Understanding deep image representations by inverting them

TLDR

Image representations—from SIFT and Bag of Visual Words to CNNs—are essential to most image‑understanding systems, yet their internal workings remain poorly understood. The study aims to determine how well an image can be reconstructed from its encoded representation. The authors propose a general framework for inverting image representations. The framework accurately inverts HOG and CNN representations, revealing that several CNN layers preserve photo‑accurate image details with varying geometric and photometric invariance.

Abstract

Image representations, from SIFT and Bag of Visual Words to Convolutional Neural Networks (CNNs), are a crucial component of almost any image understanding system. Nevertheless, our understanding of them remains limited. In this paper we conduct a direct analysis of the visual information contained in representations by asking the following question: given an encoding of an image, to which extent is it possible to reconstruct the image itself? To answer this question we contribute a general framework to invert representations. We show that this method can invert representations such as HOG more accurately than recent alternatives while being applicable to CNNs too. We then use this technique to study the inverse of recent state-of-the-art CNN image representations for the first time. Among our findings, we show that several layers in CNNs retain photographically accurate information about the image, with different degrees of geometric and photometric invariance.