A Novel Deeply-Learned Image Quality Analysis Algorithm for Clustering

Abstract

Addressing the limitations of existing deep clustering methods, which struggle with variations in image size and quality and are vulnerable to data noise and model deviations, we propose a deeply-learned clustering paradigm in an unsupervised context. This framework utilizes a multi-layer deep architecture, in which the standard fully-linked layers are replaced by the deep convolutional ones in order to intelligently calculate semantic visual representations. Furthermore, we integrate a contraction regularizer derived by the contracted autoencoders to the optimization task of our designed clustering, so as to enhance the clustering structure's quality. This model's simplified structure allows for more efficient end-to-end training and shows enhanced stability and resilience against variations in initial network parameter settings. Experimental evaluations on the MNIST (Modified National Institute of Standards and Technology database) and STL-10 (Self-Taught Learning 10) indicate that our model surpasses other leading clustering architectures in terms of clustering efficacy. Besides, furthermore, the learned deep features can facilitate generic visual recognition substantially, as demonstrated by our empirical comparison with plenty of successful visual recognition models.