In this paper, we tackle the issue of estimating the noise level of a camera, on its processed still images and as perceived by the user. Commonly, the characterization of the noise level of a camera is done using objective metrics determined on charts containing uniform patches at a given condition. These methods can lead to inadequate characterizations of the noise of a camera because cameras often incorporate denoising algorithms that are more efficient on uniform areas than on areas containing details. Therefore, in this paper, we propose a method to estimate the perceived noise level on natural areas of a still-life chart. Our method is based on a deep convolutional network trained with ground truth quality scores provided by expert annotators. Our experimental evaluation shows that our approach strongly matches human evaluations.