SAID: Segment All Industrial Defects with Scene Prompts

SAID:利用场景提示对所有工业缺陷进行细分

阅读:1

Abstract

In the field of industrial inspection, image segmentation is a common method for surface inspection, capable of locating and segmenting the appearance defect areas of products. Most existing methods are trained specifically for particular products. The recent SAM (Segment Anything Model) serves as an image segmentation foundation model, capable of achieving zero-shot segmentation through diverse prompts. Nevertheless, SAM's performance in special downstream tasks is not satisfactory. Additionally, SAM requires prior manual interactions to complete segmentation and post-processing of the segmentation results. This paper proposes SAID (Segment All Industrial Defects) to deal with these issues. The SAID model encodes single-annotated prompt-image pairs into scene embedding via Scene Encoder, achieving automatic segmentation and eliminating the reliance on manual intervention. Meanwhile, SAID's Feature Alignment and Fusion Module effectively addresses the alignment issue between scene embedding and image embedding. Experimental results demonstrate that SAID outperforms SAM in segmentation capabilities across various industrial scenes. Under the one-shot target scene segmentation task, SAID also improves the mIoU metrics by 5.79 and 0.87 compared to the MSNet and SegGPT, respectively.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。