Density Modeling of Images using a Generalized Normalization Transformation

نویسندگان

  • Johannes Ballé
  • Valero Laparra
  • Eero P. Simoncelli
چکیده

We introduce a parametric nonlinear transformation that is well-suited for Gaussianizing data from natural images. The data are linearly transformed, and each component is then normalized by a pooled activity measure, computed by exponentiating a weighted sum of rectified and exponentiated components and a constant. We optimize the parameters of the full transformation (linear transform, exponents, weights, constant) over a database of natural images, directly minimizing the negentropy of the responses. The optimized transformation substantially Gaussianizes the data, achieving a significantly smaller mutual information between transformed components than alternative methods including ICA and radial Gaussianization. The transformation is differentiable and can be efficiently inverted, and thus induces a density model on images. We show that samples of this model are visually similar to samples of natural image patches. We demonstrate the use of the model as a prior probability density that can be used to remove additive noise. Finally, we show that the transformation can be cascaded, with each layer optimized using the same Gaussianization objective, thus offering an unsupervised method of optimizing a deep network architecture.

برای دانلود رایگان متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

Normalization of qPCR array data: a novel method based on procrustes superimposition

MicroRNAs (miRNAs) are short, endogenous non-coding RNAs that function as guide molecules to regulate transcription of their target messenger RNAs. Several methods including low-density qPCR arrays are being increasingly used to profile the expression of these molecules in a variety of different biological conditions. Reliable analysis of expression profiles demands removal of technical variati...

متن کامل

On Infinitesimal Conformal Transformations of the Tangent Bundles with the Generalized Metric

Let  be an n-dimensional Riemannian manifold, and  be its tangent bundle with the lift metric. Then every infinitesimal fiber-preserving conformal transformation  induces an infinitesimal homothetic transformation on .  Furthermore,  the correspondence   gives a homomorphism of the Lie algebra of infinitesimal fiber-preserving conformal transformations on  onto the Lie algebra of infinitesimal ...

متن کامل

Gaussian Density and HOG with Content Based Image Retrieval System – A New Approach

Content-based image retrieval (CBIR) uses the contents of images to characterize and contact the images. This paper focus on retrieving the image by separating images into its three color mechanism R, G and B and for that Discrete Wavelet Transformation is applied. Then Wavelet based Generalized Gaussian Density (GGD) is practical which is used for modeling the coefficients from the wavelet tra...

متن کامل

Analysis of a Problem Using Various Visions

 In this paper an applied problem, where the response of interest is the number of success in a specific experiment, is considered and by various visions is studied. The effects of outlier values of response on results of a regression analysis are so important to be studied. For this reason, using diagnostic methods, outlier response values are recognized. It is shown that use of arc-sine ...

متن کامل

Image Retrieval and Reversible Illumination Normalization

We propose a novel approach to retrieve similar images from image databases that works in the presence of significant illumination variations. The most common method to compensate for illumination changes is to perform color normalization. The existing approaches to color normalization tend to destroy image content in that they map distinct color values to identical color values in the transfor...

متن کامل

ذخیره در منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید

عنوان ژورنال:
  • CoRR

دوره abs/1511.06281  شماره 

صفحات  -

تاریخ انتشار 2015