Unsupervised learning of compositional sparse code for natural image representation

Abstract

Abstract. This article proposes an unsupervised method for learning compositional sparse code for representing natural images. Our method is built upon the original sparse coding framework where there is a dictionary of basis functions often in the form of localized, elongated and oriented wavelets, so that each image can be represented by a linear combination of a small number of basis functions automatically selected from the dictionary. In our compositional sparse code, the representational units are composite: they are compositional patterns formed by the basis functions. These compositional patterns can be viewed as shape templates. We propose an unsupervised learning method for learning a dictionary of frequently occurring templates from training images, so that each training image can be represented by a small number of templates automatically selected from the learned dictionary. The compositional sparse code approximates the raw image of a large number of pixel intensities using a small number of templates, thus facilitating the signal-to-symbol transition and allowing a symbolic description of the image. The current form of our model consists of two layers of representational units (basis functions and shape templates). It is possible to extend it to multiple layers of hierarchy. Experiments show that our method is capable of learning meaningful compositional sparse code, and the learned templates are useful for image classification.

Learning Sparse FRAME Models for Natural Image Patterns

“…Compared to our own previous work, this paper can be considered a fusion of the original FRAME model [66] and the active basis model [59] [29]. While the active basis model focuses on the "sketching" aspect, this paper adds the "painting" aspect.…”

“…The two-stage learning algorithm can serve as a basis for learning a codebook of sparse FRAME templates from nonaligned images without any annotation and labeling, so that the training images can be represented by spatially translated, rotated, scaled and deformed versions of templates selected from the learned codebook. Here we follow the learning scheme in our previous work on compositional sparse coding [29].…”

“…So after the template matching pursuit process, the observed images are reconstructed according to (29).…”

“…The number of templates in the codebook as well as the numbers of basis functions in the templates can be selected by BIC-like criteria, as suggested by [29]. In this paper, we hand picked these parameters.…”

Unsupervised Learning of Dictionaries of Hierarchical Compositional Models

“…In this paper, we learn dictionaries of hierarchical compositional templates from unaligned natural images without annotations, which is more challenging. In comparison to our past work, [11] is concerned with learning templates with only one layer of deformations, while [25] is concerned with learning a single template instead of learning a dictionary of multiple templates.…”