An Efficient Source-Free Domain Adaptation Method

View the PDF file for the paper entitled translating the feature for identifying

PDF HTML (experimental) view

a summary:Facial expression identification forms are used in many video -based emotional computing applications, such as human computer interaction and health care monitoring. However, deep Fer models are often struggled with hidden expressions and high changes between the topic, which limits their performance in the real world applications. To improve its performance, SFDA -free domain methods have been proposed to customize a pre -exported source model using only the target field data that is not named, thus avoiding the privacy of data, storing it, and transmission restrictions. This paper deals with a difficult scenario, as the source data is not available for adaptation, and only the inexpensive targeted data that consists of only neutral expressions is available. SFDA methods are usually designed to adapt using the target data of only one category. Moreover, the use of models to create face images with non -neutral expressions can be unstable and intense. In this paper, the PFT feature is suggested to SFDA. Unlike the current SFDA photos, the light weight method works in the underlying space. We first train the translator on the source of the source field to convert the feature features of the topic from one source to another. Expression information is preserved by improving a set of consistency and elegance goals. Next, the translator is adapted to the neutral target data, without using the source data or the image creation. Through translation in the underlying space, PFT avoids the complexity and noise of the face expression, which improves the discriminatory inclusion of the classification. The use of PFT leads to canceling the need to create the image, reduces the general account expenditures (using a lightweight translator), and adapts only part of the form, which makes the method effective compared to the translation based on images.