Research On Scene Classification Based On Object Detection

With the fast development of network and computer technology, digital images occupyan increasingly important position in our daily life. The conventional scene classificationmethods mostly use low-level image features such as color, texture, and so on. Robustlow-level image features have been proven to be effective representations for a varietyrecognition tasks such as object recognition and scene classification. For high-level visualtasks, such low-level image representations are potentially not enough, because they carrylittle semantic meanings. A high-level image representation called the Object Bank(OB)carries high-level semantic information rather than low-level image feature information,making it more suitable for high-level visual recognition tasks. Based on the study of theOB method, improvement has been done in order to overcome the disadvantages of the OBmethod in this paper.Firstly, a new improved OB scene classification method based on Fisher weight wasproposed in order to overcome the disadvantage of using too high-dimensional featurevector in OB method. The use of high-dimensional feature vector greatly affects theclassification efficiency of OB method, however, low-dimensional feature vectors reducesthe classification accuracy. In improved method based on Fisher weight, the new featurevector, which is obtained by adding the Fisher weight to the low-dimensional feature vectorin OB without spatial pyramid processing, maximizes the differences of the between-classdata and minimizes the differences of the within-class data. It can greatly improve theclassification efficiency of OB method at the same time improve the classification accuracy.Secondly, an improved OB scene classification method based on within-class andbetween-class scatter weight was proposed in order to overcome the disadvantage of notdistinguishing the contribution of different objects. The contribution of different objects forclassification is not the same, the object which has a small within-class scatter and a largebetween-class scatter is beneficial to classification and it should be given a large weight; onthe contrary it is given a small weight. The results show that the proposed method canimprove the classification accuracy.Thirdly, a high-level image representations method based on high-dimensional space coordinates was proposed. The method calculates the clustering centers of differentcategories by K-means clustering algorithm, and then calculates the distances between thefeature vector and the clustering centers of different categories, finally replaces the featurevector with these distances. The new feature vector directly describes the relationshipbetween it and each category, and it is beneficial to classification. The experiments showthat the proposed method can improve the classification performance.Finally, a high-level image representations method based on the O2C distance wasproposed. Compared with the high-level image representations method based on thehigh-dimensional space coordinates, the method selects the minimum distance between thefeature vector and different cluster centers as O2C distance to represent an image. It cangreatly reduce the dimension of the feature vector and improve the classification efficiency.The superiority-inferiority of the proposed method was analysed in detail in the end.