Learning texture discrimination masks

Texture segmentation using multichannel filtering involves applying a set of masks to an input image, and then grouping the pixels based on the responses to these masks. We solve the problem of finding an optimal set of masks by designing a neural network which is trained to maximize a relevant function. Two algorithms, the centroid algorithm and the gradient descent algorithm, are used to train the network. Experimental results on segmenting two natural textures and extracting barcodes in an image are reported, and the error rates compared for both the algorithms with different network configurations. The centroid algorithm gives better results in small parameter spaces, whereas the gradient descent algorithm works better with more parameters. Our method of automatically generating texture discrimination masks not only results in a good segmentation performance, but also reduces the dimensionality of the feature space compared to previously published multichannel filtering methods.<<ETX>>