Unsupervised Wound Image Segmentation

The purpose of our research is to introduce two unique approaches for the unsupervised segmentation of wound images. The first method of segmentation is by using the texture of the image and is performed using the multi-channel filtering hypothesis for the processing of visual or optical data during the preliminary stages of the Human Visual System. We obtain the different channels from an image by filtering the image using a Gabor Filter Bank. The textural features are obtained from each filtered image and the final segmented image is acquired by reconstructing the original input image from these filtered images. The second method of segmentation was performed using parametric kernel graph cuts. Using a kernel function we transform the image data implicitly such that a piecewise constant model of the graph cut interpretation is now applicable. The objective function comprises of an original data term in order to assess the deviance of the transformed data from the initial input data within each partition. This method avoids sophisticated modelling of the input image data while availing of the computational advantages of graph cuts. By using a conventional kernel function, the energy minimization boils down to image partitioning via graph cut iterations and assessments of region parameters by means of fixed point calculations. The efficacy and flexibility of both the methods are established by carrying out investigations on real wound images.