Study On The Feature Extraction And Recognition Of The Pressed Protuberant Characters On A Metal Label

The industry product label contains much important information, widely used in all kinds of mechanical and electrical products. It is the major carrier of the enterprise for production management, quality control and product tracking. The research on automatic product label recognition not only has important theoretical significance but also has a broad market prospect.Characters in industry label are comprised by the printing characters and the protuberant characters. The recognition of the latter is more difficulty and important than the former. Because the protuberant character is the 'achromatism' character and has essential difference with other character recognition, the existing character research results do not suit for the protuberant characters. It is necessary to study the recognition method which is suitable to the charcteristis of these charcters.The 'Ring Illumination Image' and the 'Direct CCD Image' pressed protuberant characters are two kinds of characters which are acquired by different methods. The former image effect is better than the latter, but the image acquisition process is more complex than the latter and the image quality is not very stable. On the contrary, the image acquisition of the latter is easy, but the image quality is not ideal. The image characteristic has decided the image processing method and the character recognition algorithm. In the article, different research methods have been adopted to these two kinds of characters.Considering the characteristics of the 'Ring Illumination Image', such as, small size, hard to extraction the internal features as well as having clear edge, we proposed a method which performs the recognition only through the character outlines. The complete frame based the character outline for protuberant characters recognition has been constructed in the paper. The detailed method for character outline extraction, outline features extraction as well as the solution for the existence matters has also been studied.The character edge is foundation to acquire the character outline. In order to overcome the break problem in character edge extraction, a novel edge conjunction method is proposed. The edge conjunction is performed on the second-order gradient template and the boundary tracing has been used as the conjunction tool. Tests show that the method has a good effect on the large break edge conjunction.The features which have invariability on translation, revolving and scale are studied. The Outline Moment extraction and computation method has also been discussed. Studies on the number of the Fourier Descriptors which is necessary to the character construction and recognition also have been performed. The suitable neural network has been constructed for the outline features. Tests show that the two features have bad robustness to the character outline changes though they all have invariability on translation, revolving and scale. Therefore, a feature extraction method which fusing the outline moment and the Fourier descriptor features is proposed. The feature fusion has been studied on the aspects of fusion basis, method, as well as strategy. Experiments have shown that he fusion features have gathered the merits of the two types of features. The recognition rate has been improved to a high level. Another edge detection method which is based on facet model and topography structures has also been studied.The overall features based on the subspace have been adopted to recognition of the 'Direct CCD Image' pressed protuberant characters. The novel method based on PCA subspace has been proposed. The method is carried on steps followed. First, constructs subspace for every character class. Then, withdraw the common statistical features of each subspace. Finally, classifies the test samples by their reconstructing errors under each subspace. The research show that the method utilized the sole features of each class and has gotten a higher recognition rate than a commonly PCA method. The Fisher Discriminant Analysis has also been introduced to the protuberant characters recognition. The recognition tests using LDA and KFDA on protuberant character database have been performed. In order to overcome the widely existed matters of small sample question, the suboptimal of Fisher criterion as well as the large amount of computation problems, a novel Fisher Analysis method- 2D Weighted Linear Analysis(2DWLDA) is proposed. 2DWLDA has applied the idea of weighted to the 2D Linear Analysis(2DLDA). The main idea, the method to compute the analysis vectors as well as the method to decide the weighted function have all been provided. Tests show that the proposed method has not only overcome the problems mentioned above, but also improved the recognition rate and reduced the time of features extraction.The layout analysis of the industrial label, the character segmentation and normalization has also been studied. The research on the layout analysis which is suitable to the specific industry label is performed. The largest-variance characters segment method has been applied to the characters segmentation under the low gray contrast circumstances. The normalized method which is fit for the protuberant has also been studied in the paper.The paper has been supported by the 2006 Doctoral Project Fund of the Ministry of Education.