Efficient FPGA-based implementation of machine learning models for high-speed lightweight binary image classification

초록

The increasing demand for swift and reliable medical diagnosis in resource-constrained environments underscores the shortcomings of current image-processing techniques regarding efficiency. Conventional platforms, like CPUs, frequently lack adequate parallelism, whereas GPUs, despite their power, tend to consume significant energy. In contrast, Field-Programmable Gate Arrays (FPGAs) provide a low-power solution with real-time processing capabilities, making them ideal for embedded diagnostic systems. Machine learning algorithms typically employed for binary image classification often yield less than optimal accuracy and introduce delays when utilized on standard hardware. To tackle these issues, this paper introduces an FPGA-based implementation of lightweight classifiers — specifically k-Nearest Neighbors (k-NN), Convolutional Neural Networks (CNN), and Decision Trees — designed to facilitate quicker malaria diagnosis in urgent scenarios. The image preprocessing pipeline consists of converting input images to grayscale, followed by binarization and resizing to fixed dimensions that are optimized for hardware execution. Each algorithm was developed in Verilog HDL and implemented on a ZYNQ-7 ZC702 FPGA board, taking advantage of its parallel processing capabilities. The proposed system achieved real-time image classification with energy efficiency and an accuracy rate of up to 80%. This research highlights the promise of FPGA-based machine learning solutions for low-power, real-time embedded medical diagnostics of malaria.

키워드