Repository files navigation

This project investigates the performance improvements possible by integrating an Inception-ResNet-based feature extraction module within the patch merging stage of the Swin Transformer architecture.

Motivation

• Swin Transformers employ linear embedding for patch merging. Could a more sophisticated feature extraction technique improve performance?
• Inception-ResNet modules excel at capturing features at multiple scales. This might enrich the representation learned during patch merging.

Modifications

• The original linear embedding layer in the Swin Transformer patch merging is replaced with an Inception-ResNet module.

Performance on Retinal OCT Dataset

• Our study significantly advances medical picture classification and opens the door for more developments.
• Interestingly, our improved Swin model outperforms other well-known models like Efficient-Swin, VGG16, ResNet18, and AlexNet.