Image Classification Based On Lightweight And Multi-scale Attention Fusion

Image classification is a fundamental task in computer vision.With the rapid development of Internet and artificial intelligence technology,a large number of image data is generated everyday.Image classification technology is applied to many scenarios in life and work,therefore many Internet companies and research institutions focus their researches on image classification.At present,image classification based on deep learning algorithms has become the mainstream,classification performance is improved at the cost of more model parameters.How to reduce the amount of parameters while maintaining high classification accuracy is a challenging issue.Therefore,this paper proposes a lightweight split convolution,an hourglass module and a multi-scale attention fusion model.The main contributions are as follows:(1)A deep learning image classification method based on a novel lightweight split convolution is proposed.In order to reduce feature redundancy in the standard convolution process,this paper proposes a lightweight split convolution,which is integrated into the deep neural network for features extraction.The input feature maps are splited into two parts: characteristic features and self-calibration features then the two parts pass the refinement and the lightweight process respectively,which effectively extracts higher-level feature information and reduces feature loss.In addition,the two parts are combined with a non-parametric module to enhance the features.The experiments validate that this convolution improves the classification accuracy and reduces the parameter quantity.(2)A deep learning image classification method based on a novel lightweight hourglass module is proposed.In order to reduce the information loss in the process of dimension reduction and the parameters of deep convolution networks,this paper proposes a lightweight hourglass module,which is integrated with the deep neural network to obtain effective features.Channel transformation in hourglass module consists of two branches:identity mapping and linear transformation operation.Through two-stage channel transformation,dimension is reduced and increased to extract characteristic features.The features after dimension increase are integrated with input information,which reduces the feature loss.Linear activation function is introduced after the dimension reduction to keep more useful information.The experimental results show that the hourglass module improves the performance of the model and reduces the parameters.(3)A deep learning image classification method based on multi-scale attention fusion is proposed.Aiming at the problem that attention extraction methods based on global feature information ignore the function of local features,this paper proposes a multi-scale attention fusion model.In order to establish the interdependence between multi-scale channels,the spatial attention is integrated into the image features extracted by deep neural networks.The attention fusion model uses two branches to obtain the global attention weight and local attention weight of the feature map,and the two weights are aggregated as the weight of the input feature map.In the process of local attention extraction,the receptive field is increased by dilated convolution to enhance the correlation of local attention.The experiments show that the fusion model can improve the accuracy of image classification when embedded into different backbone networks.