License Plate Detection And Recognition In Complex Scenes

License plate detection and recognition in complex scenes has always been the focus and difficulty of license plate detection and recognition.The complexity of the natural environment such as illumination changes,weather changes,the complexity of the monitoring scene such as shooting background,shooting distance,shooting angle,equipment's pixel,and the complexity of car license itself such as license plate fouling,license plate occlusion,all the factors above have seriously influence the accuracy and reliability of license plate detection and recognition algorithm.Relevant studies about license plate location and detection algorithm,license plate character segmentation algorithm and license plate character recognition algorithm in complex scenes so as to improve their robustness to the complex environment,has important theoretical and practical significance.Based on the existing license plate recognition methods and considering the complexity of the scenes,this paper has studied respectively on license plate location and detection,license plate character segmentation and license plate character recognition.Details are as follows:For license plate location and detection,this paper proposes a license plate rough location algorithm based on edge features and hierarchical clustering,and a candidate license plate detection algorithm based on HOG features and SVM.Firstly,detecting the edge of image and extracting the edge features,combining the edges by hierarchical clustering,filtering the clustering results and leaving candidate license plate areas,so the license plate rough location is completed.Then,extracting the HOG feature of the candidate license plate areas,and a SVM classifier is trained to eliminate the non license plate areas and complete the detection of the candidate license plate.For license plate character segmentation,this paper proposes a horizontal correction method based on connected region analysis and angle histogram,and a character segmentation algorithm based on character length feature and vertical projection.First,license plate image is processed by two valued and morphological processing,and the connectivity of its region is analyzed.By constructing the angle histogram between regions,we get the approximate tilt angle and complete the horizontal correction of license plate image.Then,using vertical projection to segment characters,and using character length feature to optimize segmentation results,disconnect the bonding characters and combine disconnected Chinese characters.For license plate character recognition,this paper proposes a character recognition algorithmbased on multi classification SVM and weighted revoting.According to the distribution of license plate characters in specific locations,four kinds of SVM combination classifier are designed for Chinese character recognition,alphabet recognition,alphanumeric recognition and special character recognition.When calling SVM combination classifier for character recognition,this paper designs a weighted revoting method,using the initial voting results to weigh each classifier and recount the number of votes.Experiments show that the weighted revoting method can effectively avoid the problem of classification overlap.Compared with the traditional simple cumulative voting method,the weighted revoting method has higher recognition accuracy.