The perception of mixed material conditions on cigarette production lines directly affects product quality and operational efficiency. However, current monitoring methods mainly focus on raw material attributes and are difficult to comprehensively reflect production line status. To address this limitation, a novel anomaly detection method using tobacco tar residue as a key indicator is proposed. By integrating an adaptive feature enhancement (AFE) module and a multi-scale convolutional attention mechanism (MSCAM), detection precision and real-time performance for small tobacco tar residues in complex backgrounds are improved, resulting in the development of an efficient YOLO11-AM detection network. Ablation experiments conducted on standard-sampled tobacco mixtures under orthogonally optimized environmental parameters show that the proposed YOLO11-AM model achieves a mean average precision (mAP) of 97.8%, while the inference speed is improved by 24.6% compared to the baseline model, reaching 2.16 ms per image. Further industrial deployment demonstrates that the prediction error for tobacco tar residue mass is controlled within ±5%, meeting the performance requirements of online monitoring systems in cigarette factories. This study provides efficient technical support for intelligent quality control in the tobacco industry and holds significant theoretical and practical value.