融合多分支特征与前景挖掘的异常检测方法OA

In the automated anomaly detection of supermarket goods packaging,due to the complexity and diversity of packaging design,anomaly features are often concealed within complex background information,significantly increasing the difficulty of identification.Focus on the limitations of existing methods that overly rely on predefined universal feature extraction modules,which results in ineffective extraction of anomalous features,imprecise analysis of anomaly patterns,and inaccurate localization of abnormal regions in complex scenarios,this paper proposes an anomaly detection method integrating Multi-Branch Features with Foreground Mining(MBFFM).Firstly,the proposed method uses synthetic anomalies to enhance the student network's abil-ity to perceive anomalies,and promote it to build more accurate decision boundaries in the feature space.Secondly,a multi-branch feature extractor is introduced to extract and fuse multi-level and multi-scale features to obtain rich texture and semantic information,and enhance the adaptability of the model to the actual scene.Then,the foreground mining loss is used to reduce the error of foreground reconstruction,supplement local structural information,and improve the detection ability of the model to structural anomalies.The comprehensive performance of the proposed method on the GoodsAD dataset outperforms eight SOTA methods.The qualitative,quantitative and ablation experimental results verify the effectiveness of the internal modules within the overall framework of the proposed method.