Research On Method And Application Of Target Recognition Based On Feature Salience

The pattern recognition in image processing is a theory and technique which extracts real-time the features of target and searches the optimal matching target, after building the steady feature model. In target recognition, the Feature Abstraction and feature selection (i.e. how to express feature) are an important problem, which affect the performance of classifier, especially relate the effect of whole recognition system. Aiming at the problems of view changing in target detection and recognition, this paper changes multi-scale and multi-change features into building modeling of selection, abstraction, organizing and fusion. On the basis of this theory, the feature salience is proposed to press multi-features of target. And then, the arrangement model of multi-features should be built.For the multi-features fusion, this paper selects the parallel model and series model to compare them. In this basic, this paper analyses their advantages and disadvantages each other. In parallel model, as to the different salience features, the system assigns the different powers. The more salient the feature is, the larger power system assigns, which shows the different salience features have various contributes to target recognition. In series model, as to the recognition result, this paper uses the dynamic Bayesian net or D-S evidence theory to estimate and compute the target believe degree. And then, according to the salient feature model, the system selects preferentially the next feature that lead to minimum probability of error to update the believe degree. When the believe degree exceeds the setting threshold, the recognition procedure will be finished. This method not only improves the recognition performance, but also increases the robustness.In factual application, aiming at the auto-charge system on the park, vehicle management system in community and vehicle peccancy note system, this paper designs the static state vehicle recognition system. This system studies and builds the experiment worktable about three key techniques-license plate location, character segment and character recognition in License Plate Recognition (LPR). The license plate location and license plate character recognition are typical target detection and recognition problems, and the feature selection and recognition algorithm are the kernel problems in target recognition. Firstly, in this paper, the importance of feature selection in target recognition is introduced. And then, some problems about feature selection are analyzed in brief. On the base of this, the theory of feature salience is presented and the advantage of this algorithm is introduced in detail.According to the feature salience, in this paper, the shape feature, texture feature and color feature are selected to locate the license plate. And the system utilizes the parallel model to fuse the salient features that have been assigned to different power. As to the computation of believe degree, the system makes for length-to-width radio, texture density and hue weight. And the candidate region that corresponds to the maximum confidence function is the vehicle license plate.In the license plate character segment, there is a problem of license plate incline. So that, in this paper, a simple and effective method is proposed to solve this problem. And this algorithm avoids the computation complexity of Hough transform. In order to segment the whole the license plate to single characters, a novel segment algorithm named one dim is presented. And this method scatter effectively disturb of the rivet and the noise. When the license plate includes disturbance, this paper uses the resegmentation based on local threshold.In the license plate character recognition, this paper uses the theory of feature salience also, and selects the moment feature, contour feature, configuration feature and gridding feature in sequence to recognize the license plate characters. In this part, the paper utilizes the D-S evidence theory to fuse the salient features. As to the basic believe assignation, paper uses the believe degree based on the minimum Euclid distance. In fusion procedure, when the believe degree is larger than the setting threshold, the recognition procedure will be finished.Finally, this paper proposed the performance estimate method to whole license plate recognition system. And then, on the basic of this method, system makes an analysis aiming to the scale transform, rotating transform and added-noise transform.