Research On Image Scene Recognition Method Based On Deep Learning

Image scene recognition is a basic research in the field of computer vision.The scene is rich in semantic information,which can provide support for target detection,motion recognition,automatic driving and other visual tasks.It is also widely used in navigation and positioning,the vision of unmanned aerial vehicle vision and other fields.Due to the diversity of scene,it is difficult for traditional manual features to fully express the hidden information of scene images.However,in recent years,deep learning has achieved remarkable success in the field of computer vision.Therefore,this paper adopts the deep learning method that can automatically extract image features to conduct research on image scene recognition.The main contents include:This paper makes a comparative study of different image classification tasks,and mainly analyzes the differences in data sets used for object classification and scene recognition.Firstly,the image samples from the two data sets were compared,and it was found that the number and category of objects in the two data sets were significantly different.Then,through the statistical analysis of the number and size of the objects in the image,it is found that there are still differences in the size changes of the objects in the two data sets.Finally,the influence of these differences on scene recognition task is explored by using the method of deconvolution calculation class activation diagram.In view of the difference between scene recognition task and object classification task,a scene recognition method based on multi-level feature integration of convolutional neural network is proposed in this paper.The purpose is to reduce the influence of the number scale change of objects in the scene on the training and learning of convolutional neural network and improve the accuracy and generalization performance of the network.The main work of this method is to design features of different levels,corresponding to local details and global scene descriptions in the image respectively,and improve AlexNet network model to enable it to extract multilevel features of the design.Finally,the features of different levels were integrated and verified experimentally on the relevant data sets.This paper also proposes a scenario recognition method based on Inception structure optimization.Based on the extraction of multi-level features,this method adopts global mean pooling instead of the full connection layer,which reduces network parameters and improves the training speed of the network.In addition,this paper studies the feature integration method under different scene categories,with using fscore and confusion matrix to measure the classification effect of network for different scene categories,and finally proposes a classification confidence network based on Inception structure,so as to improve the accuracy of multi-category scene recognition.The proposed scene recognition based on deep learning in this paper has been evaluated on three public datasets.Experimental results demonstrate that the proposed algorithm outperforms most of state-of-the-art methods,it can effectively improve the accuracy of scene recognition.