Enhancing Object Detection in Low-Quality Images via Color-Spatial Consistency Distillation

To address the core challenges of object detection in low-quality images, this paper proposes a novel framework named Color-Spatial Consistency Distillation (CSCD). Unlike existing approaches that directly train detectors on lowquality images or apply cascaded image restoration followed by detection, CSCD introduces a feature-level interaction mechanism between the detection task and the image enhancement branch, enabling joint optimization while avoiding additional computational overhead and domain gap issues. The framework consists of two key components: Color Consistency Distillation (CCD) and Spatial Consistency Distillation (SCD). CCD aligns the color features between low-quality and high-quality images in the RGB space. At the same time, SCD extracts high-frequency texture features from the super-resolution branch and injects them into the detection pipeline to enhance structural perception. This mechanism enables the detector to decouple image quality from semantic content under degradation conditions and effectively recover discriminative features. Extensive experiments on the COCO dataset show that the proposed method achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A P, A P_{50}$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A P_{75}$</tex> scores of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$61.5 \%, 79.1 \%$</tex>, and 65.7%, respectively. When integrated into YOLOv8, CSCD improves AP by 5.0% and also provides consistent gains on YOLOv5 (+3.7%), YOLOv6 (+4.5%), and YOLOv7 (+5.2%) without increasing the inference cost of the original detectors. The method is highly modular, compatible with various object detection frameworks, and demonstrates strong potential for practical deployment across diverse application scenarios.