StyleConsensusLearning

Description

TCFL (Triangular/Diamond-shape Consensus Feature Learning) — an artistic style classification framework based on community feature learning. It uses pretrained CNN backbones (VGG16 / ResNet50) for feature extraction, and a novel Style Learning AutoEncoder (SLAE) with a configurable triangle or diamond network topology to learn style-consensus features across same-class images. A lightweight linear classifier is trained on top.

---

Architecture

Image
  └─► CNN Backbone (VGG16 / ResNet50)          feature extraction
        └─► StyleLearningAutoEncoder (SLAE)     style consensus learning
              ├─ Triangle topology  (ae1→ae2→ae3)
              └─ Diamond topology   (DiamondStyleLearningCell, Gram-loss)
                    └─► Linear Classifier       top-1/top-3 prediction

Key modules:

File	Description
model_NEW.py	SLAE, CDAutoEncoder, StyleLearningCell1/2/3, DiamondStyleLearningCell
SCLdataSet_NEW.py	TCFLDataset — loads features / images, computes per-sample style levels
run_NEW.py	Training loop (feature mode or image mode)
run_test_NEW.py	Evaluation, batch prediction, dataset image filtering
style_predict.py	Single-image / folder-level style prediction
MlpNet.py	Optional MLP classifier
utils/CustomLoss.py	KL, Cosine, Cityblock, Chebyshev, DotProduct, Gram loss
utils/styleLevelCal.py	Centroid-based Euclidean style-level assignment
utils/trainTestSplit.py	Train / test label file generation
utils/trainTestValSplit.py	Train / val / test label file generation
utils/copy_file.py	Copy images to class sub-directories
utils/image_processing.py	Image I/O helpers

---

Style Level Assignment

Each sample is assigned a style level (0–3) according to its Euclidean distance from the class centroid:

Level	Distance range
3 (strongest)	< 10 % of range from centroid
2	10 %–30 %
1	30 %–50 %
0 (weakest)	≥ 50 %

The assigned style level determines which same-class nearest-neighbour feature is used as the consensus target for the SLAE cell during training.

---

Supported Datasets

Dataset	# Classes	Notes
Painting91	13	Classical painting styles
Pandora	12	Diverse artistic styles
WikiArt3	15	WikiArt subset
FashionStyle14	14	Fashion photography
AVAStyle	14	Aesthetic visual analysis
Arch	25	Architectural styles
webstyle	10	Web-UI design styles

---

Installation

python==3.9.18
pytorch==1.12.0
torchvision==0.13.0
pillow==10.0.1
cudatoolkit==11.3.1
scipy==1.11.4          # pip install scipy
tqdm==4.66.1           # pip install tqdm
matplotlib==3.8.2      # pip install matplotlib
opencv-python          # pip install opencv-python

---

Directory Setup

StyleConsensusLearning/
├── data/               # dataset images and label files
│   └── <DatasetName>/
│       ├── Images/
│       └── Labels/     # train.txt  test.txt  label.txt
├── features/           # extracted .npy feature files and .pkl style-set files
├── pretrainModels/     # pretrained VGG16 / ResNet50 weights
├── temp/               # checkpoint models saved during training
└── imgs/               # (optional) intermediate reconstructed images

All data/, features/, pretrainModels/, and backup/ files are available on Baidu YunPan.

---

Usage

1. Extract features (run once per dataset/backbone combination)

Edit the dataset and backbone settings in run_NEW.py, then set isImage=False and choose model_type.

2. Generate train / test label splits

python utils/trainTestSplit.py
# or
python utils/trainTestValSplit.py

3. Train

python run_NEW.py          # uses feature mode (isImage=False) by default

Key arguments in runStyleConsensusLearning():

Parameter	Default	Description
num_epochs	3000	Training epochs
batch_size	256	Batch size
isImage_	False	True for raw images, False for pre-extracted features
model_type	'resnet'	'vgg' or 'resnet'

Checkpoints are saved to temp/ whenever validation accuracy improves.

4. Evaluate

python run_test_NEW.py

5. Predict a single image

python style_predict.py

---

Experiment Results vs. SOTA

	Painting91	Pandora	WikiArt3	Arch	FashionStyle14	AVAStyle
VGG16	58.42	49.73	40.02	61.41	68.22	39.94
VGG19	58.11	46.44	39.93	60.11	66.14	40.02
ResNet50	64.93	51.65	47.01	65.12	71.13	40.05
ResNet101	65.50	52.61	46.11	66.42	70.00	47.02
InceptionV3	53.41	42.83	36.68	61.52	62.70	33.33
DAE	58.82	48.71	41.48	58.55	61.48	41.46
SAE	63.65	48.64	41.53	59.61	74.33	40.29
SSCAE	64.07	49.38	43.65	60.48	75.02	45.77
DDS	62.21	52.35	43.17	/	/	/
MCFFNet	66.60	51.39	45.51	66.12	68.38	42.69
STACLE	60.41	55.80	47.21	60.81	64.47	46.38
TCFL+VGG16 TOP1	67.39	56.67	47.85	65.57	71.67	47.22
TCFL+VGG16 TOP2	83.19	74.92	66.99	78.12	85.69	63.92
TCFL+VGG16 TOP3	92.27	84.52	78.22	85.05	91.11	75.74
TCFL+ResNet50 TOP1	69.12	56.98	51.62	69.03	77.17	53.94
TCFL+ResNet50 TOP2	85.29	76.15	69.11	81.82	87.44	68.68
TCFL+ResNet50 TOP3	92.27	84.52	78.22	85.05	91.10	75.74

---

Citation

If you find this work useful, please cite our paper:

@article{cui2025community,
  title     = {Community transferrable representation learning for image style classification},
  author    = {Cui, Jia, Shen Jinchen, Wei Jialin, Liu Shiyu, Ye Zhaojia, Luo Shijian, Qin Zhen},
  journal   = {ACM Transactions on Multimedia Computing, Communications and Applications},
  volume    = {21},
  number    = {6},
  pages     = {1--20},
  year      = {2025},
  publisher = {ACM}
}

Cui, Jia, et al. "Community transferrable representation learning for image style classification." ACM Transactions on Multimedia Computing, Communications and Applications 21.6 (2025): 1–20.