Color Image Classification on Edge by Fusing Visual Feature Extractors and Weightless Neural Networks

Abstract

Weightless Neural Networks (WNNs) are a promising alternative for the image classification problem in an edge computing environment due to their low computational cost. However, the direct use of this model requires converting the image to grayscale to facilitate the conversion of an image into its binary representation. This work investigates the performance of WNNs for the color image classification problem without adopting the grayscale conversion strategy. Several image preprocessing and representation strategies are evaluated, including grayscale conversion, feature extraction techniques widely adopted in the field of computer vision such as VLAD and Fisher Vector, and Convolutional Neural Networks (CNNs). Several experimental scenarios were conducted with the CIFAR-10 dataset, evaluating the models with 7 feature extractors (5 CNN architectures) in terms of accuracy, precision, recall, F1-score and execution time. The results show that hybrid models with CNNs outperform other feature extractors, achieving 89.57% accuracy with the ConvNeXt model, surpassing the accuracy of pure WNN (44.01%) and resulting in lower computational costs compared to pure CNNs.