Crop the Way You Like: Personalized Image Cropping by Integrating Subjective and Objective Features

Image cropping aims to find more attractive crops for users by recomposing images. With the development of mobile social media, the significance of this technology is increasingly evident. Most existing methods focus on learning from average cropping results and rely on objective image features for image cropping. However, the subjectivity of image cropping and the preferences of people are often overlooked in this paradigm. In this work, we analyze the difference between individual users and average users, proposing a new task called personalized image cropping (PIC). Compared to existing methods, the PIC model not only requires generic cropping features but also captures personalized preferences from a small set of user data. Nevertheless, as a typical few-shot learning problem, direct training is inherently challenging. To address this issue, we first construct a prior model that includes both user subjective features and image objective features. Considering the difficulty in explicitly modeling subjective features, we propose an individual preference module to embed user attributes into a continuous latent space and explore their cropping preferences. Then, the model is jointly trained on PIC tasks to integrate two features, facilitating the transfer of prior knowledge to evaluate user-specific cropping. Finally, by fine-tuning with a small set of annotations, we can obtain the individualized PIC model for each user. Extensive quantitative and qualitative experiments demonstrate the necessity of the PIC task and the effectiveness of our proposed method.