A Kernel-Aware Regularization Model for Chromatic-Robust Detection: Analysis of Grayscale-to-RGB Generalization

Object detection models can experience substantial performance degradation when the input representation at deployment differs from that used during fine-tuning. In this study, we investigate a specific grayscale-to-RGB distribution shift in which detectors fine-tuned on grayscale-derived images are subsequently evaluated on RGB inputs. Experimental observations reveal a pronounced asymmetry in this setting: RGB-trained models generally retain stable performance on grayscale inputs, whereas grayscale-trained models exhibit severe degradation and substantial run-to-run variability on RGB inputs. We analyze the chromatic information processing mechanism in the first layer of multiple YOLO-family detectors and show that grayscale fine-tuning induces chromatic variance collapse in specific channels, which is closely associated with RGB detection mismatch. Based on this analysis, we introduce ColorScore to quantify kernel chromatic sensitivity and propose a kernel-aware regularization method that controls the chromatic-luminance sensitivity allocation of the first input layer, thereby mitigating variance collapse without modifying the inference architecture. Across four YOLO backbones, the proposed method reduces the mean absolute mAP50 gap between grayscale and RGB evaluation from 0.140 to 0.014. The method also alleviates performance degradation in RT-DETR-L and Faster R-CNN under the same grayscale-to-RGB evaluation setting.