Image Object Retrieval Based On Content And Semantic Information

With the rapid development of network and e-commerce, searching based text description cannot meet person's needs. So how to effectively retrieve image based on the information of the image itself from the large-scale database has become an urgent issue. Therefore, the Content-based Image Retrieval, Semantic-based Image Retrieval and how to use them in goods image retrieval have received great attention.In low-level feature selection and extraction, we analyze different features which can be summed up in four categories:color feature, texture feature, shape feature, key-point feature. Based on the feature of goods image itself, advanced block HSV color histogram and SIFT (Scale-invariant Feature Transform) feature are extracted to compose the low-level feature of goods image. Bag-of-Words algorithm is used to effectively quantify the high-dimensional key-point feature and make key-point feature easy to cascade other feature. The experiment performed on self-test data set and public test data set shows that the fusion of features is robust and effectively.In semantic concepts extraction, we define special semantic feature set based on properties and features of goods image. Color feature, Gabor feature, edge histogram and contour feature are extracted to obtained semantic concepts. Support vector machine is used to build multi-class classifiers. Prior probability model is used to optimize the classifier results.In luggage retrieval system, we combine semantic-based image retrieval and content-based image retrieval. The image retrieval in large-scale image database is sped by cascading the two parts together. The precision is up to 90.3% and speed is 3 seconds per image. Experiment results performed on self-test set show that proposed system has a fast speed and high precision.