Enhancing the Performance of YOLOv9t Through a Knowledge Distillation Approach for Real-Time Detection of Bloomed Damask Roses in the Field

Harvesting Damask roses by hand is especially challenging because of the thorns on their stems, which not only complicate the process but also pose a risk of injury to workers. This problem highlights the need for automation solutions to facilitate the harvesting process. To carry out agricultural operation, particularly for picking fully bloomed Damask roses, using harvesting robots offers significant potential to reduce labor costs while simultaneously improving crop quality. Recent developments in deep learning algorithms, especially in convolutional models, have shown significant promise for object detection, highlighting strong possibilities for improving the efficiency of this process. The substantial computational demands and processing times associated with many deep learning models present a significant obstacle to their implementation in real-time applications. To address this challenge, knowledge distillation (KD) has emerged as a valuable model compression technique. This approach enables complex "teacher" models to pass essential insights to more streamlined "student" models, making them more suitable for immediate, real-world applications. In this study, we trained YOLOv9t model as a student model and YOLOv9c model as a teacher model. To facilitate this learning, two different techniques including online distillation (OD) and offline distillation (OFD) were explored. The results demonstrated that applying both online and offline KD increased the mAP0.5 of YOLOv9t by 0.3% and 0.2%, respectively, and boosted the detection speed by 5.1 and 1.8 frames per second (FPS), respectively. The results showed that the YOLOv9t model, trained as a student with both OD and OFD methods, performed better than the YOLOv9t model. This distilled version of YOLOv9t shows strong potential as an effective model for real-time detection of fully bloomed Damask roses.