Abstract
<jats:p>In application scenarios such as medical imaging and industrial inspection, object detection models are often trained on grayscale images but deployed under mixed grayscale–RGB inputs. Experimental observations reveal a pronounced asymmetry in this setting: models trained on RGB datasets generalize well to grayscale detection tasks, whereas models trained on grayscale datasets exhibit severe performance degradation and substantial inter-run variability when applied to RGB inputs. This phenomenon significantly limits practical deployment and cannot be fully explained by the conventional assumption that grayscale training lacks chromatic information.In this work, we analyze the chromatic information processing mechanism in the first layer (layer 0) of YOLO-based object detection models. We reveal that training on grayscale samples induces variance collapse in chromatic channels, and further elucidate how this mechanism leads to RGB detection mismatch. Based on this analysis, we propose a kernel-aware regularization method that adjusts the chromatic sensitivity of the first layer by modulating the ratio between chromatic and luminance channels, thereby effectively mitigating the impact of variance collapse.Experimental results demonstrate that the proposed method reduces the mAP50 drop of YOLO grayscale models on RGB detection from 10% ~ 20% to 1% ~ 2%. It also effectively alleviates performance degradation in RTDETR and FasterRCNN under the same cross-modal scenarios, significantly improving the robustness of model training with respect to variations in input data distribution.</jats:p>