Figure 7: Original image (a), clustered image (b) and segmented image (c)
Simulation of Kohonen's maps
Simulation of competitive networks
Simulation of competitive networks Adopting this model of competitive learning, an application suitable to solve the problem of colour clustering has been added to the environment. Much work in computer vision still focuses on gray scale images even though these techniques cannot usually be applied to colour images. For example, the segmentation of a colour image is a more complicated problem than that of a gray scale image. Arbib [Arbib 94] shows theoretically and experimentally that competitive learning converges to approximate the optimum solution. The net architecture will be affected by the characteristics of the image space and the number of input neurons has to be the same as the dimension of the colour space. In particular, we adopt the Red/Green/Blue space to visualize the colour distribution in the images. The number of output neurons depends on the number of the colour clusters into which we would like to subdivide the image colours. During the learning phase, the Red/Green/Blue values of random pixels of the input image are assigned as state to the input nodes and their positions in the colour space are visualized in a 3D visualization window together with the positions of the output neurons. This animation shows the learning process actually taking place in the network. When the training phase is over the original image is processed (see for instance Fig. 7-a); the Red/Green/Blue coordinates of its pixels are fed into the network and their values are substituted by the colour of the corresponding cluster (see Fig. 7-b). This new output image will present as many colours as the number of output neurons in the network. Moreover the application can use the output image to produce the segmented image by applying an algorithm to detect the borders in the colour regions, as shown in Fig. 7-c. Image 7-b is obtained by a network of 3 input neurons totally connected to only 4 output neurons.
Figure 7: Original image (a), clustered image (b) and segmented image (c)
Simulation of Kohonen's maps
Simulation of competitive networks
Simulation of competitive networks