Automatic Quantization of Convolutional Neural Networks Based on Enhanced Bare-Bones Particle Swarm Optimization for Chest X-Ray Image Classification

Deep convolutional neural networks (CNNs) have achieved high accuracy in classifying chest X-ray images to diagnose lung diseases, including COVID-19. However, their large number of parameters makes them impractical for deployment on limited-resource devices. To address this issue, we propose a neural network quantization (NNQ) method that reduces the complexity of CNNs for X-ray image classification, while maintaining their original accuracy. Our method, called mixed-precision quantization (MPQ), aims to quantize the layers with different bit widths to achieve good results. To find the optimal bit-width of each layer, we propose an enhanced bare-bones particle swarm optimization algorithm called EBPSOQuantizer. Our experiments on the COVID-19 X-ray dataset with different CNN architectures demonstrate the effectiveness of our method.