Research On Technology Of Model Compression For Convolutional Neural Networks

Convolutional neural networks are everywhere,their applications cover computer vision,speech recognition and natural language processing.Convolutional neural networks require a lot of computation,take up much memory and are difficult to be deployed on embedded systems which greatly limit their applications.This paper takes the object detection algorithm as an example,proposes a model compression method which can help the compressed model detects objects in real time on the embedded side.Since the main compression strategy is for the convolutional layers,this compression method can be extended to other computer vision algorithms which use a large number of convolutional layers.Firstly,this paper replaces convolutional layers which contain a lot of parameters and computations with depthwise separable convolutions.This method can greatly reduce the parameters and computations while ensuring the performance of the model.Secondly,this paper proposes a dynamic pruning method which is adaptive to step size.This method can dynamically determine the pruning step size and order which can greatly improve pruning efficiency according to the sensitivity of each convolutional layer.And if the model is retrained with the pruned structure,the model will perform better than before.Finally,this paper simulates the operation time of each layer in the heterogeneous computation platform ZCU102 and proposes a model compression strategy for different platforms.According to the experimental results,the object detection algorithm using the compression method proposed in this paper performs well on the PASCAL VOC dataset.The model size has decreased from 33.26 M to 9.7M,the final mAP has increased from 54 to 58.In the TX2 embedded system,the FPS increases from 7.23 to 9.56,which basically meets the requirements of real-time detection.Therefore,the model compression method proposed in this paper can effectively solve the problem of deploying real-time object detection algorithms on the embedded side.And the method can be extended to other Computer vision algorithms which contain a large number of convolutional layers.