Integrating Channel Fusion and Channel Attention in Night Time Object Detection: An XAI Analysis

The increasing adoption of artificial i ntelligence in critical applications necessitates eXplainable AI (XAI) techniques for greater transparency and trust. Object detection in challenging low-light nighttime conditions remains a significant challenge. This research employs XAI to provide empirical explanations for the effects of a simple input channel fusion method utilizing a YGB color space derived from YCbCr and RGB, and the integration of a Convolutional Block Attention Module (CBAM), which incorporates both Channel Attention and Spatial Attention. Using a standard YOLOv8 baseline and a YOLOv8 variant enhanced with CBAM, we investigate their behavior on nighttime dashcam imagery under different input color space configurations (RGB and YGB). To explain how the YGB color space fusion influences feature processing and model performance, and how the channel and spatial attention mechanisms within CBAM impact model focus and detection robustness with varying inputs, we apply XAI techniques, including Class Activation Maps (CAMs), such as Grad-CAM, and analyze raw feature maps from key layers. By interpreting the models’ activation patterns and correlating them with detection outcomes. This work offers XAI-driven insights into the specific contributions and interactions of YGB channel fusion and CBAM attention for improving nighttime object detection using readily applicable techniques.