Study Of Improving The Target/Image Classification Performance

The human eye is the primary access to information,90%of outside information is obtained through eyes. A human eye is a natural recognizer that can recognize any things of wider world. What is computer vision? It mean that computer have the function which is similar to’eye’. The function make computer can help people to do work which must be process by the eyes and brain of human. In recent years, computer vision based on optical image becomes a hot topic.When turn scene into digital image, which need from3D to2D imaging process, it will lead to distortions and inconsistencies. So we need good descriptor to help us expression of image. Good descriptor must be invariance for every situation. We also need classifier to give same responses to similar descriptor. Like this, we complete entire classification system.According to the system of classification, we find that different details of each step will lead to different classification performance, such as different normalization methods, different descriptors dimensions, different organizational methods and different classifier selection. During exploring the impact of these details, this paper presents image descriptors which can be widely applied-Feature Level Co-occurrence Matrix. The initial purpose of the feature co-occurrence matrix is to fill the gaps that bag of words does not included positional relationship of feature words. However, when we study the theory of FLCM, we find that FLCM can be extended in other feature space. For example, in gray feature space, FLCM can be seen as GLCM algorithm.BOW can be seen as histogram algorithm. In local binary patterns feature space, FLCM can be seen as new term LBP algorithm.BOW can be seen as the original LBP algorithm. So we can conclude that FLCM not only can be used for classification but also improve classification accuracy of other descriptor.During study of improving the target/image classification performance, we found that there is big difference between the authenticity of the seal of recognition and general classification problem. Therefore, a complete classification system is proposed for this problem. Firstly, we use scanner to capture images; secondly, we combine SIFT and global features as the descriptors of image; Finally, SVM classifier and KNN classifier respectively achieve effective printed text extraction and classification work. Experimental results show that the accuracy of the classification system can reach97.62%...