To address the issues of low accuracy, false positives, and missed detections in conventional blood cell detection methods, we propose TBV-YOLO, an improved blood cell detection algorithm based on the YOLOv7 network. This algorithm incorporates our lightweight feature extraction module, BP-ELAN. By employing bi-level routing attention (BRA) and task-specific context decoupling (TSCODE), TBV-YOLO enhances the extraction of key local features and introduces richer semantic information, thereby improving the detection accuracy for dense and small targets. Additionally, the use of partial convolution (PConv) and the VoVGSCSP lightweight feature fusion module reduces computational complexity, further lightening the network model. Experimental results on the blood cell count and detection (BCCD) dataset demonstrate that the proposed model achieves a mean average precision at intersection-over-union (IoU) of 0.5 (mAP@0.5) of 93.9%, precision of 86.2%, and recall of 94.1%, representing improvements of 2.6%, 1.8%, and 5.7% over YOLOv7, respectively, with a model parameter size of only 8.8M.