With the rapid development of digital capture devices, multimedia storage technology, and network technology, image data are growing up in recent years. These data have become an important and indispensable role in our everyday life. But faced with a growing number of images, finding the desired results successfully for users is becoming ever more difficult. To find an effective way for indexing and realizing image retrieval is a major challenge for us. In this dissertation, we study primarily on image data. We thoroughly study the semantic attribute representation of the image data and the classification and retrieval based on active learning methods. Through the analysis of cross-category nature of the semantic attributes and sufficient network resources, we have established an intermediate layer between the high-level semantic properties and the low level features. By considering the sampling criteria, such as uncertainty, diversity and density, and the user's response on the relevance feedback information, we improve the performance of active learning when the labeled instances are scarce. The main contributions of this paper are as follows: 1.With respect to the shortcomings of semantic attributes which cannot adapt to different applications, especially poor performance in the absence of training samples in image classification (zero shot learning), we propose two semantic attribute augmentation methods. With the analysis of the relationship between the features of a small amount of training samples to be classified (small shot learning) and corresponding categories, we propose two complementary learning: the sequence supplemental feature learning, and discriminative supplementary features learning. Both the methods take into account the extreme lacking in training samples in different ways and use different approaches to extend the semantic attributes space. We improve the performance of the image discriminating feature representation. 2.Early active learning is a very challenge task with respect to the few labeled samples. This dissertation proposes an active learning method based on the semantic attribute space. In conventional active learning method, the user simply serves as a sample with a label, the system simply put this sample into the training sample set and retrain the model. For this conventional active learning mode, with the user defined semantic attribute and a large number of images provided by the network, we have established a s...
修改评论