Deep Learning Based Image Recognition Algorithm For Mobile Devices

Deep Learning based model compression technology has been a popular research direction in recent years.Nowadays,deep learning technology is very hot.Due to strong learning ability and non-linear feature expression ability,the image recognition model based on deep learning is constantly approaching the limit accuracy of computer vision task.At the same time,the depth and size of the model are exponentially increasing.Especially in the vision application of the mobile devices,deep model undoubtedly poses a great challenge to the storage and the computing speed of the mobile devices.In this paper,we mainly focus on the model compression technology based on deep learning for the mobile devices.Deep learning model compression technology for mobile devices in this paper is divided into two mainly approaches:the compression of parameters based on a pre-trained model;designing a light-weight convolution neural network with small number of parameters and small amount of computation.In the aspect of parameter compression,this paper firstly studies the influence of different parameter prune strategies on the performance loss of the compressed model based on pre-training,and then proposes a method of non-linear segmented prune.Based on the sparse model,this paper then proposes a uniform parameter quantization method to further compress model.In the aspect of light-weight network,this paper proposes a CNN model with small amount of parameters and computation,as well as a good recognition performance.In particular,this paper first summarizes current methods of deep learning model compression for mobile devices,and then improves the compression method based on a pre-trained model.In this paper,the effect of different prune strategies on the loss of accuracy is compared.The result shows that compared with one-step prune,the iterative prune method has less compression accuracy loss.Based on this conclusion,this paper proposes a more concrete non-linear segmented prune strategy,which can further reduce the loss of accuracy by alleviating the variation of sparse ratio at the later stage.Then,based on the sparse model obtained by parameter pruning,an effective uniform quantization method is proposed in this paper,by selecting quantization centers at equal intervals within the effective parameters distribution range and then fixed center values during fine-tune training to update the parameter center index.Finally,this paper introduces the method of parameter sparse storage.This paper focuses on the latest large CNN model.Combined with the above method,the pre-trained Inception-v3 model is about 8 times compressed with an accuracy loss less than 1%,and the actual storage is only 12.8MB on the ARM platform.In the aspect of light-weight network,this paper designs the network from the light-weight structure and enhancing the performance,then introduces the depthwise separable convolution layer to reduce the computation and the channel weighted branches of local feature fusion to improve the performance of the model.Experiment shows that this light-weight CNN model's actual storage is 27 times smaller than VGG-16 model,and the computation speed is 8 times faster than VGG-16 on mobile ARM platform.