Wavelet packet based feature extraction and recognition of license plate characters

To study the characteristics of license plate characters recognition, this paper proposes a method for feature extraction of license plate characters based on two-dimensional wavelet packet. We decompose license plate character images with two dimensional-wavelet packet and search for the optimal wavelet packet basis. This paper presents a criterion of searching for the optimal wavelet packet basis, and a practical algorithm. The obtained optimal wavelet packet basis is used as the feature of license plate character, and a BP neural network is used to classify the character. The testing results show that the proposed method achieved higher recognition rate than the traditional methods. Keyword: license plate, characters recognition, feature extraction, wavelet packet, optimal wavelet packet basis. DOI: 10.1360/982004-429 The automatic license plate recognition technique has important practical value in the intelligent transportation system, in which the key points are to locate license plate and recognize the license plate characters. Many researchers have done a large amount of research work 6] on this realm. In this paper the focus is laid on license plate characters recognition, which extracts the feature of license plate characters and classifies characters according to license plate location. License plate recognition system works outdoors. images acquired under real conditions are often anamorphic, and their resolution is low and easily suffers from various interferences. It is difficult to get ideal recognition results using the traditional characters recognition method. The typical method for license plate recognition is based on template matching, but it is sensitive to the size, incline and background interference of the character images, and the result is not satisfactory. The details of character image at different resolutions characterize different structural features. Lee applied two-dimensional wavelet transform to handwritten digit recognition and verified the performance of the proposed scheme with three different numeral databases. The recognition results were satisfactory. But, with the increase of spatial resolution, frequency resolution decreases, which is a disadvantage of wavelet transform. By further dividing frequency domain, wavelet packet transform can overcome the deficiency of wavelet transform and has better time-frequency characteristic. Analyzing the characteristics of the license plate characters recognition, this paper applies two-dimensional wavelet transform to license plate character recognition, proposes a feature extraction method of license plate characters based on two-dimensional wavelet packet, and presents a criterion of searching for the optimal wavelet basis and a practical algorithm. With the optimal wavelet packet basis as the feature of license plate characters, a BP neural network is used to classify the characters. 1 Wavelet packet quad-tree for images Lee proposed a new multiresolution recognition scheme for recognizing handwritten numerals using two-dimensional wavelet transform. Firstly the character image is normalized to 16 16 image and decomposed with Harr wavelet, and then multiresolution feature vector is composed with all wavelet coefficients at resolution 2 and 2. Finally a neural network is used to classify the characters. This method has the following shortcomings: 1) it fails to further decompose high frequency information 2) all wavelet coefficients are used as the multiresolution feature, the feature number is big and the selection of feature is not flexible. Considering the above disadvantages, in this paper a feature extraction method based on wavelet packet is presented. Through wavelet packet decomposition of the normalized character images, the feature vector is composed of wavelet packet decomposition coefficients. In order to reduce computation work and the dimensions of feature vector, a method for searching for optimal wavelet basis is proposed. The obtained optimal wavelet basis is used as the feature vector of license plate characters. Two-dimensional wavelet packet quad-tree is composed of separable wavelet packet spaces. Each node of the quadtree is labeled by a scale 2 j and two integers p and q ( 0 2 , j L p − < 0 2 j L q − < ), corresponding to a separable space. , p q p q j j j W W W = ⊗ . (1) For 1 2 ( , ) x x x = , the separable wavelet packet is , 1 2 ( ) ( ) ( ) p q p q j j l x x x ψ ψ ψ = . (2) An orthogonal basis of , p q j W can be obtained with a separable product of the wavelet packet basis of p j W