Multi-View Diffusion Process for Spectral Clustering and Image Retrieval

主讲人:李奇林
讲座时间:2024-01-02 10:00:00
讲座地点:致远楼304 (线上:腾讯会议ID: 836-716-839)
主办单位:自动化工程学院
主讲人简介:Qilin Li received the BSc degree in Computer Science from Sun Yat-Sen University, PR China, in 2013, the MPhil and PhD degrees from Curtin University, Australia, in 2016 and 2020. He is currently a Lecturer in the School of Electrical Engineering, Computing and Mathematical Sciences at Curtin University. His research interest is mainly in computer vision and pattern recognition.
讲座内容:

In this presentation, we explore a cutting-edge method in multi-view graph learning, an approach that innovatively integrates weight learning with graph learning within an alternating optimization framework. This technique, pivotal in scenarios lacking prior data distribution knowledge, constructs a unified affinity graph from diverse data representation sources. Our unique fusion-and-diffusion strategy amalgamates multiple affinity graphs through a weight learning scheme, which is developed from unsupervised graph smoothness. These fused graphs then serve as a consensus foundation for further diffusion.

The cornerstone of our methodology is a novel multi-view diffusion process. It innovatively learns a manifold-aware affinity graph by propagating affinities across tensor product graphs. This process not only captures high-order contextual information but also significantly enhances pairwise affinities. A key advantage of our approach is its independence from the limitations that typically constrain existing multi-view graph learning methods, such as the reliance on the quality of initial graphs or the presupposition of a latent common subspace among multiple views. Our method adeptly identifies consistencies across views and adaptively fuses multiple graphs.

We present a unified framework that encompasses both weight learning and diffusion-based affinity learning, accompanied by an alternating optimization solver with a convergence guarantee. This novel approach has been applied to image retrieval and clustering tasks across 16 real-world datasets. Our extensive experiments reveal that this method surpasses current leading techniques in both retrieval and clustering in 13 out of these 16 datasets.