Last Friday I presented the paper “PicSOM – content-based image retrieval with self-organizing maps ” of Laaksonen et al. This authors uses T-SOMS (Tree-based SOMs) to classify images based in their content: using distinct measure types, like colour, sFFT (shape Fast Fourier Transform) and others several SOM maps are created (one per measure). The solution to how to give weight to that measures is simple: the user feedback. A set of images is presented to the user in a web-based application, so he can select the interested ones (example: Wynona Ryder’s face) to obtain the next set of images (an example can be seen in Figure 1). The algorithm learns with the selected images and gives weight to select the images of an specific map (in our example, shape measures are more important than colour measure).
Figure 1: Winona Ryder’s face in PicSom interface
To test the performance of every map and the global performance of the whole system the authors use sumatory things to stablish if the selected images belongs to a determined class (i.e. Planes, dinosaurs of faces). The conclusion is that it is necessary to use the whole set of measures at the same time to acquire the best performance. But the most interesting part is that you can read the paper and test the algorithm by yourself, a not so common practice.