关键词: 模型压缩, 卷积神经网络, 模型剪枝, 知识蒸馏, 模型量化

Abstract:

Convolutional neural networks require substantial computational and storage resources in practical applications,which renders model compression essential for efficient deployment.However,single compression techniques often lead to performance degradation,reduced generalization ability,or increased computational complexity.To address these issues,this paper proposes a compression framework that integrates model pruning,knowledge distillation,and model quantization.It starts with sparsity training to prune redundant channels,continues by leveraging the original model as a teacher to distill knowledge into the pruned student network,and concludes with quantization training to further optimize the compressed model’s usability.The proposed method was tested on classification and object detection models in convolutional networks.Experimental results demonstrate that the framework effectively reduces storage and computation requirements.Model sizes are reduced by over 90%,inference speeds increases by 3-4 times,and accuracy loss is controlled within 2%.The proposed multi-method compression framework ensures the performance of convolutional neural networks while reducing the model size and improving the inference speed,and is suitable for efficient deployment in resource-constrained environments.

Key words: model compression, convolutional neural networks, model pruning, knowledge distillation, model quantization