Stable Visual Pattern Mining Via Pattern Probability Distribution.

Yifan Ma,Xuefeng Liang,Xiaoyu Lin,Guanghui Shi

DOI: https://doi.org/10.1007/978-981-99-8543-2_23

2024-01-01

Abstract:Visual patterns are the fundamental elements that compose an image and often convey higher-level semantics. Visual pattern mining can be widely applied to real-world applications and various downstream tasks, such as tourist destination recommendation, scenic area promotion, etc. Therefore, it is a fundamental issue in computer vision. Visual pattern has two properties: frequency and discrimination. Most emerging studies first perform self-supervised representation learning and a supervised classification task to extract discriminative feature, and then mine the visual patterns through clustering algorithms to ensure the frequency. However, these methods require the manual setting of the hyperparameters (e.g. clustering radius eps in DBSCAN, etc.) to which the mining results are sensitive. Improper hyperparameters will result in obtaining redundant visual patterns or missing visual patterns, and the optimal hyperparameters differ among categories. To address these problems, we propose a new stable mining method which directly obtains visual patterns in each category without using clustering algorithms. First, we propose a Pattern Distribution Extractor (PDE) to ensure the discrimination and extract the pattern probability distribution, which represents the probability of the image belonging to each candidate pattern cluster. Experiments demonstrate pattern probability distribution provides more stability in mining stage than the previous methods. Further, we propose a Semantic Density Mining (SDM) strategy which first adaptively removes the non-frequent images to ensure the frequency and then directly obtains the visual patterns in each category according to the pattern probability distribution. Extensive experiments demonstrate the effectiveness of our proposed method.