Towards the Automatic Evaluation of Visual Balance for Graphic Design Posters

One of the requirements for developing reliable and independent computational creativity systems is the ability to autonomously evaluate aesthetics. However, finding objective metrics to do it effectively is still an open problem.

In Graphic Design (GD), to properly evaluate artefacts, it may be necessary to take into account a number of factors, e.g. the visual relationship of the given artefact to its concept, how legible it is, how innovative it is, and even whether it fits the personal taste of the target audience. Furthermore, visual balance is often a relevant feature to take into consideration.

In this work, we have presented and tested a practical method for evaluating visual balance in 2-dimensional GD artefacts, especially focusing on posters. The proposed method takes as input one \textsc{png} image of any size and ratio. However, to better performance, we automatically resize the input image to 400 pixels wide. Height is set proportionally.

The Centre of Mass (CM) of the given image is calculated considering the brightness and the position of each pixel. First, the default CM is assigned to a vector in the centre of the page (centred vertically and horizontally). Then, each pixel is assigned a weight equal to its inverted normalised brightness, i.e. 0 standing for brighter values and 1 standing for darker ones. This value is then raised to the pow of 2, emphasising the differences between lighter and darker pixels. A vector referring to the position of the given pixel is then multiplied by its respective weight. Lastly, the resulting vector is added to the default CM vector. This way, darker pixels will more strongly attract the CM in their direction to the detriment of lighter ones.

After assessing the CM of the image, the distance to one or more axis can be calculated to estimate balance. In the following experiments, we evaluated posters considering the centre-vertical, centre-horizontal, leftmost-vertical (left-margin) and the lowest-horizontal (bottom-margin) axes (see Figure 1), either alone or mixing two axes together. To mix axes together, their distances to the CM were weighed and then summed up.

Figure 1

Page axes tested.

To test the presented approach, a set of 120 GD posters created by different authors and gathered from various sources were evaluated manually by graphic designers and CC practitioners working on GD, by means of a user survey. Refer to Figure 2 for examples of evaluated posters.

Figure 2

Examples of evaluated posters, grouped by type of source. (1) posters designed on purpose by our research team; (2) posters from the archive; (3) posters from well-known GD studios, gathered from multiple websites; (4) posters from


The results of the survey were then compared to the ones performed by the developed method (see Figure 3). In addition, the respondents were asked how visually appealing they found the posters to be, hoping to retrieve some insights into a supposed correlation between page balance and visual pleasantness.

Figure 3

Average similarity between manual and automatic evaluation values. Maximum value highlighted in bold.


The results suggested the proposed method could match, at around 80%, the balance evaluation made by the respondents. Moreover, the results indicated a possible correlation between page balance and visual pleasantness, at least, for the current experimental setup (for more information, please, refer to the paper).

Future work must focus on testing the proposed approach as a fitness assignment method for an automatic evolutionary system. As assessing balance alone may be reductive to evaluate GD posters, we must complement it with other metrics, such as for assessing legibility or innovation degree. Lastly, we must further study how other visual features, such as apparent movement, may impact the calculation of visual balance.