Rethinking multi-pattern mining from a perspective of pattern prototype learning

Guanghui Shi, Xuefeng Liang, Xiaoyu Lin

Abstract: Visual pattern mining aims to discover discriminative and frequent visual elements in images. Earlier studies have adopted a two-step framework comprising representation learning followed by pattern mining. However, the reliance on clustering algorithms in the mining step introduces hypersensitivity to parameter selection, resulting in either redundant or missing visual patterns. Furthermore, the optimization based on two non-interactive steps prevents the framework from adapting itself to diverse data. In this work, we rethink the multi-pattern mining task from the perspective of prototype learning, and introduce an end-to-end Visual Pattern Prototypes Proposing (VP3) method with two losses: class-wise and pattern-wise contrastive loss. VP3 eliminates the reliance on clustering algorithms by treating prototypes as visual patterns, and integrates the mining results (pattern prototypes) into representation learning. Consequently, VP3 can achieve more robust results. Extensive experiments conducted on four benchmark datasets demonstrate the robustness and superior performance of VP3 over six state-of-the-art methods. Code is available at https://github.com/BeCarefulOfYournaoke/VP3.

External IDs:dblp:journals/pr/ShiLL26