Method for constructing deep adaptation network based on robust soft label

Abstract

The invention belongs to the field of computer vision, and provides a deep adaptation network construction method based on robust soft labels. According to the method, the nuclear norm of a probability matrix is predicted to the maximum, the noise probability is filtered out, and meanwhile, the important probability is reserved, so that a new label form is generated, and then label induced losses such as CMMD, intra-class distance and inter-class distance are expressed again through the new label form in a probability weighting mode. The robust soft label constrained by the nuclear norm is proved to have more reliable characterization capability compared with the traditional hard label and soft label, and the label induced loss is re-expressed in a probability weighting mode and is also proved to be loss function modeling with higher interpretability and rationality; therefore, the problem that a traditional hard tag and a traditional soft tag are too large in deviation during tag correlation loss modeling is solved.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to a construction method of a deep adaptive network based on robust soft labels. Background technique [0002] Although deep learning techniques have been widely used, the performance and gains of deep learning rely heavily on large amounts of labeled data. In practical scenarios, manually annotating enough training data often takes a lot of time and cost, and the problem of domain drift between different sources of data is also a big obstacle. Therefore, the concept of domain adaptive (DA) is proposed to utilize the knowledge of label-enriched domain (i.e., source domain) to help the learning of related but unlabeled domain (i.e., target domain). [0003] Domain adaptation works by reusing a classifier trained on a source domain to annotate another unlabeled target domain, which follow different distributions. A crucial issue in the domain adaptation task is how to redu...

AI Technical Summary

Problems solved by technology

The method produces a form of robust soft labels by maximizing the kernel norm of the predicted probability matrix, while filtering out the noisy probabilities, and then reformulates the labels in a probability-weighted manner. Induced loss, which solves the problem of excessive deviation of traditional hard labels and soft labels when constructing label-induced losses

Images