Scene Recognition Based On Deep Neural Network

Since the 21st century,the rapid development of Internet technology,image data has become an indispensable resource,it is full of our daily life,its quantity is also rising in a hierarchical way.How to classify and manage massive image data quickly and efficiently has become a necessary problem that we are facing.Scene image recognition task is also a branch of image recognition.The purpose of scene image recognition is that,according to the environment of image content,the object and the layout,to classify the scene image into one of the predefined scene categories.It has a wide range of applications,including intelligent video monitoring,disaster detection and robot navigation and so on,gradually developed into an important research direction in the field of computer vision.At present,the traditional scene classification cannot achieve high accuracy like the image classification,because most of the existing method of the scene classification task ignores the complicated relation between multiple local characteristics.Images used for scene classification usually contain multiple typical objects with flexible spatial distribution.In addition to the global scenario representation,the relationships between local features at the object level should also be considered.Therefore,two new scene recognition methods are proposed in this paper,which are respectively considered from the vertical and horizontal aspects of local feature processing.In this paper,a scene recognition method based on graph neural network is proposed.The graph structure is constructed by mining several feature regions in the global feature map of the scene image,and the feature regions are processed vertically.The graph neural network updates the state of the nodes by relying on the state of the surrounding edge nodes.In the pooling phase,a graph adaptive pooling method named GSAPool is adopted,which takes into account two aspects of edge structure information and node characteristic information in the graph structure.And then we designed the structure more reasonable and accurate graph topology.We conduct feature aggregation before discarding some nodes,so that the information of partially discarded nodes can be retained and effective information with distinguishing power can be prevented from being contained in the partially discarded nodes.Furthermore,the relationship between the local image features of different locations in the scene image is mined to improve the accuracy of classification.Secondly,a new end-to-end scene recognition network framework,namely recurrent memory location network,is proposed.Horizontal object-based scene classification is performed by repeatedly locating and memorizing objects in the target region of each iteration.This method uses a multitask mechanism,which can continuously mine different local objects in the scene image and classify after repeatedly making memory fusion on the characteristics of the objects.And the corresponding loss function is designed to optimize the problems in the circular location.The final experimental results show that the proposed recurrent memory location network model can effectively improve the accuracy of the classification system.