CG-YOLOv11: A Smoke-Removal-Enhanced Target Detection Method for Indoor Smoke Scenes

Abstract

To address the challenges of low detection accuracy, missed detections, and false alarms in indoor fire scenarios caused by smoke, diverse postures of trapped individuals, partial occlusions by obstacles, and cluttered backgrounds, this paper proposes CG-YOLOv11, a smoke-removal-enhanced target detection method for indoor smoke scenes. Firstly, CAA-CycleGAN is employed to remove smoke and enhance image visibility. Specifically, a Color Attenuation Attention (CAA) sub-network is designed, and cyclic perceptual consistency loss together with Color Attenuation Prior (CAP) loss is introduced to improve smoke removal performance for non-uniform smoke images. Secondly, to enhance YOLOv11's feature representation and multi-scale fusion capability under occlusion and small-target conditions, we integrate the Multi-Scale Edge Enhancement Feature (MSEF) module into the original C3k2 module of YOLOv11 to form the C3k2-MSEF module, and further design a Multi-Scale Edge-Enhanced Feature Pyramid Network (MSEFFPN) to improve multi-scale feature fusion. Finally, CAA-CycleGAN and MSEF-MSEFFPN-Modified YOLOv11 (MM-YOLOv11) are cascaded to form the complete CG-YOLOv11 method, thereby further improving overall target detection performance in indoor smoke scenes. Experimental results demonstrate that CG-YOLOv11 achieves a precision of 83.0%, [email protected] of 75.5%, and a detection speed of 106.2 FPS, satisfying the accuracy and real-time requirements for rescue target detection in indoor smoke environments and validating the effectiveness of the proposed method.