In this research, we present advancements in precision agriculture by improving fruit classification for automated harvesting through a novel deep learning approach, the Detail-Semantics Enhancement You Only Look Once (DSE-YOLO) model. This model was created to solve the problem of identifying fruits in situations where flora frequently blocks the view of fruits, irrespective of their size or growth stage. Existing methods were constrained by the requirement for manual feature design. The DSE-YOLO model combines substantial feature extraction with semantic information to improve fruit detection at various scales. We offer two loss functions: Double Enhanced Mean Square Error (DEMSE) and Exponentially Enhanced Binary Cross Entropy (EBCE), in addition to correcting class imbalances. Using a large collection of annotated fruit photos, our approach showed considerable gains in detection accuracy. The DSE-YOLO model is an extremely effective instrument for automating fruit picking, outperforming earlier techniques and greatly increasing the productivity and efficiency of modern agriculture.

Index Terms- Precision Agriculture, Fruit Classification, Automated Harvesting, Deep Learning, YOLOv5, DSE-YOLO, Multi-Stage Detection.