Deep Multi-task Image Clustering with Attention-Guided Patch Filtering and Correlation Mining

Deep Multi-task image clustering endeavors to leverage deep learning techniques for the simultaneous processing of multiple clustering tasks. Current multi-task deep image clustering approaches typically rely on conventional deep convolutional neural networks and transfer learning technologies. However, suboptimal clustering results are produced in the execution of each task including irrelevant redundant information. This paper proposes a novel end-to-end deep multi-task clustering framework named Deep Multi-Task Image Clustering with Attention-guided Patch Filtering and Correlation Mining (APFMTC) that eliminates redundant information between different tasks while extracting relevant information to achieve improved cluster division. Specifically, APFMTC partitions image samples into several patches, treating each patch as a word thus each image is regarded as an article, and the process of determining the cluster to which an image belongs is likened to categorizing articles. During the clustering process, several parts of each image sample generally carry more significance. Therefore, a weights estimation module is designed to evaluate the importance of different visual words extracted by the key patch filter for different categories. Ultimately, in each task, the final cluster division is determined by assigning weights to the words contained within the image samples. To evaluate the effectiveness of the proposed method, it is tested on multi-task datasets created from four datasets: NUS-Wide, Pascal VOC, Caltech-256, and Cifar-100. The experimental results substantiate the efficacy of the proposed method.