.png){kind=icon}
.png){kind=icon}
Return
Software Optimization Techniques for CNN-Based Object Detection Systems in Zynq FPGA
.png){kind=icon}
OA
AI
DOI：10.1049/ell2.70455.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Convolutional neural networks (CNNs) have widely been adopted in object detection systems to achieve high performance. The CNN-based object detection systems consist of three stages: preprocessing, neural network processing and postprocessing. The preprocessing stage becomes a critical bottleneck as lightweight CNNs have been developed and the neural network processing stage is generally accelerated in hardware. As multiple operations are performed in the preprocessing stage, large memory access is required for intermediate data between the operations. In addition, a number of computations are required to resize an input image using image interpolation where interpolation ratios are calculated for each pixel. This paper proposes software optimization techniques to reduce the memory access and computations in the preprocessing stage. The operations performed in separate nested loops are efficiently fused using local registers to remove memory access. In addition, redundant computations are removed by caching the interpolation ratios across frames. To evaluate the proposed techniques, they were applied to an object detection system implemented on the Xilinx Zynq-7000 FPGA platform. The experimental results show that the proposed techniques reduce the preprocessing time by 2.19 to 2.67 and increase the overall inference throughput by 1.61 to 1.79 compared to the baseline.
Keywords:
computer vision
convolutional neural nets
field programmable gate arrays
optimisation
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.