The Research Of Automatic Method For Natural Scene Image

In recent years, with the rapid development of Internet, the quantity of image data presents geometric growth. Huge amounts of image bring convenience but also pose a challenge for human being, for example, how to retrieve the needed resources from large image database accurately and efficiently is a challenged problem. But due to the existence of the "Semantic gap" problem, which expresses that lower dimensional image features cannot describe rich semantic information. This restricts the development of content-based image retrieval and automatic image annotation, in addition, to deal with the manual image annotation problem, automatic image annotation technology is becoming the hot topic research.The process of the automatic image annotation contains two phases. Firstly, computer learns the relationship model between low dimension visual feature and high-level image semantics by training the annotated images, then, the model finishes the annotation task by applying the model for unknown semantic images. We propose an approach based on fuzzy association rules and decision tree in the paper. Fuzzy association rules is an important method in data mining and classification, it solves the problem of fuzziness attributes in image classification and low accuracy problem, Which improves the efficiency of the algorithm and annotation accuracy greatly. In addition, we use the decision tree algorithm to reduce fuzzy association rules'amount, it improves the algorithm when facing large amounts of fuzzy association rules. On the other hand, we introduce "semantic denoising" thought in this paper. By this thought, the approach calculates the semantic similarity between annotated words. Also, It can improve the precision of semantic annotation after deleting the unrelated words for images.The main contributions of the dissertation are summarized as follows:1. Firstly, the membership function is used to transform the numeric image low dimensional visual feature into semantic fuzzy feature, we can obtain the fuzzy feature vector, then the approach can generate effective fuzzy association rules which are used to capture correlations between image features and semantic concepts in training images. Lastly, the well-known decision tree is added to reduce the number of rules. The method we propose has certain advantages, it crosses the limit of semantic gap problem, furthermore, the cost of time and the size of the algorithm are reduced greatly. Experiments are based on Corel5k and IAPR-TC12 datasets, the evaluation measures include precision, recall and F-measure, experiments prove that the proposed method performs higher accuracy and efficiency in comparison with other automatic annotation methods.2. Furthermore, we introduce "semantic similarity" idea based on fuzzy association rules and decision tree model, adding the denoising process for image annotation. The paper proposes a weighted semantic similarity method which calculates the distance and the depth between each pair words with a weighted value parameter. At the same time, the proposed method removes the annotated word which has lower similarity value. The experiments results show that this denoising process will improve the precision of semantic annotation.