Hardware Acceleration of Pre-trained Networks for Colorectal Polyp Classification

The objective of the research is to identify a pre-trained Convolutional Neural Network (CNN) with faster inference time for classification of colorectal polyps on a Field Programmable Gate Array (FPGA) device. Basically, Deep Learning(DL) models are trained and executed using Graphical Processing Units (GPUs) having high processing capacity, but those devices consume more power during execution. DL models that are appropriate for real-time miniature instruments must be implemented with low-power consumption. This paper evaluates the performance of relevant polyp classification models using FPGA (Avnet Ultra96-V2) and GPU (Quadro RTX 5000). The CNN models are trained on GPU and then deployed for inference on FPGA using the VITIS-AI platform and PYNQ framework. The MobileNetV2 performs classification with 88% accuracy and least inference time of 2.60 seconds at a throughput of 176 frames per second on FPGA. A precise comparison of the throughput and power consumption during execution on GPU and FPGA are provided. The FPGA device is found to perform
$\mathbf{1}.\mathbf{25}\times$
faster than GPU in terms of throughput while using
$\mathbf{40}\times$
less power with a negligible reduction in accuracy.