Smartphone-based fruit counting for early capsicum yield estimation

Portable, accurate and low-cost fruit counting is critical for estimating early capsicum yields. On the basis of lightweight YOLOv8n deep learning model, an improved YOLOv8n model is proposed. Firstly, by integrating the Ghost module in the backbone and head, the number and complexity of model parameters can be effectively reduced. Then, by integrating the SKA attention mechanism in the C2f module, the model ability to capture spatial information is enhanced, and a good balance between computational efficiency and computational performance is realized. By replacing the Conv module with the ODConv module, the convolution kernel is dynamically adjusted, which effectively enhances the feature extraction ability of the model for the target region. By replacing the CIoU loss function with the EIoU loss function, the convergence speed in the model training process is accelerated and the model detection accuracy is improved. Finally, experiments are carried out based on the capsicum dataset under complex natural environment conditions to verify the generalization ability of the improved YOLOv8n model. Compared to the original YOLOv8s model, the improved Yolov8s model improves the precision, recall and mean average accuracy (mAP) by 2.4%, 4.8% and 3.9% respectively, with an average detection time of 5.69 milliseconds and a model parameter of only 5.3M. The experimental results show that the model proposed in this study has superior automatic counting performance, and can provide farmers with a fast, accurate and cost-effective tool for estimating early capsicum yield.