An Explainable AI (XAI) Approach to Assessing the Significance of Color in Object Detection

Object detection is a fundamental task in computer vision with applications ranging from autonomous driving to surveillance systems. While considerable research has focused on model architectures and training strategies, less attention has been paid to the impact of color space selection on detection performance. This is significant because color s paces determine how visual information is encoded, directly affecting feature extraction and model robustness in diverse lighting and environmental conditions, which are critical for applications like autonomous driving. Therefore, this paper examines how the choice between RGB and YGB color spaces influences object detection capabilities, with particular focus on the significance of color information in feature extraction and recognition. Two identical YOLO models were trained: one on standard RGB images and another on YGB images, where the Y channel from YCbCr color space replaces the R channel. Through comprehensive analysis of detection patterns, heatmaps of attention mechanisms, faithfulness, color channel relevance, object detection sensitivity, color channel contributions, and masking effects, in this paper, Layer-wise Relevance Propagation (LRP) is used to demonstrate differences in how these models process and utilize visual information. While an established XAI tool was adopted, a structured pipeline combining LRP, channel masking, and custom-developed metrics provides a novel framework for quantifying the role of luminance in object detection. Results show that the YGB model achieves more precise feature localization, demonstrates superior performance in variable lighting conditions, and exhibits a more efficient hierarchical encoding of information.