Conference paper
A Visual lexicon to Handle Semantic Similarity in Design precedents
The adequate use of existing knowledge, and not only the creation of completely new solutions, is also an important part of creative thinking. When conceiving a solution, designers oftentimes report having a vague image of the form that will embody the final solution to the design task at hand. The shaping and establishing of this image is supported by creating sketches, collages, models and other types of external representations.
This process is largely supported by the use of design precedents. In design, precedents provide the frame of reference for the development of new solution principles and product forms. They act as sources of knowledge to generate an image of the possible solution space, to get an impression of modes, styles, trends, applications of materials and production/assembly techniques, etc.
Finding the correct precedent can be a very challenging task. This has been recognized by many researchers that have tried to support the process by providing computer tools. Many of these tools have the drawback of relying on metadata added to the image by the editor of the collection. In all the tools studies for this research, such metadata contained, among other things, descriptions of the image using adjectives, bringing along a series of difficulties: it is impractical for large collections; considers only the viewpoint of the editor; fixes descriptions in time and restricts attribution of meaning.
As a solution to this problem, the author has proposed before an approach that uses image recognition techniques to index and retrieve visual information called Content Based Image Retrieval (CBIR). In this approach, the designer gives the computer tool an image and the computer searches for images that are similar to the example given.
For this, the computer looks for geometrical features such as color, texture, shape, mass distribution, etc. The aim of those studies was to completely eliminate human mediated indexing and description of the images. CBIR systems allow the retrieval of results without having to describe, organize and index each image, as is necessary in current systems to handle design precedents.
This allows escaping the subjectivity of the interpretations, escaping the imprecisions of language and avoiding differences in opinion between users. However promising, as was proved in previous tests with 30 designers, this approach falls short of fulfilling all the designer’s needs for visual information.
The reason is that the algorithms available cannot recognize what the image contains (in semantic terms) but humans can, and with great facility. This ability was reflected in the searching process of the designers in our studies. It is very natural for them to expect living room furniture if using a sofa and a lamp as seeds for a query, because a user can understand that these two are related, and that the common aspect is that they are both elements of a living room.
To the system, they are geometrically so different that the results are completely incoherent. Image recognition is useful, but not enough! This paper concludes proposing a strategy to solve this problem in future systems to handle design precedents. Humans have great ability to recognize if two things are related to each other.
For instance, if one takes a vacuum cleaner, an iron and a toaster, it is easy to say that they are all electric household appliances, and that ‘iron’ is more related to ‘toaster’ than to say, ‘scooter’. This was demonstrated by the user that expected to find household products by giving the system a sofa and a phone.
Organizing precedents in a way that allows this ability to be exploited seems to be the next step. This idea has been implemented in a prototype that is currently being tested. Preliminary results show that this is a significant improvement over the system developed for this research.
Language: | English |
---|---|
Publisher: | The Design Society |
Year: | 2007 |
Proceedings: | 16th International Conference on Engineering Design |
ISBN: | 1904670016 and 9781904670018 |
Types: | Conference paper |