COF-YOLO: A Contour-Oriented Framework for Rocket Detection Under Visual Degradation Conditions

Abstract

Rocket detection in launch imagery remains difficult when the target undergoes rapid scale variations and is obscured by smoke, plumes, cloud cover, or strong illumination. Under these conditions, local texture cues are often weak or unstable, while the visible shape of the target remains relatively more informative. This makes contour-related representation important for reliable detection. To address this problem, this paper proposes COF-YOLO, a contour-oriented detection framework built upon YOLOv12. The proposed method establishes a progressive contour enhancement pipeline that combines image-level augmentation, backbone feature modeling, and pre-prediction feature refinement within a deep detection framework. Specifically, a contour-preserving data augmentation strategy (CPDA) is introduced to improve sample diversity while maintaining geometric integrity; a C3k2_Shape module is designed to enhance contour-aware representation by integrating deformable convolution and frequency-domain decomposition; and a lightweight ShapeRefinement module is employed to capture multi-scale contour context and refine features before prediction. Experimental results on public rocket datasets, together with qualitative tests on real-scene optical launch images, demonstrate that COF-YOLO consistently outperforms the baseline YOLOv12 and several mainstream lightweight detectors, especially under small-target and occluded conditions, while maintaining practical inference efficiency. These results indicate that introducing a contour-oriented inductive bias is an effective way to improve robust detection in weakly textured and feature-degraded visual environments.