Comparison of Image Processing Models in Quark Gluon Jet Classification

TL;DR

The study systematically compares CNN, ViT, and Swin Transformer architectures for quark–gluon jet classification using three-channel jet images from Pythia8, revealing that Swin-Tiny with fine-tuning of the last two blocks delivers the best efficiency–accuracy balance (ACC about $81.4\%$ and AUC about $88.9\%$). Self-supervised MoCo pretraining improves feature robustness, particularly in data-scarce scenarios, and a custom, compact Swin model achieves comparable performance with far fewer parameters and reduced training time. The results highlight the value of hierarchical attention models for jet substructure and suggest strong potential for domain transfer to real collision data. These insights have practical implications for efficient, scalable jet analyses in high-energy physics and for future domain-adaptation studies.

Abstract

We present a comprehensive comparison of convolutional and transformer-based models for distinguishing quark and gluon jets using simulated jet images from Pythia 8. By encoding jet substructure into a three-channel representation of particle kinematics, we evaluate the performance of convolutional neural networks (CNNs), Vision Transformers (ViTs), and Swin Transformers (Swin-Tiny) under both supervised and self-supervised learning setups. Our results show that fine-tuning only the final two transformer blocks of the Swin-Tiny model achieves the best trade-off between efficiency and accuracy, reaching 81.4% accuracy and an AUC (area under the ROC curve) of 88.9%. Self-supervised pretraining with Momentum Contrast (MoCo) further enhances feature robustness and reduces the number of trainable parameters. These findings highlight the potential of hierarchical attention-based models for jet substructure studies and for domain transfer to real collision data.

Figures (10)

• Figure 1: (a) Quark jet and (b) gluon jet. Gluon jets exhibit a broader spatial radiation pattern and higher particle multiplicity compared to quark jets. For visualization purposes, the pixel display size is enlarged relative to the actual input resolution ($72 \times 72$).
• Figure 2: Simplified schematic of the self-attention mechanism. A $4 \times 4$ image is split into $2 \times 2$ patches, which are projected into Query, Key and Value embeddings. Attention score are computed and applied to obtain the output representation for a single attention head. The class token is omitted for clarity.
• Figure 3: ViT employs globl self-attention with a flat stack of Transformer blocks, whereas the Swin Transformer uses window-based local self-attention and shifted windows to enable cross-window interactions. Patch merging between stages further builds a hierarchical multi-scale representation.
• Figure 4: Illustration of the Momentum Contrast (MoCo) framework for jet images. Different views of the same jet image form positive pairs ($q$, $k^{+}$), while views from different images form negative pairs ($q$, $k^{-}$). By maximizing similarity between positive pairs and minimizing similarity to negative pairs, MoCo encourages the encoder to learn discriminative representations for visually similar jet images.
• Figure 5: Overview of the block-wise fine-tuning strategy applied to the pretrained ViT-Tiny and Swin-Tiny models.The number of unfrozen transformer blocks is progressively increased from the classification head, while the remaining blocks are kept frozen.