Binarization and Classification of RGB Images

Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in various real-world information processing tasks, such as object classification. However, CNNs tend to be computationally and power intensive, making them less suitable for implementation on wearable and embedded systems. One potential solution is using binary models, which have been proven to significantly reduce memory and computational requirements. Nevertheless, this reduction in resources often leads to a decrease in accuracy and raises questions about the most effective way to handle non-binary inputs such as color images. While a direct mapping from binarization to accuracy loss is almost impossible to predict analytically, we empirically study the problem of input binarization and its impact on the classification task. In particular, we explore how multiple-valued binarization, as opposed to single value binarization, will affect the neural model classification ability. We conducted a study in which we explored three different color channel feature fusion methods and found that extracting features from each color channel separately and then fusing those features at later stages is the most effective approach for preserving classification accuracy in the binary network with multiple-valued thresholding inputs. We conducted experiments on two color image datasets in classification tasks with increasing difficulty. The results demonstrate that while the accuracy for the GTSRB datasets remain comparable to its full-precision counterpart, the binarization process noticeably affects the representational power and accuracy of CNNs when applied to the more complex in textures CIFAR-10 dataset.