the same number of nonzero weights, performing retraining with distillation, while retaining comparable accuracy. Further, filter-wise pruning enables our accelerator to exploit inter-filter parallelism, where a processing block for a layer executes filters concurrently, with a straightforward architecture. We also propose an overlapped tiling algorithm, where tiles are extracted with overlap to prevent both accuracy degradation and high utilization of BRAMs storing high-resolution images. Our evaluation using semantic-segmentation tasks showed a 1.8 times speedup and 18.0 times increase in power efficiency of our FPGA design compared with a desktop GPU. Additionally, compared with the conventional FPGA implementation, the speedup and accuracy improvement were 1.09 times and 6.6 points, respectively. Therefore, our approach is useful for FPGA implementation and exhibits considerable accuracy for applications in embedded systems.