Based on LAMP and Channel Pruning and Knowledge Distillation Model Algorithm

Abstract

To solve the problem of limited computing power of mobile devices for tea picking robots, this study proposed a hybrid compression algorithm based on LAMP pruning and channel pruning, and combined with knowledge distillation technology to optimize the performance of the lightweight model. The LAMP-adaptive normalization strategy of LAMP pruning was used to balance the proportion of pruning at different levels to avoid the over-pruning problem of traditional amplitude pruning. The channel pruning is further introduced to eliminate redundant convolution kernels, reducing the computational effort by 37.3% and the parameter count by 20.3%. The Mimic loss and linear attenuation distillation strategies were proposed, and the reasoning speed was increased by 19.7% at the pruning rate 1.7 through weighted multi-layer feature difference and dynamic adjustment of loss weight. Experiments show that the proposed method reduces the average delay to 8.26ms while maintaining the accuracy of the model, which is significantly better than the traditional pruning methods such as L1 and group_taylor, and provides an efficient solution for real-time tea recognition on edge devices.