Scene Classification Based On High-level Image Semantics

Scene classification is always a very challenging topic in computer vision,it is also an important foundation for image retrieval,object detection,and image segmentation.With the ever-growing popularity of computer internet,a large amount of vision data spring up every day in the life,how to manage that effectively is becoming a serious problem for us.In order to classify and retrieve these huge amounts of image data,they should be labeled according to scene semantics,scene classification is just right an effective way to solve the problem.Although there are a lot of methods that have shown great success in scene classification,it is still faced with many difficulties due to the complexity and variability of scene images.Based on above,we in this paper proposed two novel methods with high level scene semantics for scene recognition.Then,image can be labeled on the basis of analysis and understanding of semantic of scenes:1)As the complexity of the scene structure and content grows,scenes can't be classified correctly based on simple image color-texture features.As for a scene,how to understanding the scene semantics and describing structural layout is the key for scene classification.Motivated by this,we have proposed a scene representation based on deformable part models.It can reveal the semantic meaning of a scene by response maps of many object detectors on multi-levels,while describing structure information of a scene by specific scene detectors.So our proposed image representation becomes more robust and effective on the task of scene recognition.2)Convolutional neural network have learned the convolution kernel based on color and texture appeared frequently on images,so deep convolutional activation features is an effective approach for scene recognition based on semantics analysis.However,global CNN features lack geometric invariance,which caused misclassification with highly complicated and variable scenes.So we have proposed a novel method combining salient regions and deep activations.In order to make it more expressive,we encoded the representation with Fisher Vector.The proposed representation can be used as generic feature for complex scene classification.These two methods have been evaluated on some common database on scene classification,such as 15 Scene,UIUC Sports,and MIT Indoor.Experimental results show that our proposed scene classification methods outperform several representative classification approaches.