Research On Scene Recognition On Deep Learning

Scene recognition is a challenging subject in the field of computer vision.It is the basis of computer vision tasks such as image retrieval,object recognition and image semantic understanding.With the rapid development of network technology and the popularity of new intelligent devices,the number and categories of scene images are constantly increasing,and the intra-class difference and inter-class similarity of scene images have an increasingly obvious impact on the accuracy of scene recognition.A single feature is difficult to express rich image information and obtain satisfactory accuracy.At present,it is urgent to construct an effective representation of scene image to make the image features more discriminative.In view of the above problems,this paper makes an in-depth study on scene recognition by using the deep learning method.The main work is as follows:(1)In order to better represent the effective information worthy of attention in the scene image and obtain more discriminative features,this paper introduces the attention mechanism into the scene recognition task and proposes a multi-scale attention network model for scene recognition,Through multiple scale branches to construct a scene image attention representation that takes into account channels and spaces.The improved residual channel attention structure is used in the model to enhance the information of key objects that are worthy of attention;for the problem of information loss in the process of spatial attention calculation,a spatial attention structure based on different scales is proposed to enhance the key areas of attention Position information,and then use the feature complementation between multiple scales to obtain the final representation of the scene image;and the central loss function joint supervision strategy is introduced to further reduce the misjudgment of intra-class differences,effectively improve the discrimination of features.(2)In order to further overcome the influence of inter-class similarity,a network model based on multi-level context information is proposed,which describes the mutual position relationship between objects through context characteristics and effectively classifies the categories with high similarity among objects.First,the scene segmentation network is migrated to the scene recognition task,and the context relationship of the image is preliminarily modeled.Secondly,in view of the diversity of the relationship between objects,put forward the direction LSTM module,to extract context information from four directions,and connect to multiple convolution middle layer,forming processing level characteristics,more make the extraction to the context of the feature space layout information,both from the bottom from the top abstract semantic information.Finally,an effective representation of image context information is formed by combining multi-level context features.(3)In order to enrich the characteristics of the images,said several fusion strategy research and comparative analysis,this will eventually attention characteristics and context information fusion images into comprehensive said,give full play to the characteristics of complementary,make the final characteristics of said both to focus on identifying information effectively,and can represent significant relative position relations between objects,effectively solve the scene said the problem of inadequate recognition task characteristics,improve the identification accuracy.Finally,the models and method proposed in this paper are experimentally verified in the Scene15,MIT indoor67,and SUN397 datasets which with a large number of complex scenes,and the effectiveness of the proposed algorithm is verified,the overall accuracy reached 95.83%,85.69% and 72.65%,respectively.The experimental results show that the fusion of attention and context information recognition method,effectively improve the scene recognition accuracy and compared with other scene recognition algorithm also has certain advantages.