Researches On Multi-feature Fusion For License Plate Detection

License Plate Recognition(LPR) is an important component of Intelligent Transportation System(ITS). It has many applications in our daily life, such as electronic toll collection, parking entrance management, traffic law enforcement. There are mainly three parts in the LPR system: license plate detection, character segmentation, character recognition. License plate detection is the most signif icant step of all. The results of detection impact the subsequent process and the performance of the whole system. Many factors in the outdoors may interfere the detection, such as complex background, part cover and non-uniform illumination, so how to detect the license plate accurately is still challenging.This paper focuses on license plate detection algorithm based on multi feature fusion. The main work are written as follows:Firstly, it describes some technology about license plate detection as well as features of license plate, at the same time, analyses the feasibility and the advantage of the feature-based method. Besides, it introduces all kinds of feature-based method existing and their advantages and disadvantages.Secondly, as for the fact that the already existing methods based on edge density usually have a relatively low location rate and can detect license with the same size only. On the basis of the mu lti-scale object detection, we propose a new method EM-LPD which depends on edge and multi-scale theory. After establishing the image pyramid, it detects license plate on each scale image with a size-fixed window, and selects some regions with a high edge density as candidate license plates, and finally combines the detected results on each scale.The results show that EM-LPD can get a high detection rate on the test dataset.Thirdly, as for the fact that the already existing methods based on edge density usually have a relatively high false positive rate, we add the process of filter ing the false positives. Based on the characteristics of the characters, we select texture feature to describe license plate. Considering the advantage of multi feature in object detection, we introduce two texture features: T-HOG and MB-LBP and T-HOG is more sensitive to image noise, while MB- LBP is robust to noise, so we combine these two features to describe the license plates. Experimental results show that they perform well in the application of license plate detection.With data fusion technology, we propose two new methods to license plate verif ication, which are DF-LPV and FF-LPV. Combined with the two features above, they filter the candidate license plates in the way of decision fusion and feature fusion. Using the technology of information fusion can filter the false positives well.Multi feature fusion reflects in two ways: on one hand, from an overall point of view, this paper uses edge feature in the license plate detection, uses texture feature in the false positives filter ing process, on the other hand, in the local process of filtering, it selects two different texture features: T-HOG and MB-LBP based on the characteristic of license plate so as to enhance the discrimination of license plate. These features are complementary. Experimental results show that method based on multi feature fusion can improve the performance of the license plate detection.