An effective litchi detection method based on edge devices in a complex scene

Due to the random distribution and irregular growth of fruit clusters on trees, litchi mainly depends manual picking harvesting. Given this situation, visual-guidance automatic system is utilised in existing studies reduce time effectively improve economic benefits. However, methods are primarily based high-performance computing equipment laboratory. This equipment's volume energy consumption makes it challenging deploy, severely limiting application picking. study adopts portable, low-power edge devices proposes an effective detection method solve shortcomings mentioned above. First, images orchard were collected, data augmentation was adopted generalisation images. Then, model constructed detect litchi. Finally, trained compressed remove redundant parameters according channel pruning layer algorithms deploy devices. The knowledge distillation leveraged fine-tune model. experimental results demonstrated that proposed could construct a only 7.6 MB with compression rate 96.8%. can achieve average precision 95.3% recall 97.3% at 117.6 frames per second (FPS), 1.8 times original be easily deployed portable low-computing rapid accurate litchi, which great significance for biological systems.