Model adaptation via credible local context representation

Conventional model transfer techniques, requiring the labelled source data, are not applicable in privacy-protected medical fields. For challenging scenarios, recent data-free domain adaptation (SFDA) has become a mainstream solution but losing focus on inter-sample class information. This paper proposes new Credible Local Context Representation approach for SFDA. Our main idea is to exploit credible local context more discriminative representation. Specifically, we enhance model's discrimination by information regulating. To capture context, discovery method developed that performs fixed steps walking deep space and takes features this path as context. In epoch-wise adaptation, clustering-like training conducted with two major updates. First, all target data constructed then context-fused pseudo-labels providing semantic guidance generated. Second, each weighting fusion its forms anchored neighbourhood structure; thus, clustering switched from individual-based coarse-grained. Also, regularisation building drive coarse-grained learning. Experiments three benchmarks indicate proposed can achieve state-of-the-art results.