Object Recognition Based On K-fans Model

K-fans model is part-based model of the special form. According to space position,part-based model can be divided into thin flexible model, layer model, the constellation model,k-fans model and so on. Based on the model of target recognition, the main target is sparseflexible model used by the exterior and space of the position relations renewable model todescribe the visual classes. By assuming a part of model (from the local features and part ofthe geometry to part of the relationship between the position, proportion, and cardinaldirections) to realize the target recognition, this model involved the use of less limitingcondition and the number of local characteristics and each part of the model of geometricshape depends on its k closest to the elements of geometric shape. The positions of objectsand structure of model in layer model are unknown and can be regarded hidden variables.This method completely ratio unchanged, then all variables of the model get by the maximumlikelihood and the only needs to adjust every layer is the number of objects. Because ofreproducibility, the model can coding geometry and common vision object classificationappearance for part of the tools, shape, use spatial relationship of probability codingconnection and its part, the proportion is not variables in the lowest level of key points, aslocal features basis. Constellation model gives a probabilistic way to mixed appearance andlocal descriptor appearance and its main limitation needs specific items (model features andimage characteristics of matching probably).So the model limits the detection of relatedcharacteristics, based on the local part of representation is the focus of the characteristics ofthe correlation of simulation position. K-fans model is from tree-model by extending. Thismodel has k reference parts and the remaining parts can be seen as the reference section. Insimulation of k-fans space position relations, k-fans model presents graphics limit, whenusing gaussian distribution to simulate part of mutual position, considering only positionsrelation of reference parts and positions between non-reference parts and each one of thereference parts, without considering the position relationship between the reference parts.In this paper, we studied single target recognition in static images based on k-fans model,using the maximum possible planning, while estimating appearance and spatial parameters.First run canny edge extraction on local interested part, get a certain number of key points,assign each key point a score based on edge density, and assign each point scale, record theabscissa x value and the vertical axis y values of every point, sort every point based on thescore of every key point, take a number of points whose score is within a certain thresholdfrom these key points as the initial parts of the model, look the highest score of the k points asthe reference part, the remaining parts as the non-reference part, and save the extracted parts of the model as the initial appearance model, then analog the spatial relationship between theparts by the Gaussian distribution, and record the relative position of each non-reference partbetween each reference part as well as non-reference part between non-reference part to getthe initial Gaussian space model, save the initial appearance model and the initial space model,so we get the initial k-fans training model; then regard the appearance model of the initialk-fans training model as a template, run the template on training images, extract a number ofpoints whose correlation is within a certain threshold between each part of the initial k-fansmodel, save these extracted points as interest points, use the weakly-supervised learningmethod for the training of the initial k-fans model, combine the appearance model and thespatial model to do iterating contrast to get select the model parts, save these parts as theappearance model, and the relative space of these parts as a space model, and ultimately getthe trained k-fans model, this method is considered as studying of the spatial characteristics,format a learning model (the consistency of the detection characteristics). The last of thepaper is to change the related parameters to do longitudinal comparison, use1-fan model usedfor the identification of the motorbike target class to design two sets of experiments to studythe effect to experimental results of the training part of the number and the scale of trainingparts, and the parameters changing to the identification effect, verify the effectiveness fortarget recognition based on1-fans model of weakly-supervised learning methods.