<p>Colorectal cancer commonly originates from polyps. Timely detection and removal of these polyps can significantly reduce mortality rates associated with colorectal cancer. Nevertheless, conventional medical imaging methods suffer from low accuracy and slow detection speeds, rendering them inadequate for practical applications. For the above reasons, we propose a real-time polyp detection model, YOLOv5-CE-DAFF, which is based on the YOLOv5 model and incorporates the CBAM-ECA attention mechanism and dropout-based adaptive feature fusion. Firstly, to pay more attention to polyp features, we incorporate the CBAM-ECA attention module into the local cross-layer fusion structure. Secondly, we design an adaptive feature fusion of the prediction feature layer based on dropout, which increases the number of high-quality prediction boxes. Thirdly, we improve the positive and negative sample matching strategy to accelerate model convergence and improve detection accuracy. Additionally, we combine Soft-NMS and WBF to replace the original NMS algorithm, which retains as many high-quality prediction boxes as possible. The model is tested on two colonoscopy datasets. Numerical experiments show that YOLOv5-CE-DAFF has high detection accuracy and a low missed detection rate for colon polyps. Compared with the original YOLOv5 model, YOLOv5-CE-DAFF significantly improves recognition accuracy and reduces missed detection rate.</p>