Impact of color space on neural networks

Abstract

The choice of color space can significantly affect the performance and interpretability of neural networks in image-based tasks, but its impact remains underexplored in many deep learning applications. This study investigates how different color representations (such as RGB, HSV, LAB, and YCbCr) affect the accuracy and convergence speed of convolutional neural networks (CNNs). Through systematic experiments on benchmark datasets, we evaluate the effectiveness of these color spaces in classification and semantic segmentation tasks. Experimental results show that when using single color space on tasks like classification and semantic segmentation, traditional RGB still hasits advantages.