Scene Classification Based On Sparse Representation

Scene image classification has always been a difficult point in the field of computer vision,and has many applications in digital image management,video surveillance,remote sensing image processing and other fields.With the rapid development of social imaging detection technology and the construction of high-speed communication networks,massive data images and accurate information flow behind image streams are generated.In order to facilitate the management of massive digital images,the need for image classification becomes more and more urgent.The scene image classification mainly has three difficulties:1.There is large number of content in the image;2.The highly similarity between the scene image classes;3.The numerical value of interclass variance of scene image is large.Therefore,scene image classification needs to take into account the content features and image structure characteristics.Sparse representation is an effective method to take features,and it has strong advantages in anti-noise.The sparse representation-based scene classification method proposed in this paper is studied from the aspects of dictionary construction,target image processing,and classifier training.The main research results achieved by the major work undertaken are as follows:1.The over-complete dictionary based on the pouch model is used to sparsely represent the scene image,and preserve the structural features of the image with slicing images.First,the scene to be classified is determined,the lexical image blocks of the scene to be classified are roughly cut,the image blocks are rotated and expanded,and the image blocks are classified to form an over-complete dictionary of sparse representation.The target image is cut in a ratio of 1:1:1:1 according to the top,bottom left,bottom middle,and right,and the cut images are sparsely represented in combination with the over-complete dictionary of the word bag model.After the sparse representation of the target image,there is no need to reconstruct the image.Only by analyzing the sparse representation vector,can we know which kind of lexical features are included in the image;The content contained in each part of the image can be obtained,which can not only obtain the content semantic features of the scene but also take into account the structural features of the scene.2.A naive Bayesian classification algorithm based on the gravitational model is used.The squared term in the gravitational model is used as the anti-jamming term:due to the square of the distance in the gravitation model,when there is external interference,the target will not deviate from the orbit,but will be in the original orbit.A gradually attenuating disturbance is produced on it.Similar distances are used to simulate the anti-jamming items in the gravitational model to enhance the anti-jamming ability of Naive Bayes algorithm.Compared with Sparse Representation-based Classification,Nuclear sparse multi-block rotation sparse classification algorithm,Bag-of-Words classification algorithm,Segmented sparse representation classification algorithm based on word bag model is more effective in analyzing the internal features of images.It is more efficient to classify scenes in small sample.