TANet-AVA

Model Card: TANet-AVA (Theme-Aware Network)

Model Details

• Architecture: Theme-Aware Network (TANet)
• Task: Image Aesthetics Assessment (IAA)
• Training Dataset: AVA (Aesthetic Visual Analysis)
• Original Authors: Shuai He et al.
• Source: The weights and architecture are ported from the official TANet GitHub repository for streamlined accessibility on Hugging Face.

Intended Use

TANet is designed to computationally assess the aesthetic quality of images. It addresses the inherent challenge of visual attention dispersion by adaptively extracting theme information from an image and applying theme-specific perception rules. It is suited for applications such as computational photography, automated image curation, and recommendation systems.

Architecture Highlights

TANet operates using a specialized multi-branch architecture to capture complex aesthetic criteria:

• Theme Perception: An explicitly trained branch extracts the visual theme of the image, allowing the network to self-adapt and establish context-aware rules for aesthetic evaluation.
• RGB-distribution-aware Attention (RGBNet): A custom module designed to help the network perceive color distributions and relationships within the RGB space, resolving complexities associated with standard spatial attention mechanisms.

Training Data

This specific model checkpoint was trained and evaluated on the AVA (Aesthetic Visual Analysis) dataset, a standard large-scale benchmark for image aesthetic assessment containing images with dense aesthetic score distributions and varied photographic styles.

Citation and Attribution

If you use this model in your research or applications, please cite the original authors and their paper:

Original TANet Code: https://github.com/woshidandan/TANet-image-aesthetics-and-quality-assessment Paper: https://www.ijcai.org/proceedings/2022/132

@inproceedings{he2022rethinking,
  title={Rethinking Image Aesthetics Assessment: Models, Datasets and Benchmarks},
  author={He, Shuai and others},
  booktitle={Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22},
  year={2022}
}