Personalized image annotation via class-specific cross-domain learning

●We present a class-specific weighted nearest neighbour (cs-WNN) model for generic image annotation by directly maximizing the log-likelihood of class-specific tag predictions.●Class-specific modeling and multiple kernel learning are integrated to exploit the nonlinearity of features in the visual space of each tag.●We present a novel personalization method, namely class-specific cross-domain learning (cs-CDL), to build users’ own annotation profiles that correspond to the user-specific parameters of cs-WNN models.●The effectiveness of our proposed method has been validated on three standard datasets and two new generated datasets.

In this paper, we propose to learn users’ own profiles for image annotation with the purpose of facilitating image searching towards users’ intentions. Considering that the size of a user’s annotation vocabulary is usually small and different users have different visual understanding towards a specific tag, we perform personalization in a class-specific form to summarize a user’s annotation profiles for each tag in his vocabulary. In particular, we first exploit a generic annotation dataset with a class-specific weighted nearest neighbor (cs-WNN) model by combining the techniques of multiple kernel learning and nearest neighbor modeling. Next, a new personalization method, namely class-specific cross-domain learning (cs-CDL), is proposed to achieve users’ own annotation profiles (i.e. the user-specific parameters of cs-WNN models) by exploiting users’ annotation datasets. Experimental results have been reported over several challenging image databases to validate the effectiveness of the proposed method for both generic and personalized image annotation.