Efficient Convolutional Neural Networks on Raspberry Pi: Enhancing Performance with Pruning and Quantization

Abstract In the evolving realm of edge computing, adapting Convolutional Neural Networks (CNNs) for deployment on resource-limited devices like the Raspberry Pi presents significant challenges. This study delves into the effectiveness of pruning techniques and TensorFlow Lite conversion in optimizing CNN models for such platforms, focusing on widely-used datasets such as MNIST, SVHN, 'Horses or Humans', 'Cats vs Dogs', and 'TF Flowers'. Our comprehensive analysis reveals that pruning effectively maintains model accuracy up to a certain level of sparsity, particularly in models like MNIST, where only a marginal decrease in accuracy was observed with increased sparsity. This underscores the delicate balance between model complexity and performance. Moreover, our findings demonstrate notable improvements in inference times following the application of pruning and quantization techniques. This was especially evident in the significant reduction of inference time for the MNIST model, highlighting the feasibility of deploying more efficient CNN models on the Raspberry Pi. These optimizations not only enhance the practicality of using CNNs in resource-constrained environments but also extend their applicability in real-world scenarios. Our study contributes important insights into model optimization for edge devices and lays a foundation for future research, emphasizing the need for tailored approaches in neural network optimization for edge computing.