Deep Comprehensive Correlation Mining for Image Clustering

Abstract

Recent developed deep unsupervised methods allow us to jointly learn representation and cluster unlabelled data. These deep clustering methods mainly focus on the correlation among samples, e.g., selecting high precision pairs to gradually tune the feature representation, which neglects other useful correlations. In this paper, we propose a novel clustering framework, named deep comprehensive correlation mining (DCCM), for exploring and taking full advantage of various kinds of correlations behind the unlabeled data from three aspects: 1) Instead of only using pairwise information, pseudo-label supervision is proposed to investigate category information and learn discriminative features. 2) The features' robustness to image transformation of input space is fully explored, which benefits the network learning and significantly improves the performance.3) The triplet mutual information among features is presented for clustering problem to lift the recently discovered instance-level deep mutual information to a triplet-level formation, which further helps to learn more discriminative features. Extensive experiments on several challenging datasets show that our method achieves good performance, e.g., attaining 62.3% clustering accuracy on CIFAR-10, which is 10.1% higher than the state-of-the-art results 1 .

Deep Semantic Clustering by Partition Confidence Maximisation

“…(1) CIFAR-10(/100) [28]: A natural image dataset with 50,000/10,000 samples from 10(/100) classes for training and testing respectively. We adopted the clustering setup same as [24,44,7]: Using both the training and test sets (without labels) for CIFAR10/100 and STL-10, and only the training set for ImageNet-10, ImageNet-Dogs and Tiny-ImageNet; Taking the 20 super-classes of CIFAR-100 as the ground-truth.…”

“…Existing deep clustering approaches generally fall into two categories according to the training strategy: (1) Alternate training [49,46,7,44,6,50,15] and (2) Simultaneous training [23,36,35,16] Alternate training strategy usually estimates the ground-truth membership according to the pretrained or upto-date model and in return supervises the network training by the estimated information. DEC [46] initialises cluster centroids by conducting K-means [32] on pretrained image features and then fine-tunes the model to learn from the confident cluster assignments to sharpen the resulted prediction distribution.…”

“…The method in [49] jointly optimises the objectives of Auto-Encoder [2] and K-means [32] and alternately estimates cluster assignment to learn a "clustering-friendly" latent space. DAC [7], DDC [6] and DCCM [44] exploit the inter-samples relations according to the pairwise distance between the latest sample features and train the model accordingly. Whilst explicit local learning constraints on cluster assignment computed from either the pretrained or up-to-date models usually lead to a deterministic clustering solution, these approaches suffer from the problem of more severe error-propagation from the inconsistent estimations in the neighbourhoods during training.…”

“…The 1 st /2 nd best results are indicated in red/blue. The results of previous methods are taken from [44,24]. † : The best result among multiple trials.…”

“…Recent deep clustering models either iteratively estimate cluster assignment and/or inter-sample relations which are then used as hypotheses in supervising the learning of deep neural networks [44,6,7,50], or used in conjunction with clustering constraints [24,16,23]. In ideal cases, such alternation-learning methods can approach the performance of supervised models not the least benefiting from their robustness against noisy labels [18].…”