Robust Image Clustering Based On Low Rank Representation

Image clustering is an important unsupervised learning method in computer vision.The high dimensionality and the complexity of image have brought great challenges to image clustering.On the one hand,by imposing a prior constraint on the image data,the traditional clustering methods are robust when the image is corrupted by noise.On the other hand,deep learning has proved to have superhuman abilities in numerous fields such as image classification and image processing,which provides a new direction for image clustering.This thesis studies the methods of traditional clustering and deep clustering,and proposes improvement based on low-rank theory respectively.Images of the same target may have different characteristics due to factors such as illumination and shooting Angle.In this paper,the low-rank matrix recovery algorithm is firstly used to remove the existing noise in the image,so as to improve the clustering accuracy.Aiming at the over-penalization problem of nuclear norm and7)₁norm,we propose a matrix recovery model based on quasi-norm and7)₀norm.The model uses quasi-norm in place of nuclear norm as a constraint for low rank structures and uses7)₀norm instead of7)₁norm as a constraint for sparse structures which alleviates over-penalization.We further decompose the quasi-norm and design an alternating proximation optimization algorithm to solve the non-convex objective function effectively.The matrix recovery experiment shows that the proposed method can effectively remove the noise in the image,and the face clustering experiment shows that the algorithm can improve the robustness and accuracy of the image clustering.Most of existing methods are based on a linear model which may depress the performance when the image is grossly corrupted.Kernel based methods map original image into latent feature space,but a golden rule for choosing the kernel functions is still absent in practice.To solve these problems,a novel deep clustering method with low-rank prior is proposed in this paper.Convolutional neural network has powerful feature extraction ability.It can automatically extract features from the image so that Manual designing of kernel functions or feature selection is not necessary any more.Autoencoder can preserve local structures of image and compress the input into a lower-dimensional representation.On the other hand,low-rank prior helps autoencoder capture global structures of the whole image dataset.The framework learns to preserve local and global structure simultaneously by taking low-rank reconstruction error of features into consideration.Therefore,low-dimensional feature space learned by autoencoder is more suitable for clustering methods based on low-rank representation.In addition,lower dimension feature space is clustering-friendly.The experimental results show that the proposed method is robust and effective.