基于双空间协同异常合成的卷烟外观缺陷检测方法OA

Cigarette appearance defect detection is a critical technical process for ensuring product quality and brand image.Addressing the challenges of weak defect recognition and insufficient cross-domain generalization in existing detection methods,we propose CigDefect-CAS—a cigarette appearance defect detection framework based on dual-space collaborative anomaly synthesis.This framework constructs local anomalies in the image space and global anomalies in the feature space,enhancing the model's ability to learn weak defects.During feature space anomaly synthesis,gradient ascent and truncated projection mechanisms are introduced.This ensures synthesized anomalies cluster near the decision boundary while preventing excessive feature shifts,thereby enhancing the distribution's plausibility and realism.Furthermore,the detection framework integrates an attention mechanism to enhance perception of fine-grained structural features.Feature distribution matching technology is introduced to reduce distribution discrepancies between training and testing datasets,thereby improving the model's robustness and generalization performance across domains.Experimental results demonstrate that the proposed method achieves 99.5%,99.2%,and 99.4%on image-level AUROC,pixel-level AUROC,and PRO metrics respectively.It significantly outperforms existing mainstream methods in both weak defect detection accuracy and cross-domain generalization performance,validating the effectiveness and advanced nature of this framework.