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