Multi-Level Transitional Contrast Learning for Personalized Image Aesthetics Assessment

Introduction：

PyTorch implementation for the paper

Model weight：(Hugging Face) (Baidu Netdisk)

Inference Guide：

1. Overview

This guide will help you get started with the MTCL inference code.

2. Model Architecture

MTCL consists of three main components:

**GIAA Model**: General Image Aesthetic Assessment backbone (ResNet-50 based)
**Contrast Model**: Contrastive learning encoder for personalized features
**PIAA Model**: Fusion of GIAA and Contrast features with personalized regression head

3. Directory Structure

project_root/
├── code/
│   ├── GIAA/
│   │   └── train_GIAA_model.py       # GIAA model definition
│   ├── MTCL/
│   │   └── Contrast_Database         # Contrast data for training
│   │   └── FlickrAES_TrainUser       # Train user of FlickrAES
│   │   └── train_Contrast_model.py   # Contrast model definition
│   └── PIAA/
│       └── ├── FlickrAES_PIAA/
            │       └── image/        # Flickr-AES images
            │       └── label/
            │           ├── test_worker.csv                   # Test Worker information
            │           └── image_labeled_by_each_worker.csv  # Image ratings by workers

            └── test_PIAA_model.py          # This inference script

4. Download Model Weight

Pre-trained PIAA Model: Place at

./model/ResNet50/ResNet50-FlickrAes-PIAA.pt
./model/ResNext101/ResNext101-FlickrAes-PIAA.pt

Flickr-AES Dataset:

Images: ./FlickerAes_PIAA/image/
Labels: ./FlickerAes_PIAA/label/

5. Running Inference

python test_PIAA_model.py

Citation

If you find our work is useful, pleaes cite the paper:

@article{yang2023multi,
  title={Multi-level transitional contrast learning for personalized image aesthetics assessment},
  author={Yang, Zhichao and Li, Leida and Yang, Yuzhe and Li, Yaqian and Lin, Weisi},
  journal={IEEE Transactions on Multimedia},
  volume={26},
  pages={1944--1956},
  year={2023},
  publisher={IEEE}
}