| In recent years,with the support of big data sources and high-capacity,highperformance computing devices,deep convolutional neural networks(CNNs)have made significant progress.However,the remarkable performance of neural networks is often accompanied by high computing and storage costs,which poses challenges for their application on resource-limited devices.To facilitate the deployment and application of CNNs on such devices,this paper studies two model compression algorithms for compressing deep CNN models.(1)A fusion pruning algorithm based on information entropy layering.Firstly,this method finds similar filters by affinity propagation clustering and removes redundant parts.Secondly,it further prunes channels by using information entropy layering and batch normalization layer scaling factors.The combination of the two methods compresses the network model,and fine-tuning is conducted to reduce the size of the network without sacrificing accuracy.Experiments were performed on VGG16 and ResNet56 networks using the CIFAR-10 dataset.The results show that in VGG16,compared with the original model,the proposed algorithm reduces the number of parameters by 90.69% and computing cost by 24.46%.In ResNet56,63.53% of the parameters were removed,resulting in a 63.82% reduction in computational cost.The new model has better memory usage and computation speed than the baseline model while maintaining high network accuracy.Compared with similar algorithms,this method has advantages in terms of computation speed and model size.(2)A pruning algorithm combining attention mechanism.Firstly,one basic training is performed,and then the network baseline model is retrained based on the attention mechanism.After that,the model is pruned based on channel correlation,and continuously iterated until obtaining a simplified model with minimal accuracy difference from the baseline model.This algorithm was tested on ResNet with different datasets,and the results showed that it had strong adaptability to different datasets and network structures.On the CIFAR-100 dataset,ResNet50 was trimmed,which reduced the number of model parameters by 80.3% and computation cost by 69.4%,while preserving accuracy.On the ImageNet dataset,the parameter volume of ResNet50 was compressed by 2.49 times,and the computation volume was compressed by 3.01 times.The parameter volume of ResNet101 was reduced by 61.2%,and the computational cost was reduced by 68.5%.Compared with traditional fixed threshold methods,this approach achieved better results in detection accuracy,compression efficiency,and inference speed.In summary,the proposed method for compressing CNNs aims to reduce the number of parameters and improve computing efficiency.Experimental verification and analysis demonstrate that the proposed method is effective and applicable for at least two different CNN models,indicating its feasibility and usefulness.The algorithm-compressed models were also deployed on resource-limited devices and showed good performance through experimental verification. |
