Objective To improve the efficiency of recognition for banana ripeness.Methods A method for recognizing banana ripeness is developed based on improved YOLOv11n. A modified polarized self-attention mechanism is introduced into YOLOv11n to enhance the feature extraction capability of the backbone network across various banana distribution scenarios. The original upsampling is replaced with a module of content-aware reassembly of features, which enlarges the receptive field to more effectively aggregate contextual information. Scylla intersection over union (SIoU) is adopted as the new bounding box loss, which calculates the vector angle between ground truth and predicted boxes to better address the matching problem between them and reduce instances of missed and false detection.Results The improved method achieves increases of 1.4% and 3.0% in mean Average Precision 0.50 (mAP_0.50) and mean Average Precision 0.50~0.95 (mAP_0.50~0.95), respectively, with the recognition accuracy surpassing other existing methods.Conclusion The proposed method effectively enhances the accuracy and efficiency of recognition for banana ripeness, demonstrating high practical value.