Self-Taught Semisupervised Dictionary Learning With Nonnegative Constraint

Abstract

This paper investigates classification by dictionary learning. A novel unified framework termed self-taught semisupervised dictionary learning with non-negative constraint (NNST-SSDL) is proposed for simultaneously optimizing the components of a dictionary and a graph Laplacian. Specifically, an atom graph Laplacian regularization is built by using sparse coefficients to effectively capture the underlying manifold structure. It is more robust to noisy samples and outliers because atoms are more concise and representative than training samples. A non-negative constraint imposed on the sparse coefficients guarantees that each sample is in the middle of its related atoms. In this way the dependency between samples and atoms is made explicit. Furthermore, a self-taught mechanism is introduced to effectively feed back the manifold structure induced by atom graph Laplacian regularization and the supervised information hidden in unlabeled samples in order to learn a better dictionary. An efficient algorithm, combining a block coordinate descent method with the alternating direction method of multipliers is derived to optimize the unified framework. Experimental results on several benchmark datasets show the effectiveness of the proposed model.

Label propagation dictionary learning based process monitoring method for industrial process with between-mode similarity

“…Since the sample labels are represented by linear combinations of atom labels, we apply non-negative constraints on the sparse coding matrix Z to make this more credible [30]. The non-negative constraint ensures that the sample label is in the convex hull of the atom label.…”

“…It can be considered that the profile and the corresponding atom have the same label. We follow the idea in [30] to construct an affinity matrix W = ZZ T to measure the similarity between profiles. The graph Laplacian regularization based on W can be expressed as…”

Target tracking based on standard hedging and feature fusion for robot

“…Generally, now the main strategy of tracking algorithm is to build a faultless appearance model of the target through generative-based (Zhou et al, 2014;Hare et al, 2016;Ning et al, 2016;Gao et al, 2014;Zhou et al, 2013) or discriminative-based (Chan et al, 2017b(Chan et al, , 2017aZhang et al, 2017Zhang et al, , 2020Zhang et al, , 2015bZhang et al, , 2019Zhang et al, , 2015a methods and to update it effectively in subsequent trajectory estimates. However, since the appearance of the object changes differently in each frame of the image sequence, this not only makes it difficult to build a complete description of the appearance of the target object but also leads to the accumulation of errors in the tracking process.…”

Sequence-Dropout Block for Reducing Overfitting Problem in Image Classification

“…A variety of techniques [6]- [11] have been developed to tackle overfitting problem. However, overfitting still remains a major challenge when training large neural networks or having very small amounts of data, and tackling this challenging well can benefit many computer vision tasks, such as classification [12], [13], denoising [14] and tracking [49], [50].…”