Image Segmentation on Convolutional Neural Network (CNN) using Some New Activation Functions

Image segmentation means subdividing the image into different objects. We use different methods for the segmentation of images. For getting different objects from a single image, we generally apply old methods, such as Fuzzy C-means, and K-means clustering techniques. In the case of getting different objects from huge amounts of image datasets, we generally apply the CNN technique. The Activation Function (AF) plays a major role during the segmentation of images through CNN. We can increase the power of CNN using activation functions. Generally, the exponential activation function is used for very complex and non-separable types of problems. So, in this paper, we segment the triangle and CamVid datasets on CNN using NewSigmoid AF and root2Sigmoid AF. Both AFs are exponential activation functions. We made a 12-layer CNN for the triangle segment and a 13-layer CNN for the CamVid segment. After that, we compare four activation functions on both CNNs. These activation functions are NewSigmoid, Relu, root2sigmoid, and Tanh. On the triangle image dataset, we have achieved global accuracy of 96.29% through NewSigmoid AF, 95.77% through Relu AF, 97.02% through root2sigmoid AF, and 96.47% through Tanh AF. On the CamVid dataset, we have achieved global accuracy of 93.97% through NewSigmoid AF, 94.01% through Relu AF, 94.19% through root2sigmoid AF, and 94.13% through Tanh AF. We also check the validation accuracy of the CamVid dataset. We take 10% of images for validation purpose. For the same dataset and same CNN network, we found validation accuracy of relu is 96.04%, root2simgoid is 96.26%, tanh is 96.20% and NewSimgois is 96.26%. Therefore, we can say that the root2sigmoid AF gives a better result as compared with NewSigmoid, Relu, and Tanh AFs.