To address the difficulty in identifying cucumber diseases caused by numerous varieties and similar symptoms, this study proposes an improved YOLOv9 model for cucumber disease recognition, named BiFEL-YOLOv9, to enhance detection accuracy in natural backgrounds. Initially, a weighted bidirectional feature pyramid network module is incorporated into critical network layers to enhance the model’s multi-scale feature fusion capability. Following that, the original RepNCSPELAN4 module is enhanced by integrating a feature enhancement module and a large selective kernel block (LSKBlock) to obtain the RNFEL module, which improves the model’s feature representation capability and robustness to complex backgrounds. Experimental results indicate that the BiFEL-YOLOv9 model achieves an accuracy of 97.96%, a recall rate of 95.51%, and mean average precision scores of 98.21% for mAP_0.5 and 95.12% for mAP_0.5:0.95, all of which surpass the performance of the original YOLOv9 model. The proposed model effectively accomplishes the detection and recognition of cucumber diseases.