Research Of Scene Parsing Based On Multi-feature And Exemplar-SVM

Scene parsing is an important and challenging task in the field of computer vision.It is studied for marking a category label for each pixel in the image.It has extensive application value in many fields,such as automatic driving,robot navigation,environmental monitoring,content based image search and so on.Scene parsing usually includes image segmentation,target class detection,and image multi-label recognition.In the early computer vision research,these problems have been proposed.Many researchers have studied these problems,but it has not been solved well.The difficulty of scene parsing is to acquire the semantic features with strong expressive force in the image and to improve the recognition rate of small image targets(occupying fewer pixels).Aiming at these two problems,this paper proposes a scene parsing method based on multi-feature fusion.Firstly,the retrieval set are established by calculating the image which is most similar to the query image in the training set image,which is similar to the scene,the spatial layout is similar to the object.Therefore,the number of images in the retrieval set is less than the number of images in the training set.After such processing,the class labels of each pixel in the query image are limited to the class labels of the pixels in the retrieval set image,thereby the computation of the algorithm is reduced.Secondly,constructing the global and local features based on the combination of deep convolutional features and traditional SIFT and GIST.The query image and the retrieval set image are matched at the superpixel level,and the classified likelihood value of the superpixel in the query image is obtained,and the maximum likelihood value is taken as the superpixel annotation.Finally,support vector machine(SVM)is used to classify each exemplar and the Markov random field(MRF)energy function is used as the objective function.In order to minimize the energy function and execute the context inference,the MRF energy function combines the data items and the smooth items,thus fine-tuning the superpixel class labels in the query image to get the final image parsing results.In this thesis,the proposed algorithm is verified on an open SIFT Flow dataset and the experimental results show that the proposed scene parsing method based on multi-feature fusion has a high recognition rate.Besides,the algorithm is applied to the real video images and has achieved good results,which verifies the effectiveness of the proposed method.