Authors - Seema Shrawne, Aishwarya Jadhav, Saniya Gupte, Sayali Panch, Sakshi Thombre, Vaibhav Dhore, Vijay Sambhe Abstract - Accurately detecting and counting fruits is crucial for improving yield estimates in Precision Agriculture. However, challenges such as varying fruit types, overlapping objects, and cluttered backgrounds make this task difficult. In this study, we use RetinaNet, a deep learning model known for detecting dense objects, to address these issues. By replacing its ResNet50 backbone with SEResNet18, we improve both accuracy and speed, increasing mAP by 1.65%at IoU 0.5 and 1.79%at IoU 0.7. We also fine-tune the model using Differential Evolution to optimize anchors, further boosting mAP by 0.61% and 7.3% at IoU 0.5 and 0.7, respectively. Our enhanced model performs better on KFuji apples, mangoes, and holly fruits, while the default model slightly outperforms ours on strawberries and oranges at higher IoU thresholds.
