A lightweight SOD-YOLOv5n model-based winter jujube detection and counting method deployed on Android is proposed. The model improves upon the YOLOv5n model by replacing strided convolution and pooling layers with SPD-Conv, optimizing the upsampling module with CARAFE, and using GSConv in the neck to reduce model size while maintaining accuracy. The SOD-YOLOv5n model achieves higher precision, recall, and mAP compared to YOLOv5n, with reductions in RMSE and MAPE. The model size is reduced to 3.64 MB using float16 quantization. The quantized model is used to develop the JujubeDetector app for Android, enabling real-time detection and counting of winter jujubes. Experimental results show the app achieves a root mean square error (RMSE) of 1.46, coefficient of determination (R²) of 0.97, and detection time of 30-90 ms. The model performs well in complex environments with low light and overlapping jujubes. The SOD-YOLOv5n model outperforms other models in accuracy and efficiency, making it suitable for Android deployment. The study concludes that the SOD-YOLOv5n model effectively improves winter jujube detection and counting, with practical applications for yield prediction.A lightweight SOD-YOLOv5n model-based winter jujube detection and counting method deployed on Android is proposed. The model improves upon the YOLOv5n model by replacing strided convolution and pooling layers with SPD-Conv, optimizing the upsampling module with CARAFE, and using GSConv in the neck to reduce model size while maintaining accuracy. The SOD-YOLOv5n model achieves higher precision, recall, and mAP compared to YOLOv5n, with reductions in RMSE and MAPE. The model size is reduced to 3.64 MB using float16 quantization. The quantized model is used to develop the JujubeDetector app for Android, enabling real-time detection and counting of winter jujubes. Experimental results show the app achieves a root mean square error (RMSE) of 1.46, coefficient of determination (R²) of 0.97, and detection time of 30-90 ms. The model performs well in complex environments with low light and overlapping jujubes. The SOD-YOLOv5n model outperforms other models in accuracy and efficiency, making it suitable for Android deployment. The study concludes that the SOD-YOLOv5n model effectively improves winter jujube detection and counting, with practical applications for yield prediction.