Research And Implementation Of Deeplearning Model Optimization Based On Network Compression

In recent years,deep convolutional neural networks(CNNs)have performed well in the field of computer vision.At the same time,the demand for engineering applications of artificial intelligence systems such as computer vision has also increased,and the research on neural network compression and acceleration methods has become increasingly active.However,due to its over-parameterized design,the convolutional neural network has led to a huge amount of parameters and expensive computational consumption of the network model,making it very difficult to deploy the model to some resource-constrained devices,such as embedded devices,Mobile phones,etc.How to deploy a highly available and highly reliable network model on resource-constrained devices has become the main bottleneck for the widespread application of artificial intelligence in real life.The research topic of this thesis is the compression and acceleration algorithms of deep convolutional neural network models,focusing on the compression and acceleration algorithms based on filter pruning,model reconstruction after pruning,and model mobile deployment.The main research contents are as follows:1.Research a filter pruning algorithm based on the similarity of feature maps.Different from the traditional filter pruning algorithm,this paper does not use the amount of information contained in the filter as the evaluation criteria for filter selection,but analyzes the similarity of the output feature map of each layer of convolutional layer,and uses spatial distance to calculate the filter.Therefore,the filter that can be replaced most is selected to cut off.2.Combined with lightweight network design,a network reconstruction algorithm after pruning is proposed to restore model capacity.The traditional pruning algorithm directly performs fine-tuning and retraining to restore the accuracy of the model after pruning,but it is often difficult to achieve the original accuracy of the model.This paper believes that the main reason is that pruning damages the original capacity of the model.This paper proposes a reconstruction algorithm based on the remaining feature maps after pruning,using lightweight operations to generate new feature maps based on the remaining feature maps and splice back to the original output feature map to restore the original capacity of the model.As shown in the experimental results,the pruning algorithm in this paper can maintain the accuracy of the original model or even improve the accuracy while reducing the amount of parameters and calculations.Compared with some excellent pruning algorithms,it shows excellent performance.3.Android mobile deployment based on the optimized model of pruning.This paper uses the deep learning mobile terminal inference framework Pytorch Mobile and the Android application development framework Android Studio to deploy the model to the Android mobile terminal,compares the performance of the compressed and optimized model with the original model,verifies the effectiveness of the algorithm in this paper,and completes the prototype implementation of a highly available,high-precision,and high-security offline image recognition system and testing.