Image Classification Based On Lightweight Bidirectional Ghost Network

With the development of deep neural network structure,image classification task has obtained more and more accurate results,but the network models widely used in image classification currently have the problem of large number of parameters and computation.With the popularization of embedded devices and mobile devices,the design and research of lightweight network model has become a trend.Lightweight networks aim to reduce computational redundancy,which can be specifically divided into spatial redundancy and channel redundancy.Among them,spatial redundancy refers to the information in the spatial dimension of the feature map obtained by the network that does not play a role in classification.Channel redundancy refers to the existence of many highly similar or repetitive information in the channel dimension of the feature graph.Both of the above lead to the increase of network computing.In order to solve the redundancy problem,many lightweight network models have been proposed.However,these models often have the following problems: on the one hand,the existing lightweight network model is not ideal in the space or channel direction of redundancy modeling,and can not balance the relationship between the number of model parameters and the accuracy;On the other hand,these methods only start from the redundancy characteristics of a single space or channel direction,without considering both at the same time.Based on the above problems,this paper proposes Bidirectional GhostNetwork(Bi-GhostNet),which is mainly composed of bidirectional phantom modules based on space redundancy and channel redundancy.The main work of this paper includes:(1)Study on feature map block technology based on learnable sampling.Firstly,a learnable and differentiable sampling module based on Gumbel Softmax distribution is designed to generate a block mask of the feature map so that the model pays more attention to the important information blocks in the feature map.At the same time,in order to make better use of the excellent parallel computing capability of the GPU,the feature graphs are divided and stacked according to the block size.The important information blocks cost more computing resources,while the non-important information blocks are operated cheaply by linear operation.(2)Research on de-redundancy feature extraction algorithm based on both spatial and channel directions.In this paper,a bidirectional phantom module is designed to eliminate the redundancy of features in both spatial and channel directions.In terms of space,by processing important information and non-important information respectively,the network focuses on important information in the direction of space.In the aspect of channels,a series of low-cost linear transformation operations are carried out on the feature graphs of important channels to obtain the output feature graphs of samples,thus reducing the amount of computation and parameters.(3)Comparative and ablation experiments were carried out on the Bi-GhostNet proposed in this paper in CIFAR-10 data set and ImageNet data set.In the CIFAR-10 data set,the VGG-16 network with the bidirectional phantom module inserted achieved the same accuracy as the VGG-16 network with 49% reduction in the number of parameters and 61% reduction in the amount of computation.In ImageNet data set,when compared with networks of the same level of computational complexity,Bi-GhostNet can also achieve the optimal Top-1 accuracy and Top-5 accuracy with a lower number of parameters and computations.The ablation experiments of Bi-GhostNet were conducted in the CIFAR-10 dataset,which mainly included the experiments on the internal hyperparameters of the spatial orientation module and the effectiveness of the bidirectional phantom module.The above experiments prove that BiGhostNet reduces the number of parameters and the amount of computation compared with the previous methods,and can run smoothly in the equipment with limited resources.