Object Recognition And Localization Based On Part-based Model

When part-based model applies to target recognition, with regard to the representation ofan image, we need to consider (1) computational complexity,(2) the position relationshipbetween the various parts of research entities,(3) the appearance of the various parts of theentity description,(4) image occlusion and background clutter. Target recognition ultimategoal is learning training model and to identify training model object entity, part-based modelis a model algorithm which is a combination of part appearance and part spatial relationships.Therefore, in learning and training of the model, we need to pay attention to the problem:(1)how to simulate the position relationship between the part (2) how would you describe theappearance of the various parts, and (3) to determine based on the number of pixels or the sizeof the area is sparse or dense,(4) how to deal with occlusion and background clutter.Part-based model differs from the bag of the word model. The various components of thebag of word model assume that the object entities are independent of each other. It has a goodperformance for object classification and recognition, but for the positioning of the object, theperformance is poor. Part-based model assumes that certain spatial relationships between thevarious parts of the extracted effective combination of the appearance of the object entity andpart of space and has a good performance for object classification and recognition andpositioning. Part-based model of classification to identify and locate an image, it usually takesfour steps:(1) feature detection,(2) an object composed of some of the spatial relationship ofthe simulation,(3) the object part of the appearance of description,(4) training modeltemplates extracted feature points of the input sample image.Appearance model can be independent selected independent of the space model andinference algorithm. In the same object model, using the various types of appearance modelsis possible. In the building of part-based appearance model, we have to calculate the possibleprobability of each part at each position, instead of using the characteristics of the detector togenerate the position set of each part, in other words, when it is applied to produce theprobability of cost chart on the entire configuration space of an image. We also look theappearance of the model as a feature operation.Object identification using the computer is roughly divided into the following steps.Firstly, make the necessary pretreatment on every image, and then train a certain number ofimages using a specific method to build the training model, finally do target identification oninput test images with the training model that has been got. This paper mainly studies thepart-based model used in the recognition and localization of single object in static image,gives a space model and appearance model, and by studying part-based model based on different space models, give a tree part-based model which is based on the spatial domainprocessing and Gaussian distribution. This paper also studies how tree part-based model to getthe appearance model using spatial domain image processing method, how to get theGaussian space model by Gaussian distribution simulating the spatial relationship between theanalog part, how to deal with test images using deformable templates, and how to extractfeature points whose associated degree with every part of the training model within a certainthreshold. Finally, this paper also set area radius of every identify part, the number of everyidentify part and the number of training images as a parameter. Changing the trainingparameters can get different learning models. Build our own database, and do a lot ofexperiments on objects, such as faces. Get the impact on recognition results which are causedby parameter changing, and analysis the law in changes, lying the foundation for furtherresearch.