Knowledge distillation with group convolution mapping layer for plant pest classification

Abstract

The increasing demand for high-quality food requires efficient monitoring and control of greenhouse environments. One crucial aspect is the detection and classification of plant pests, which can be achieved using deep learning-based computer vision techniques. However, deploying these models on edge devices like Raspberry PI 4 in a stationary greenhouse setup poses significant challenges due to limited computational resources. To address this challenge, we employ knowledge distillation, a technique that transfers knowledge from a large teacher network to a smaller student network. The goal is to retain the accuracy of the teacher network while reducing the complexity of the student network.