Reproduce “CST-YOLO: A Novel Method for Blood Cell Detection Based on Improved YOLOv7 and CNN-Swin Transformer” on OpenResearch

About this paper

Ming Kang, Chee-Ming Ting, Fung Fung Ting, Raphaël Phan

Researchers at Monash University Malaysia Campus developed CST-YOLO, an object detection model combining an improved YOLOv7 with a CNN-Swin Transformer fusion, achieving enhanced precision for small-scale object detection, particularly for blood cells. The model demonstrated [email protected] scores of 92.7% on BCCD, 95.6% on CBC, and 91.1% on BCD datasets, outperforming baseline YOLOv7 by up to 3.7%.

Key insights

• A synergistic fusion of CNNs (for local features) and Swin Transformers (for global context) significantly improves small-scale object detection compared to using either architecture in isolation.
• Dynamically weighting feature maps during aggregation (W-ELAN) and employing channel attention (MCS) are effective strategies for enhancing feature information flow and multi-scale perception.
• The CNN-Swin Transformer (CST) module is particularly crucial for improving detection accuracy of the smallest objects, like platelets, due to its ability to combine different receptive fields.

Summary by alphaXiv.