This paper proposes a method to extract the emotional impact of images. Emotions are often associated with facial expressions, but we decided consider other features as first emotional characteristic of natural images, which, in general, does not contain faces. For a seek of generally we have built a new image database composed of a large variety of low semantic images. We used colour images because often colours and emotions are supposed to be linked. For the modelling of the emotions, we considered colours features completed with other recent and efficient descriptors. We supposed that different features used could also implicitly encode high-level information about emotions. The concept of emotion is not easy to model. The perception of emotion is not only influenced by the content and the colour of the images. It is also modified by some personal experiences like cultural aspects and personal semantic associated to some colours or objects. The complexity of emotion modelling was considered in classification process through psycho-visual tests. The twenty-five observers assessed the nature and the power of emotions they felt. These tests allowed us to distinguish three classes of emotions, which are "Negative", "Neutral" and "Positive" emotions. We used a Support Vector Machine for classification and the average success rate is 51,75%; that is really relevant regarding the equivalent results in the literature.