In this work, we introduce a Self Organizing Map (SOM) based system for the visual similarity analysis and visualization of art image data. In our system, two types of art images, i.e. painting and graphic design images, are focused. Different from traditional art imaging systems which barely focus on associating artistic meanings with image features and seeking artistic interpretations for image analysis results, our system aims to bridge the gaps between artistic concepts, image features and numeric solutions, which allows the numeric analysis results to be explained by art concepts and to be visualized easily. Therefore, using our system, art image data can be better organized and explored; visual similarities of art images can be better understood and explained. 

Elements of art (e.g. line, shape, form, space, color, texture) and principles of design (e.g. balance, emphasis, repetition, movement) are commonly used by art theorists for formal critique of paintings. In our work, we define painting image features that quantify color, line and texture in elements of art; balance and emphasis in principles of design. Since balance and emphasis are both related to composition, we summarize our extracted features into four generic categories: color, composition, line and texture.

• In terms of color, "visual temperature of color", "visual weight of color", "color contrast" are "interpreted" into numeric image features.
• In terms of composition, concepts such as "balance" and "emphasis" are defined using graph based saliency model.
• In terms of line, we use Hough Trasnform to capture the different styles of straight lines in different styles of paintings.
• In terms of texture, the slope of the log amplitude spectrum is used to measure the closeness of the painting texture to natural texture, and thereon to quantify the abstract level of paintings.

In design theory, any 2D graphic designs (e.g. logos, trademarks) are composed of individual basic components (referred as "Forms") of different shapes, sizes and colors. For color features, artistic color features that encodes color temperature, color weight and contrast can be used for logo images. However, since most of logo images do not have complex color schemes, in this work, we mainly focus on the shape features of logo images. Hilbert space filing curve based shape descirptor is adopted and enhanced for shape description.

For our experiments in painting style analysis, we obtain painting collections of six artists representing three art movements: post-impressionism, cubism and renaissance. Through our experimental results, artistic styles of different painting collections and the influential relationships between different paintings can be analysed and visualized.　In the images below the clustering and visualization results using SOM for clustering three styles of paintings are shown. Hierachical clustering is used to measure the relationship between artistis and styles.

 

For our experiments in Logo design analysis, similar approach is used. We use SOM to cluster two different logo image dataset. The first dataset is MPEG-7 shape image dataset, the other is graphic map symbol image dataset. After clustering, logos that share similar characteristics such as colour, shape, can be found to assist further design. The image below shows the visualization of similar logo images clustered by each nodes on SOM.

 

 

Publications: