Research On Multi-feature Scene Recognition Based On Deep Learning

Scene recognition is one of the important research directions in the field of computer vision,and it has been widely used in many aspects.Due to the rapid growth of data volume,traditional algorithms have shown many shortcomings in solving the problem of large data volume scene recognition.With the proposed convolutional neural network(CNN),the scene recognition method based on CNN has also appeared.The feature image is extracted from the scene image through the trained CNN network.Based on this,various spatial global features are extracted to make the scene.The effect of recognition has been greatly improved.Although the CNN features have achieved encouraging performance,some of them only take advantage of the spatial features in the scene.While the other part utilizes the object features between scene categories,its features include the general features caused by common objects in different scenes,which weakens the ability to distinguish between scenes.Most existing scene representation methods take advantage of the features that make up an object in a separate scene between classes,ignoring the negative effects of ordinary objects in different scenes.The general characteristics of a generic object create a generality between different scenes,thereby weakening the discriminating properties between scenes.In order to solve the problem of intra-class difference and inter-class similarity in scene recognition,this paper proposes a scene recognition system based on double convolutional neural network and a scene recognition method based on subject model object discriminant features.A convolutional neural network model is designed and applied to scene recognition.The specific work is as follows:1.Most existing CNN-based scene recognition only considers the global spatial features in the scene and ignores the local object features.At present,the recognition rate of large-scale scene recognition is still not high.This paper introduces the concept of convolution object features,proposes a scene recognition method based on double convolutional neural network,and designs two CNN networks,including scene CNN and object CNN.On the one hand,the scene CNN network is used to extract the spatial features in the scene image,and on the other hand,the object CNN network is used to extract the object features in the scene.Finally,the obtained features are input into the support vector machine for scene recognition by means of feature fusion.In order to study the distinguishing characteristics between scenes,this paper adopts the automatic method to retrieve the scene pictures in the crawling webpage,and selects three kinds of scenes: library,desert and swimming pool.After preprocessing,it consists of 5285 scenes composed with Scene-15 data set.The Scene-18 dataset of the image is used to study the discrimination method of similar scenes.3.The existence of general features caused by common objects in different scenes weakens the ability to distinguish between scenes.In order to solve the negative effects caused by ordinary objects in different scenes,this paper proposes an object feature descriptor LOD based on LDA topic model,and proposes a scene recognition method based on subject model object discriminant features.And using the double convolution model to extract the global spatial features and object features,based on the idea of the topic model,construct the probability distribution of "object-picturescene",and select the objects with discriminative characteristics in each type of scene to form the LOD features.The global spatial features are combined with the LOD features and classified using a support vector machine.The above scene recognition method is tested on the Scene-15 dataset and the Scene-18 dataset,and the results obtained by the simulation experiment are compared with the current mainstream scene recognition methods.The experimental results show that the recognition accuracy of the scene recognition method based on the double convolutional neural network on the Scene-15 dataset is higher than that of the current mainstream scene recognition method.However,due to the introduction of convolutional object features,the common objects in different scenarios are ignored.For the negative impact of scene recognition,the recognition performance of the algorithm for a specific scene is not ideal.The scene recognition method based on the subject model object discriminant feature can achieve the best recognition effect on both the Scene-15 dataset and the Scene-18 dataset,and because the method uses the object discriminant feature,it solves the negative of the general object.Impact,the best recognition accuracy is achieved in all categories of scenes.The experimental results show that the scene recognition system designed in this paper improves the scene classification performance and has good generalization ability.