Research And Implementation Of Lightweight Convolution Neural Network Based On Mobile Terminal

In the field of image classification,object detection and human posture recognition,traditional convolution neural network has disadvantages such as high computational complexity and large model size,which make it difficult to deploy on mobile terminals.In order to overcome these difficulties,some researchers proposed a mobile-based convolution neural network.This network adopts a highly efficient convolution method to adapt mobile terminal equipment with limited computing capacity and limited storage space.Although in the field of image classification,convolution neural networks based on mobile terminals have achieved good results,there is still a problem of slow forward operations in the field of object detection and human posture recognition.In order to solve these above problems,this paper focuses on lightweight convolution neural networks based on mobile terminals.Firstly,in the field of object detection and human posture recognition,this paper reconstructs a group of lightweight convolution neural network architectures based on mobile terminals.The construction method is as follows: 1)Depthwise separable convolution is used to replace conventional convolution;Depthwise separable convolution is a lightweight convolution computing technique that separates conventional convolutions in the spatial dimension.This technique not only reduces the computational complexity of convolution operations,but also reduces the number of parameters.So it's a High-efficiency convolution method;2)Fusion of different hierarchical feature maps;Determined by the characteristics of the convolution neural network,different levels of feature maps express different information,so by merging different levels of feature map information,you can Effectively improve the neural network's expressive ability;3)Introduce the approximate linear layer;Because the traditional rectified linear unit does not respond to negative values,when the input data dimension is lower,the output data loss ratio is more serious,so this paper introduces an approximation Linear module,this module not only reduces the proportion of data information loss,but also ensures the neural network's nonlinear characteristics.4)Equivalent replacement of batch normalization layers;The operation mode of the batch normalization layer in the forward operation can be equivalently replaced by modifying the bias parameter of the related convolution layer.With such an alternative,the amount of computation of the convolution neural network can be reduced.Then,in order to further speed up the forward operation of the model,this paper designs and implements a parallel acceleration platform based on lightweight convolution neural network of mobile terminal.Its main work contents are as follows: 1)The forward operations of Convolution neural network are accelerated through multi-thread technology;Since the convolution calculation is very easy to parallelized and the number of processor cores in the current mobile terminal is constantly increasing,multi-core can be fully scheduled through multi-thread programming techniques.It will reduce convolution operation time-consuming and improve forward operation speed.2)Accelerate the forward operation of the convolution neural network by using single instruction multiple data technology;Single instruction multiple data technology is a set of instruction sets that process multiple data within a single instruction cycle.Through this technology,the data throughput of the processor can be effectively improved,the memory overhead can be reduced,and the forward operation time can be further reduced.Finally,according to experiments,the object detection algorithm based on VOC dataset reconstructed in this paper of m AP value is 2.3 higher than the Mobilenet V1-SSD algorithm proposed by Google in 2017,and the operation speed is improved about 5%.The object detection algorithm based on COCO dataset reconstructed in this paper is about15% faster than the Mobilenet V1-SSD algorithm.The m AP on the COCO2017 test-dev dataset is 0.1 higher than the Mobilenet V1-SSD algorithm.The human pose recognition algorithm reconstructed in this paper is 10 times faster than Open Pose algorithm,and the model size is reduced by nearly 7 times.Based on their excellent performance in mobile terminals,this group of lightweight convolutional neural network algorithms can be applied to intelligent robots or drones and other products.