Maximum mutual information and Tsallis entropy for unsupervised segmentation of tree leaves in natural scenes

To identify plant leaves in natural scenes, the accuracy of leaf segmentation is very important. Where there are no assumptions about the colour and position of the leaf and the background, segmentation of leaves in natural scenes is very difficult. Furthermore, an image segmentation algorithm with fixed parameters will not produce correct results for all images. The fusion of the results of the unsupervised segmentation algorithms, normally leads to a better result than any of the individual ones. Generally, the image segmentation fusion, deals with a large amount of data, so its speed is very important. In this paper, a very fast method is introduced for image segmentation fusion based on the maximum mutual information and the state table. To obtain the best consensus segmentation, instead of the classical Shannon entropy, we used the Tsallis entropy and generalized the equation for mutual information which has an additional parameter. To find the best parameter value, the features of the segmentation results are compared with the predetermined shapes. Experiments are performed on tree leaves images with natural background that are part of Pl@ntLeaves dataset and no prior knowledge is used about leaf colour and position. The results show that the application of Tsallis entropy, improves the performance of tree leaves image segmentation fusion in comparison with the classical Shannon entropy.