Research On Image Classification Based On Multi-resolution Convolutional Neural Network

Since the development of convolutional neural networks,it has achieved some remarkable achievements in image classification.With the iterative update of hardware,the research and development of convolutional neural networks still have broad development prospects.Current research shows that the depth and accuracy of convolutional neural networks are still difficult to study.The classification accuracy of the deep network is high,but the complexity and verbosity will make the experiment more difficult,and the generalization of the shallow network is weak.In order to solve this problem,this paper focuses on optimizing the structure of convolutional neural network and reducing the complexity of the network.The main work includes:1.The optimization method of adding multi-resolution factors to the dense block structure is proposed,which improves the generalization ability of the output features of each layer and the efficiency of feature extraction,makes important features more prominent,and optimizes the problem of feature loss during feature extraction.Improve the integrity of future maps in the transmission process,and further improve the performance of the network model.Solved the problem of insufficient feature extraction and feature loss in the process of convolutional neural network extraction of features.2.An optimization scheme for adding a residual attention mechanism network between dense block structures is designed to enhance the salient features between dense block structures and improve the optimization effect of feature extraction.It solves the problem that current deep learning usually needs to be trained from scratch,and using Finetune to learn new tasks does not achieve the expected results.3.Using the characteristics of distillation learning,an optimization method of self-distillation is proposed,and the weight model of the network with deeper network layers is transplanted to the network with shallower network layers,so as to achieve the purpose of compressing the network.4.Application of multi-resolution network in image classification.Image classification technology has been gradually applied in industrial production and actual life.This part is applied to the recognition of road traffic signs.The features of the traffic sign pictures are extracted through preprocessing,and then the features are optimized using the optimized network structure mentioned above.The experiment and analysis show that the optimization algorithm proposed in this paper has better performance.