| In the information age,with the rapid development of artificial intelligence technology and the wide popularization of portable devices such as smart phones,the demand for mobile applications has greatly promoted the research of energy-efficient computing systems.As the mainstream deep learning algorithms,convolutional neural networks(CNN)are widely used in image classification,target detection,natural language processing and big data analysis.Most of CNNs are large in size and have high computational complexity,which are usually deployed on the server side and cooperate with graphics processing unit(GPU)acceleration to run effectively.It can’t meet the application needs where the memory space and power are limited,so that there is still a huge technical challenge for the storage and calculation of the CNN models on mobile end devices and embedded systems.Therefore,this thesis selects the lightweight CNN with low computational cost as the main research content.The specific works are as follows:(1)A design method of high-performance lightweight CNN models for low resolution image classification is proposed.With the application requirements of mobile terminal equipment for visual image processing in low resolution scene,a lightweight model called Low-res Mobile Net is constructed based on Mobile Net V2.The structure of model is simplified which can adapt to the input of low-resolution feature map.The Inception structure is used to fill the depthwise convolution and extract richer lowresolution image features.The activation function of dimension upgrading process is replaced to avoid information loss and improve the classification accuracy.The inter layer connection structure is adopted to strengthen the inter layer feature information fusion.Also,gradually decreasing expansion factors are used to reduce the scale of the model and remove the redundant structure.The experimental results show that the Lowres Mobile Net model achieves excellent low-resolution image classification performance with very few parameters and calculations.(2)A design method of multi branch lightweight CNN model based on Big-Little Net is proposed.Aiming at the problem that deep CNN is difficult to deploy to mobile terminal equipment with limited resources,an efficient and simplified lightweight CNN model called Mobile＿BLNet is constructed based on multi branch network architecture.The model puts forward the concept of modular design,and the features with different resolution can be recognized through module stacking.Structural branches with different computational complexity are designed to effectively obtain multi-scale features and reduce the amount of calculation.The depthwise separable convolution and inverse residual structure are used to reduce the size of the model and save a lot of computing resources.The experimental results show that the Mobile＿BLNet is simple and has excellent performance,which achieves a good balance between model scale,amount of calculation and classification accuracy.(3)Aiming at the low compression efficiency of lightweight model Mobile＿BLNet,a model compression and reconstruction method based on pruning is proposed.In this method,the channel pruning operation is used to compress the scale of the model,which can automatically identify the channels with low contribution to the model and cut the structure.In the pruning process,the total ratio method is used to determine the threshold,which improves the classification accuracy under the same compression effect.The model is reconstructed based on the pruning condition,and the model structure is simplified under the condition of ensuring the classification performance,which can further reduce the number of parameters and calculation required by the model.The experimental results show that the model pruning and reconstruction method proposed in this thesis can significantly reduce the cost of parameters and calculation required by Mobile＿BLNet while maintaining the network performance.The lightweight level of the model is further improved. |
PDF Full Text Request