题目：Content-based image retrieval and an attention selection model
主讲人：顾晓东 副教授
             复旦大学信息学院电子工程系
时间：2011年12月31日上午10:30
地点：物理楼521

       An image retrieval approach using combinational features is introduced to both unsupervised learning image retrieval and supervised learning one. In unsupervised learning image retrieval, compared with SIMPLIcity, FIRM and edge-based method, the proposed method respectively shows 14%, 10%,7% improvement for accuracy in a dataset of Corel image dataset. In supervised learning image retrieval, both one-against-one SVM and one-against-all SVM are implemented. One-against all SVM method achieves an accuracy of more than 95% with sufficient training.
       Topological properties play an important role in human beings visual attention. Using topological properties expression, a TPQFT (Topological properties based Phase spectrum of Quaternion Fourier Transform) attention selection model is proposed. Then TPQFT model is improved to WTPQFT (Weight Adjusted Topological properties based Phase spectrum of Quaternion Fourier Transform) model. The experimental results show that TPQFT model reflects the real attention selection more accurately than PQFT (Phase spectrum of Quaternion Fourier Transform) , and WTPQFT model further show higher accuracy than TPQFT one.