System for processing images

Abstract

The invention discloses a system (1) for processing images (4) comprising a main neural network (2), preferably convolution-based (CNN), and at least one preprocessing neural network (6), preferably convolution-based, upstream of the main neural network (2), for carrying out before processing by the main neural network (2) at least one parametric transformation f, differentiable with respect to its parameters, this transformation being applied to at least part of the pixels of the image and being of the form P' = f(V(p), Theta) where p is a processed pixel of the original image or of a decomposition of this image, p' the pixel of the transformed image or of its decomposition, V(p) is a neighborhood of the pixel p, and Theta is a vector of parameters, the preprocessing neural network (6) having at least part of its learning be performed simultaneously with that of the main neural network (2).

Description

Technical field [0001] The present invention relates to a system for processing images using neural networks, and more particularly but not only to systems for biometrics, especially facial recognition. Background technique [0002] For the recognition of faces or other targets, it has been proposed to use so-called convolution neural networks (CNN). The article Deep Learning by Yann Le Cun et al. (436NATURE, Volume 521, May 28, 2015) includes an introduction to these neural networks. [0003] In addition, it is commonplace to try to perform image preprocessing, such as gamma correction or local contrast correction, in order to correct image defects (such as lack of contrast). [0004] The biometrics of the face assumes a wide variety of image acquisition and lighting conditions, which causes difficulties in the choice of corrections to be made. Moreover, since the improvement in the performance of the convolutional neural network is related to the fully learned hidden layer, this ...

AI Technical Summary

Problems solved by technology

[0006] However, although these systems are more efficient, they are not fully robust to the presence of artifacts and to degradation of image quality

Moreover, increasing the computing power of a computing resource is quite expensive and is not always the right solution

[0007] Thus, existing solutions to the problem of image quality (independent of the problem seeking to learn, which thus consist either in enriching the learning base with examples of problematic images, or in performing image processing upstream) are not fully satisfactory

Images