Depthwise Separable Convolutional Neural Network Structure Optimization For Embedded Systems

With the good effect of convolutional neural network and the development of AIoT,the demand for its application to be grounded is also increasing.In the face of some embedded systems with high requirements on computing power,storage and real-time,the huge number of parameters and computation of convolutional neural network models are often difficult to be deployed and applied.In this paper,we study a variety of typical lightweight model structures and some model compression methods,analyze and summarize the design principles and innovation points based on these model structures.Based on these theoretical foundations,we propose an optimization method for the deep separable convolutional network structure,which can further accelerate the model computation speed at the embedded end,taking into account the characteristics of embedded devices.First,the optimization method proposed in this paper addresses the problems of memory access overhead and the number of high parameters of deep separable convolution.The method uses grouped convolution in deep separable convolution to reduce the number of parameters,and proposes an optimization method for the memory access overhead of grouped convolution.It is experimentally verified that the speed of grouped convolution is significantly improved after optimizing the memory overhead for the same amount of computation(FLOPs).Based on this,a group-level pruning method is proposed to prune the convolutional kernels of grouped convolutions,which significantly reduces the number of parameters and computation of deeply separable convolutions.The optimized depth-separable convolution is experimentally analyzed and verified to have a significant improvement in computational speed at the embedded device side.Second,an optimization method is proposed for the network structure to address the boundary problem of the feature map channel information of the grouped convolution.The optimization method proposes alternate convolution for the optimized depth-separable network based on the optimization,so that the channel information of feature maps between groups is passed in different layers.The structure-optimized model is subjected to accuracy experiments on MNIST and CIFAR-10 datasets,and it is verified that the optimization method proposed in this paper can reduce the number of parameters and the computational effort of the model to a large extent with a small loss of model accuracy.Then,we propose an embedded end-computation optimization method to address the memory handling overhead caused by the use of split and concatenate operations in the optimized structure.The optimization method reduces the operation overhead of the optimized network structure in terms of memory handling on the embedded side by making the input and output feature maps of the 1×1 convolutional layer computed in situ,and further improves the operation speed on embedded devices.Finally,we experimentally compare the actual speedup of the optimized model on embedded devices with the optimization method proposed in this paper.Finally,to address the feasibility and effectiveness of the model optimization method proposed in this paper in practical engineering applications,an embedded application system for keyword recognition is implemented for engineering verification.The system uses Deep Separable Convolutional Network(DS-CNN)for speech keyword recognition and optimizes it using the optimization method proposed in this paper.The feasibility of this paper's approach in engineering is verified by performing requirement analysis and system design for the software and completing an embedded application example for speech keyword recognition.