Precise few-shot segmentation of metal surface defects is vital for industrial production and cost control.However,mainstream few-shot segmentation algorithms following the meta-learn-ing paradigm,which employ single prototypes for feature encoding,struggle to represent the com-plex foreground-background distribution of defect samples.Hence developing a model to alleviate the foreground-background perception ambiguity remains a significant challenge.To address this is-sue,we propose a Decoupling Fusion Mechanism-based Network(DFMNet).The defect segmenta-tion task is decomposed into two semantic segmentation subtasks one for known classes(defect ar-eas)and another for base classes(background areas).The network employs a graph reasoning-based meta parsing module to generate coarse segmentation masks for seen classes,and utilizes a back-ground descriptor-based base class guiding module for base class coarse masks,achieving the objec-tives of the corresponding subtasks.The introduction of an information fusion module,grounded in a decoupling fusion strategy,processes the coarse predictions of both subtasks,mitigating the fore-ground-background imbalance and enabling finer segmentation.An adaptive adjustment factor based on the Gram matrix is introduced to focus the model more on feature variations caused by scene dif-ferences,thereby enhancing the model's generalization ability.Experiments conducted on the metal surface defect datasets,with extensive comparative analysis and ablation studies against various ex-isting few-shot segmentation networks,demonstrate that our method effectively alleviates the issue of foreground-background perception ambiguity and achieves state-of-the-art results.
metal surface defect detectionfew-shot semantic segmentationgraph reasoningback-ground descriptoradaptive adjustment factors
维普
万方数据
